Robotics and automation in the food industry: Current and future technologies 1845698010, 9781845698010

Citation preview

Robotics and automation in the food industry

© Woodhead Publishing Limited, 2013

Related titles: Computer vision technology in the food and beverage industries (ISBN 978-0-85709-036-2) Food processing technology: principles and practice Third edition (ISBN 978-1-84569-216-2) Instrumentation and sensors for the food industry Second edition (ISBN 978-1-85573-560-6) Details of these books and a complete list of titles from Woodhead Publishing can be obtained by: • •

•

visiting our web site at www.woodheadpublishing.com contacting Customer Services (e-mail: [email protected]; fax: +44 (0) 1223 832819; tel.: +44 (0) 1223 499140 ext. 130; address: Woodhead Publishing Limited, 80, High Street, Sawston, Cambridge CB22 3HJ, UK) in North America, contacting our US office (e-mail: [email protected]; tel.: (215) 928 9112; address: Woodhead Publishing, 1518 Walnut Street, Suite 1100, Philadelphia, PA 19102-3406, USA

If you would like e-versions of our content, please visit our online platform: www.woodheadpublishingonline.com. Please recommend it to your librarian so that everyone in your institution can benefit from the wealth of content on the site. We are always happy to receive suggestions for new books from potential editors. To enquire about contributing to our Food Science, Technology and Nutrition series, please send your name, contact address and details of the topic/s you are interested in to [email protected]. We look forward to hearing from you.

The Woodhead team responsible for publishing this book: Commissioning Editor: Sarah Whitworth Publications Coordinator: Anneka Hess Project Editor: Cathryn Freear Editorial and Production Manager: Mary Campbell Production Editor: Mandy Kingsmill Freelance Project Manager: Vedhapriya Badrinarayanan Cover designer: Terry Callanan

© Woodhead Publishing Limited, 2013

Woodhead Publishing Series in Food Science, Technology and Nutrition: Number 236

Robotics and automation in the food industry Current and future technologies Edited by Darwin G. Caldwell

Oxford

Cambridge

Philadelphia

New Delhi

© Woodhead Publishing Limited, 2013

Published by Woodhead Publishing Limited, 80 High Street, Sawston, Cambridge CB22 3HJ, UK www.woodheadpublishing.com www.woodheadpublishingonline.com Woodhead Publishing, 1518 Walnut Street, Suite 1100, Philadelphia, PA 19102-3406, USA Woodhead Publishing India Private Limited, G-2, Vardaan House, 7/28 Ansari Road, Daryaganj, New Delhi – 110002, India www.woodheadpublishingindia.com First published 2013, Woodhead Publishing Limited © Woodhead Publishing Limited, 2013; except Chapters 1 and 11. Chapter 1 was prepared by a US Government employee, it is therefore in the public domain and cannot be copyrighted. Chapter 11 © Buhler Sortex Ltd, 2013. Note: the publisher has made every effort to ensure that permission for copyright material has been obtained by authors wishing to use such material. The authors and the publisher will be glad to hear from any copyright holder it has not been possible to contact. The authors have asserted their moral rights. This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. Reasonable efforts have been made to publish reliable data and information, but the authors and the publisher cannot assume responsibility for the validity of all materials. Neither the authors nor the publisher, nor anyone else associated with this publication, shall be liable for any loss, damage or liability directly or indirectly caused or alleged to be caused by this book. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming and recording, or by any information storage or retrieval system, without permission in writing from Woodhead Publishing Limited. The consent of Woodhead Publishing Limited does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from Woodhead Publishing Limited for such copying. Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation, without intent to infringe. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library. Library of Congress Control Number: 2012948531 ISBN 978-1-84569-801-0 (print) ISBN 978-0-85709-576-3 (online) ISSN 2042-8049 Woodhead Publishing Series in Food Science, Technology and Nutrition (print) ISSN 2042-8057 Woodhead Publishing Series in Food Science, Technology and Nutrition (online) The publisher’s policy is to use permanent paper from mills that operate a sustainable forestry policy, and which has been manufactured from pulp which is processed using acid-free and elemental chlorine-free practices. Furthermore, the publisher ensures that the text paper and cover board used have met acceptable environmental accreditation standards. Typeset by Newgen Publishing and Data Services, India Printed and bound in the UK by the MPG Books Group

© Woodhead Publishing Limited, 2013

Contents

Contributor contact details .......................................................................... Woodhead Publishing Series in Food Science, Technology and Nutrition.............................................................................

xv

Part I Introduction, key technologies and significant areas of development..................................................................................

1

1 Automatic process control for the food industry: an introduction ..................................................................................... Y. Huang, United States Department of Agriculture, USA 1.1 Introduction .................................................................................. 1.2 Process control systems and structure in the food industry ......... 1.3 Process control methods in the food industry .............................. 1.4 Future trends ................................................................................ 1.5 References .................................................................................... 2 Robotics in the food industry: an introduction.................................. J. O. Gray, The University of Manchester, UK and S. T. Davis, University of Salford, UK 2.1 Introduction .................................................................................. 2.2 Current manufacturing procedures .............................................. 2.3 Automation in the food sector...................................................... 2.4 Specifications for a food sector robot .......................................... 2.5 Future trends ................................................................................ 2.6 Conclusion ................................................................................... 2.7 References ....................................................................................

© Woodhead Publishing Limited, 2013

xi

3 3 4 5 16 18 21

21 23 25 26 32 34 35

vi

Contents

3 Sensors for automated food process control: an introduction ......... P. G. Berrie, Endress+Hauser Process Solutions AG, Switzerland 3.1 Introduction .................................................................................. 3.2 Special considerations for food instrumentation.......................... 3.3 Measurement methods ................................................................. 3.4 Device integration ........................................................................ 3.5 Applications of sensors in automated food process control......... 3.6 Future trends ................................................................................ 3.7 Conclusion ................................................................................... 3.8 References ....................................................................................

36

4 Machine vision in the food industry ................................................... E. R. Davies, Royal Holloway, University of London, UK 4.1 Introduction .................................................................................. 4.2 Machine vision: principles and methods...................................... 4.3 Applications and case studies ...................................................... 4.4 Recent advances in the application of vision in the food industry ................................................................................ 4.5 Appraisal of the need for special hardware for food inspection applications ................................................................. 4.6 Conclusion and future trends ....................................................... 4.7 Acknowledgements ...................................................................... 4.8 Sources of further information and advice................................... 4.9 References ....................................................................................

75

5 Optical sensors and online spectroscopy for automated quality and safety inspection of food products ............................................... C. B. Singh and D. S. Jayas, University of Manitoba, Canada 5.1 Introduction .................................................................................. 5.2 Optical sensing and spectroscopic techniques ............................. 5.3 Applications in the food industry................................................. 5.4 Future trends ................................................................................ 5.5 Conclusion ................................................................................... 5.6 References .................................................................................... 6 Supervisory Control and Data Acquisition (SCADA) and related systems for automated process control in the food industry: an introduction ..................................................................... J. F. Holmes and G. Russell, Georgia Tech Research Institute, USA and J. K. Allen, The University of Oklahoma, USA 6.1 Introduction to Supervisory Control and Data Acquisition ......... 6.2 History of SCADA....................................................................... 6.3 SCADA standards and applications ............................................. 6.4 SCADA in food processing.......................................................... 6.5 Laboratory study: implementation of SCADA ............................

© Woodhead Publishing Limited, 2013

36 37 44 60 67 72 73 73

75 77 92 103 104 105 106 107 108

111 111 112 117 126 126 127

130

130 133 133 136 139

Contents vii 6.6 6.7

Future trends in SCADA.............................................................. References ....................................................................................

140 142

7 Gripper technologies for food industry robots .................................. T. K. Lien, Norwegian University of Science and Technology, Norway 7.1 Introduction .................................................................................. 7.2 Gripper challenges in food process automation ........................... 7.3 Gripping physics .......................................................................... 7.4 Pinching and enclosing grippers .................................................. 7.5 Penetrating (needle) grippers ....................................................... 7.6 Suction grippers ........................................................................... 7.7 Surface effect (freeze) grippers .................................................... 7.8 Selection of the appropriate gripping technology ........................ 7.9 Future trends: from laboratory to industry ................................... 7.10 References ....................................................................................

143

8 Wireless sensor networks (WSNs) in the agricultural and food industries ...................................................................................... N. Wang, Oklahoma State University, USA and Z. Li, South China Agricultural University, P. R. China 8.1 Introduction .................................................................................. 8.2 Current state of development of WSNs ....................................... 8.3 WSN applications in agriculture and food production ................ 8.4 Future trends in WSN technology in agriculture and food production ............................................................................ 8.5 References .................................................................................... 9 Intelligent quality control systems in food processing based on fuzzy logic......................................................................................... N. Perrot and C. Baudrit, INRA, France 9.1 Introduction .................................................................................. 9.2 Principles of intelligent control systems using fuzzy logic ......... 9.3 Current applications in the food industry..................................... 9.4 Advances in research and future trends ....................................... 9.5 References .................................................................................... 10 Advanced methods for the control of food processes: the case of bioconversion in a fed-batch reactor .............................................. D. Dochain, Université catholique de Louvain, Belgium 10.1 Introduction .................................................................................. 10.2 The basic dynamical model ......................................................... 10.3 Modelling issues: population balance modelling in food processes .............................................................................. 10.4 Monitoring issues: tuning of observer-based estimators.............. 10.5 Design of PID controllers for fed-batch processes ......................

© Woodhead Publishing Limited, 2013

143 144 147 148 154 157 163 167 169 169

171

171 172 184 195 195

200 200 203 211 220 223

226 226 229 231 233 239

viii Contents 10.6 10.7 10.8 10.9

Real-time optimization................................................................. Acknowledgements ...................................................................... Conclusion ................................................................................... References ....................................................................................

246 260 261 261

Part II Robotics and automation in particular unit operations and industry sectors ...................................................................

265

11 Automation and robotics for bulk sorting in the food industry ....... G. Hamid, B. Deefholts, N. Reynolds, D. McCambridge, K. Mason-Palmer and C. Briggs, Buhler Sortex Limited, UK 11.1 Introduction .................................................................................. 11.2 Principles of operation ................................................................. 11.3 Requirements ............................................................................... 11.4 Recent advances in technology .................................................... 11.5 Current applications ..................................................................... 11.6 Conclusion ................................................................................... 11.7 Future trends ................................................................................ 11.8 Sources of further information and advice................................... 11.9 References ....................................................................................

267

12 Automatic control of food chilling and freezing ................................ C. James and S. J. James, Grimsby Institute of Further & Higher Education (GIFHE), UK 12.1 Introduction: key drivers and challenges for automatic control of food chilling and freezing ........................................................ 12.2 Automation in refrigerated food retail display............................. 12.3 Automation of refrigeration and freezing operations in food catering ............................................................................ 12.4 Automation in refrigerated food transport systems ..................... 12.5 Automation in food chilling and freezing systems ...................... 12.6 Automation in food cold storage systems .................................... 12.7 Advances in research and future trends ....................................... 12.8 Sources of further information and advice................................... 12.9 References ....................................................................................

288

13 Robotics and automation in meat processing .................................... G. Purnell, Grimsby Institute of Further & Higher Education (GIFHE), UK 13.1 Introduction .................................................................................. 13.2 Automation of carcass production processes before primary chilling............................................................................ 13.3 Automation of carcass separation processes after primary chilling............................................................................

304

© Woodhead Publishing Limited, 2013

267 268 272 273 275 285 286 286 287

288 290 292 292 294 299 301 301 302

304 309 316

Contents 13.4 13.5 13.6 13.7

Future trends ................................................................................ Conclusion ................................................................................... Sources of further information and advice................................... References ....................................................................................

14 Robotics and automation in the poultry industry: current technology and future trends............................................................... G. McMurray, Georgia Tech Research Institute, USA 14.1 Introduction .................................................................................. 14.2 Robotics and automation in live hanging and first processing of poultry...................................................................................... 14.3 Robotics and automation in second processing of poultry .......... 14.4 Robotics and automation in bulk packing and shipping of poultry meat ............................................................................. 14.5 Future trends ................................................................................ 14.6 References ....................................................................................

ix 324 325 326 326

329 329 331 334 347 351 352

15 Robotics and automation in seafood processing ................................ J. O. Buljo and T. B. Gjerstad, SINTEF Raufoss Manufacturing AS, Norway 15.1 Introduction .................................................................................. 15.2 Technologies for robotics and automation in the seafood industry ........................................................................... 15.3 Application of robotics and automation in fish slaughtering, filleting, portioning and associated unit operations ..................... 15.4 Automation in other unit operations in fish processing ............... 15.5 Future trends ................................................................................ 15.6 Sources of further information and advice................................... 15.7 References ....................................................................................

354

16 Robotics and automation in the fresh produce industry .................. N. Kondo, Kyoto University, Japan 16.1 Introduction .................................................................................. 16.2 Machine vision system as a key technology ................................ 16.3 Vegetable preprocessing and grading systems ............................. 16.4 Information flow for food traceability and farming guidance ..... 16.5 Conclusion ................................................................................... 16.6 References ....................................................................................

385

17 Robotics and automation for packaging in the confectionery industry........................................................................... J. S. Dai, King's College London, UK 17.1 Introduction .................................................................................. 17.2 The confectionery market and its business requirements ............ 17.3 Reconfigurable mechanism technology .......................................

© Woodhead Publishing Limited, 2013

354 359 366 373 377 382 383

385 386 389 396 398 399

401 401 402 407

x

Contents 17.4 17.5 17.6 17.7 17.8

Case study of a reconfigurable system for carton folding ........... Future trends ................................................................................ Conclusion ................................................................................... Acknowledgements ...................................................................... References ....................................................................................

408 414 416 416 416

18 Automatic control of batch thermal processing of canned foods ..... R. J. Simpson, S. F. Almonacid, Universidad Técnica Federico Santa María, Chile and Centro Regional de Estudios en Alimentos Saludables (CREAS), Chile and A. A. Teixeira, University of Florida, USA 18.1 Introduction .................................................................................. 18.2 On-line control strategies ............................................................. 18.3 Validation of computer-based control systems ............................ 18.4 Industrial automation of batch retorts .......................................... 18.5 Advances in research and future trends ....................................... 18.6 References ....................................................................................

420

19 Automation for a sustainable food industry: computer aided analysis and control engineering methods ............................... A. I. Papadopoulos, Centre for Research and Technology – Hellas, Greece and P. Seferlis, Aristotle University of Thessaloniki, Greece and Centre for Research and Technology – Hellas, Greece 19.1 Introduction .................................................................................. 19.2 Definition of sustainability and links with the food industry ...... 19.3 Automation and sustainability in food manufacturing................. 19.4 Tools for automated sustainable design and operation in food engineering................................................................................... 19.5 Advanced tools and methods for sustainable food engineering with potential applications ....................................... 19.6 Software technologies for automated sustainable design ............ 19.7 Conclusion and future trends ....................................................... 19.8 Sources of further information and advice................................... 19.9 References .................................................................................... Index ............................................................................................................

© Woodhead Publishing Limited, 2013

420 421 432 433 437 439

441

441 443 445 446 466 478 479 480 481 487

Contributor contact details

(* = main contact)

Chapter 2

Editor Professor Darwin G. Caldwell Istituto Italiano di Tecnologia (Italian Institute of Technology) Via Morego, 30 16163 – Genoa Italy E-mail: [email protected]

Chapter 1 Yanbo Huang United States Department of Agriculture Agricultural Research Service Crop Production Systems Research Unit 141 Experiment Station Road Stoneville MS 38776 USA E-mail: [email protected]

Professor John Gray* School of Electrical and Electronic Engineering The University of Manchester Sackville Street Building Manchester M13 9PL UK E-mail: [email protected] Dr Steve T. Davis School of Computing Science and Engineering UG8, Newton Building University of Salford Salford M5 4WT UK E-mail: [email protected]

Chapter 3 Dr Peter G. Berrie Endress+Hauser Process Solutions AG Kaegenstr. 2 4153 Reinach Switzerland E-mail: [email protected]. com

© Woodhead Publishing Limited, 2013

xii Contributor contact details

Chapter 4

Chapter 7

Professor E. R. Davies Machine Vision Group Department of Physics Royal Holloway, University of London Egham Surrey TW20 0EX UK

Professor T. K. Lien Department of Production and Quality Engineering Norwegian University of Science and Technology S. P. Andersens vei 5 7491 Trondheim Norway E-mail: [email protected]

E-mail: [email protected]

Chapter 8 Chapter 5 Chandra B. Singh and Digvir S. Jayas* Biosystems Engineering University of Manitoba Winnipeg, MB R3T 5V6 Canada E-mail: [email protected]

Chapter 6 Jonathan F. Holmes* and Geoffrey Russell Food Processing Technology Division Georgia Tech Research Institute Atlanta Georgia USA E-mail: [email protected]. edu Janet K. Allen The School of Industrial Engineering The University of Oklahoma Norman Oklahoma USA

Ning Wang* Department of Biosystems and Agricultural Engineering Oklahoma State University Stillwater OK 74078 USA E-mail: [email protected] and Zhen Li College of Engineering South China Agricultural University Guangzhou, Guangdong P.R. China, 510642

Chapter 9 Nathalie Perrot* and Cedric Baudrit INRA, UMR782 Génie et Microbiologie des Procédés Alimentaires AgroParisTech, INRA 78850 Thiverval-Grignon France E-mail: [email protected]; [email protected]

© Woodhead Publishing Limited, 2013

Contributor contact details xiii

Chapter 10 Denis Dochain CESAME Université catholique de Louvain Euler building, 4-6 avenue G. Lemaitre 1348 Louvain-la-Neuve Belgium E-mail: [email protected]

Chapter 11 Gabriel Hamid*, Ben Deefholts, Nick Reynolds, David McCambridge, Kasper Mason-Palmer and Craig Briggs Buhler Sortex Limited Research and Development Department 20 Atlantis Avenue London E16 2BF UK E-mail: gabriel.hamid@buhlergroup. com

Food Refrigeration & Process Engineering Research Centre (FRPERC) Grimsby Institute of Further & Higher Education (GIFHE) HSI Building, Origin Way, Europarc, Grimsby North East Lincolnshire DN37 9TZ UK E-mail: [email protected]; [email protected]

Chapter 14 Gary McMurray Georgia Tech Research Institute Food Processing Technology Division 640 Strong Street Atlanta GA 30332 USA E-mail: [email protected]

Chapter 15 Chapter 12 Christian James* and Stephen J. James Food Refrigeration & Process Engineering Research Centre (FRPERC) Grimsby Institute of Further & Higher Education (GIFHE) HSI Building, Origin Way, Europarc, Grimsby North East Lincolnshire DN37 9TZ UK E-mail: [email protected]

Chapter 13 Dr Graham Purnell Senior Research Fellow

Jan Olav Buljo* and Tone Beate Gjerstad SINTEF Raufoss Manufacturing AS S.P. Andersens veg 5 7465, Trondheim Norway E-mail: [email protected]; [email protected]

Chapter 16 Naoshi Kondo Graduate School of Agriculture, Kyoto University, Kitashirakawa-Oiwakecho Sakyo-ku Kyoto 606–8502 Japan E-mail: [email protected]

© Woodhead Publishing Limited, 2013

xiv Contributor contact details

Chapter 17 Jian S. Dai Centre for Robotics Research School of Natural and Mathematical Sciences King’s College London University of London UK Email: [email protected]

Chapter 18 Ricardo J. Simpson* and Sergio F. Almonacid Departamento de Ingeniería Química y Ambiental Universidad Técnica Federico Santa María P.O. Box 110-V Valparaíso Chile E-mail: [email protected] and Centro Regional de Estudios en Alimentos Saludables (CREAS). CONICYT-REGIONAL, R06I1004 Blanco 1623 Room 1402 Valparaíso Chile Arthur A. Teixeira Department of Agricultural and Biological Engineering Frazier Rogers Hall

P.O. Box 110570 University of Florida Gainesville FL 32611-0570 USA

Chapter 19 Athanasios I. Papadopoulos Chemical Process Engineering Research Institute Centre for Research and Technology – Hellas P.O. Box 60361 57001, Thermi-Thessaloniki Greece E-mail: [email protected] Panos Seferlis* Department of Mechanical Engineering Aristotle University of Thessaloniki P.O. Box 484 54124, Thessaloniki Greece E-mail: [email protected] and Chemical Process Engineering Research Institute Centre for Research and Technology – Hellas P.O. Box 60361 57001, Thermi-Thessaloniki Greece

© Woodhead Publishing Limited, 2013

Woodhead Publishing Series in Food Science, Technology and Nutrition

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Chilled foods: a comprehensive guide Edited by C. Dennis and M. Stringer Yoghurt: science and technology A. Y. Tamime and R. K. Robinson Food processing technology: principles and practice P. J. Fellows Bender’s dictionary of nutrition and food technology Sixth edition D. A. Bender Determination of veterinary residues in food Edited by N. T. Crosby Food contaminants: sources and surveillance Edited by C. Creaser and R. Purchase Nitrates and nitrites in food and water Edited by M. J. Hill Pesticide chemistry and bioscience: the food-environment challenge Edited by G. T. Brooks and T. Roberts Pesticides: developments, impacts and controls Edited by G. A. Best and A. D. Ruthven Dietary fibre: chemical and biological aspects Edited by D. A. T. Southgate, K. W. Waldron, I. T. Johnson and G. R. Fenwick Vitamins and minerals in health and nutrition M. Tolonen Technology of biscuits, crackers and cookies Second edition D. Manley Instrumentation and sensors for the food industry Edited by E. Kress-Rogers Food and cancer prevention: chemical and biological aspects Edited by K. W. Waldron, I. T. Johnson and G. R. Fenwick Food colloids: proteins, lipids and polysaccharides Edited by E. Dickinson and B. Bergenstahl Food emulsions and foams Edited by E. Dickinson Maillard reactions in chemistry, food and health Edited by T. P. Labuza, V. Monnier, J. Baynes and J. O’Brien The Maillard reaction in foods and medicine Edited by J. O’Brien, H. E. Nursten, M. J. Crabbe and J. M. Ames Encapsulation and controlled release Edited by D. R. Karsa and R. A. Stephenson Flavours and fragrances Edited by A. D. Swift Feta and related cheeses Edited by A. Y. Tamime and R. K. Robinson Biochemistry of milk products Edited by A. T. Andrews and J. R. Varley Physical properties of foods and food processing systems M. J. Lewis

© Woodhead Publishing Limited, 2013

xvi Woodhead Publishing Series in Food Science, Technology and Nutrition 24 Food irradiation: a reference guide V. M. Wilkinson and G. Gould 25 Kent’s technology of cereals: an introduction for students of food science and agriculture Fourth edition N. L. Kent and A. D. Evers 26 Biosensors for food analysis Edited by A. O. Scott 27 Separation processes in the food and biotechnology industries: principles and applications Edited by A. S. Grandison and M. J. Lewis 28 Handbook of indices of food quality and authenticity R. S. Singhal, P. K. Kulkarni and D. V. Rege 29 Principles and practices for the safe processing of foods D. A. Shapton and N. F. Shapton 30 Biscuit, cookie and cracker manufacturing manuals Volume 1: ingredients D. Manley 31 Biscuit, cookie and cracker manufacturing manuals Volume 2: biscuit doughs D. Manley 32 Biscuit, cookie and cracker manufacturing manuals Volume 3: biscuit dough piece forming D. Manley 33 Biscuit, cookie and cracker manufacturing manuals Volume 4: baking and cooling of biscuits D. Manley 34 Biscuit, cookie and cracker manufacturing manuals Volume 5: secondary processing in biscuit manufacturing D. Manley 35 Biscuit, cookie and cracker manufacturing manuals Volume 6: biscuit packaging and storage D. Manley 36 Practical dehydration Second edition M. Greensmith 37 Lawrie’s meat science Sixth edition R. A. Lawrie 38 Yoghurt: science and technology Second edition A. Y. Tamime and R. K. Robinson 39 New ingredients in food processing: biochemistry and agriculture G. Linden and D. Lorient 40 Benders’ dictionary of nutrition and food technology Seventh edition D. A. Bender and A. E. Bender 41 Technology of biscuits, crackers and cookies Third edition D. Manley 42 Food processing technology: principles and practice Second edition P. J. Fellows 43 Managing frozen foods Edited by C. J. Kennedy 44 Handbook of hydrocolloids Edited by G. O. Phillips and P. A. Williams 45 Food labelling Edited by J. R. Blanchfield 46 Cereal biotechnology Edited by P. C. Morris and J. H. Bryce 47 Food intolerance and the food industry Edited by T. Dean 48 The stability and shelf-life of food Edited by D. Kilcast and P. Subramaniam 49 Functional foods: concept to product Edited by G. R. Gibson and C. M. Williams 50 Chilled foods: a comprehensive guide Second edition Edited by M. Stringer and C. Dennis 51 HACCP in the meat industry Edited by M. Brown 52 Biscuit, cracker and cookie recipes for the food industry D. Manley 53 Cereals processing technology Edited by G. Owens 54 Baking problems solved S. P. Cauvain and L. S. Young 55 Thermal technologies in food processing Edited by P. Richardson 56 Frying: improving quality Edited by J. B. Rossell 57 Food chemical safety Volume 1: contaminants Edited by D. Watson 58 Making the most of HACCP: learning from others’ experience Edited by T. Mayes and S. Mortimore 59 Food process modelling Edited by L. M. M. Tijskens, M. L. A. T. M. Hertog and B. M. Nicolaï 60 EU food law: a practical guide Edited by K. Goodburn 61 Extrusion cooking: technologies and applications Edited by R. Guy

© Woodhead Publishing Limited, 2013

Woodhead Publishing Series in Food Science, Technology and Nutrition xvii 62 Auditing in the food industry: from safety and quality to environmental and other audits Edited by M. Dillon and C. Griffith 63 Handbook of herbs and spices Volume 1 Edited by K. V. Peter 64 Food product development: maximising success M. Earle, R. Earle and A. Anderson 65 Instrumentation and sensors for the food industry Second edition Edited by E. Kress-Rogers and C. J. B. Brimelow 66 Food chemical safety Volume 2: additives Edited by D. Watson 67 Fruit and vegetable biotechnology Edited by V. Valpuesta 68 Foodborne pathogens: hazards, risk analysis and control Edited by C. de W. Blackburn and P. J. McClure 69 Meat refrigeration S. J. James and C. James 70 Lockhart and Wiseman’s crop husbandry Eighth edition H. J. S. Finch, A. M. Samuel and G. P. F. Lane 71 Safety and quality issues in fish processing Edited by H. A. Bremner 72 Minimal processing technologies in the food industries Edited by T. Ohlsson and N. Bengtsson 73 Fruit and vegetable processing: improving quality Edited by W. Jongen 74 The nutrition handbook for food processors Edited by C. J. K. Henry and C. Chapman 75 Colour in food: improving quality Edited by D MacDougall 76 Meat processing: improving quality Edited by J. P. Kerry, J. F. Kerry and D. A. Ledward 77 Microbiological risk assessment in food processing Edited by M. Brown and M. Stringer 78 Performance functional foods Edited by D. Watson 79 Functional dairy products Volume 1 Edited by T. Mattila-Sandholm and M. Saarela 80 Taints and off-flavours in foods Edited by B. Baigrie 81 Yeasts in food Edited by T. Boekhout and V. Robert 82 Phytochemical functional foods Edited by I. T. Johnson and G. Williamson 83 Novel food packaging techniques Edited by R. Ahvenainen 84 Detecting pathogens in food Edited by T. A. McMeekin 85 Natural antimicrobials for the minimal processing of foods Edited by S. Roller 86 Texture in food Volume 1: semi-solid foods Edited by B. M. McKenna 87 Dairy processing: improving quality Edited by G. Smit 88 Hygiene in food processing: principles and practice Edited by H. L. M. Lelieveld, M. A. Mostert, B. White and J. Holah 89 Rapid and on-line instrumentation for food quality assurance Edited by I. Tothill 90 Sausage manufacture: principles and practice E. Essien 91 Environmentally-friendly food processing Edited by B. Mattsson and U. Sonesson 92 Bread making: improving quality Edited by S. P. Cauvain 93 Food preservation techniques Edited by P. Zeuthen and L. Bøgh-Sørensen 94 Food authenticity and traceability Edited by M. Lees 95 Analytical methods for food additives R. Wood, L. Foster, A. Damant and P. Key 96 Handbook of herbs and spices Volume 2 Edited by K. V. Peter 97 Texture in food Volume 2: solid foods Edited by D. Kilcast 98 Proteins in food processing Edited by R. Yada 99 Detecting foreign bodies in food Edited by M. Edwards 100 Understanding and measuring the shelf-life of food Edited by R. Steele 101 Poultry meat processing and quality Edited by G. Mead 102 Functional foods, ageing and degenerative disease Edited by C. Remacle and B. Reusens

© Woodhead Publishing Limited, 2013

xviii Woodhead Publishing Series in Food Science, Technology and Nutrition 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141

Mycotoxins in food: detection and control Edited by N. Magan and M. Olsen Improving the thermal processing of foods Edited by P. Richardson Pesticide, veterinary and other residues in food Edited by D. Watson Starch in food: structure, functions and applications Edited by A.-C. Eliasson Functional foods, cardiovascular disease and diabetes Edited by A. Arnoldi Brewing: science and practice D. E. Briggs, P. A. Brookes, R. Stevens and C. A. Boulton Using cereal science and technology for the benefit of consumers: proceedings of the 12th International ICC Cereal and Bread Congress, 24 – 26th May, 2004, Harrogate, UK Edited by S. P. Cauvain, L. S. Young and S. Salmon Improving the safety of fresh meat Edited by J. Sofos Understanding pathogen behaviour: virulence, stress response and resistance Edited by M. Griffiths The microwave processing of foods Edited by H. Schubert and M. Regier Food safety control in the poultry industry Edited by G. Mead Improving the safety of fresh fruit and vegetables Edited by W. Jongen Food, diet and obesity Edited by D. Mela Handbook of hygiene control in the food industry Edited by H. L. M. Lelieveld, M. A. Mostert and J. Holah Detecting allergens in food Edited by S. Koppelman and S. Hefle Improving the fat content of foods Edited by C. Williams and J. Buttriss Improving traceability in food processing and distribution Edited by I. Smith and A. Furness Flavour in food Edited by A. Voilley and P. Etievant The Chorleywood bread process S. P. Cauvain and L. S. Young Food spoilage microorganisms Edited by C. de W. Blackburn Emerging foodborne pathogens Edited by Y. Motarjemi and M. Adams Benders’ dictionary of nutrition and food technology Eighth edition D. A. Bender Optimising sweet taste in foods Edited by W. J. Spillane Brewing: new technologies Edited by C. Bamforth Handbook of herbs and spices Volume 3 Edited by K. V. Peter Lawrie’s meat science Seventh edition R. A. Lawrie in collaboration with D. A. Ledward Modifying lipids for use in food Edited by F. Gunstone Meat products handbook: practical science and technology G. Feiner Food consumption and disease risk: consumer-pathogen interactions Edited by M. Potter Acrylamide and other hazardous compounds in heat-treated foods Edited by K. Skog and J. Alexander Managing allergens in food Edited by C. Mills, H. Wichers and K. Hoffman-Sommergruber Microbiological analysis of red meat, poultry and eggs Edited by G. Mead Maximising the value of marine by-products Edited by F. Shahidi Chemical migration and food contact materials Edited by K. Barnes, R. Sinclair and D. Watson Understanding consumers of food products Edited by L. Frewer and H. van Trijp Reducing salt in foods: practical strategies Edited by D. Kilcast and F. Angus Modelling microorganisms in food Edited by S. Brul, S. Van Gerwen and M. Zwietering Tamime and Robinson’s Yoghurt: science and technology Third edition A. Y. Tamime and R. K. Robinson Handbook of waste management and co-product recovery in food processing Volume 1 Edited by K. W. Waldron

© Woodhead Publishing Limited, 2013

Woodhead Publishing Series in Food Science, Technology and Nutrition xix 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178

Improving the flavour of cheese Edited by B. Weimer Novel food ingredients for weight control Edited by C. J. K. Henry Consumer-led food product development Edited by H. MacFie Functional dairy products Volume 2 Edited by M. Saarela Modifying flavour in food Edited by A. J. Taylor and J. Hort Cheese problems solved Edited by P. L. H. McSweeney Handbook of organic food safety and quality Edited by J. Cooper, C. Leifert and U. Niggli Understanding and controlling the microstructure of complex foods Edited by D. J. McClements Novel enzyme technology for food applications Edited by R. Rastall Food preservation by pulsed electric fields: from research to application Edited by H. L. M. Lelieveld and S. W. H. de Haan Technology of functional cereal products Edited by B. R. Hamaker Case studies in food product development Edited by M. Earle and R. Earle Delivery and controlled release of bioactives in foods and nutraceuticals Edited by N. Garti Fruit and vegetable flavour: recent advances and future prospects Edited by B. Brückner and S. G. Wyllie Food fortification and supplementation: technological, safety and regulatory aspects Edited by P. Berry Ottaway Improving the health-promoting properties of fruit and vegetable products Edited by F. A. Tomás-Barberán and M. I. Gil Improving seafood products for the consumer Edited by T. Børresen In-pack processed foods: improving quality Edited by P. Richardson Handbook of water and energy management in food processing Edited by J. Klemeš, R. Smith and J.-K. Kim Environmentally compatible food packaging Edited by E. Chiellini Improving farmed fish quality and safety Edited by Ø. Lie Carbohydrate-active enzymes Edited by K.-H. Park Chilled foods: a comprehensive guide Third edition Edited by M. Brown Food for the ageing population Edited by M. M. Raats, C. P. G. M. de Groot and W. A Van Staveren Improving the sensory and nutritional quality of fresh meat Edited by J. P. Kerry and D. A. Ledward Shellfish safety and quality Edited by S. E. Shumway and G. E. Rodrick Functional and speciality beverage technology Edited by P. Paquin Functional foods: principles and technology M. Guo Endocrine-disrupting chemicals in food Edited by I. Shaw Meals in science and practice: interdisciplinary research and business applications Edited by H. L. Meiselman Food constituents and oral health: current status and future prospects Edited by M. Wilson Handbook of hydrocolloids Second edition Edited by G. O. Phillips and P. A. Williams Food processing technology: principles and practice Third edition P. J. Fellows Science and technology of enrobed and filled chocolate, confectionery and bakery products Edited by G. Talbot Foodborne pathogens: hazards, risk analysis and control Second edition Edited by C. de W. Blackburn and P. J. McClure Designing functional foods: measuring and controlling food structure breakdown and absorption Edited by D. J. McClements and E. A. Decker New technologies in aquaculture: improving production efficiency, quality and environmental management Edited by G. Burnell and G. Allan

© Woodhead Publishing Limited, 2013

xx Woodhead Publishing Series in Food Science, Technology and Nutrition 179 More baking problems solved S. P. Cauvain and L. S. Young 180 Soft drink and fruit juice problems solved P. Ashurst and R. Hargitt 181 Biofilms in the food and beverage industries Edited by P. M. Fratamico, B. A. Annous and N. W. Gunther 182 Dairy-derived ingredients: food and neutraceutical uses Edited by M. Corredig 183 Handbook of waste management and co-product recovery in food processing Volume 2 Edited by K. W. Waldron 184 Innovations in food labelling Edited by J. Albert 185 Delivering performance in food supply chains Edited by C. Mena and G. Stevens 186 Chemical deterioration and physical instability of food and beverages Edited by L. H. Skibsted, J. Risbo and M. L. Andersen 187 Managing wine quality Volume 1: viticulture and wine quality Edited by A. G. Reynolds 188 Improving the safety and quality of milk Volume 1: milk production and processing Edited by M. Griffiths 189 Improving the safety and quality of milk Volume 2: improving quality in milk products Edited by M. Griffiths 190 Cereal grains: assessing and managing quality Edited by C. Wrigley and I. Batey 191 Sensory analysis for food and beverage quality control: a practical guide Edited by D. Kilcast 192 Managing wine quality Volume 2: oenology and wine quality Edited by A. G. Reynolds 193 Winemaking problems solved Edited by C. E. Butzke 194 Environmental assessment and management in the food industry Edited by U. Sonesson, J. Berlin and F. Ziegler 195 Consumer-driven innovation in food and personal care products Edited by S. R. Jaeger and H. MacFie 196 Tracing pathogens in the food chain Edited by S. Brul, P.M. Fratamico and T.A. McMeekin 197 Case studies in novel food processing technologies: innovations in processing, packaging, and predictive modelling Edited by C. J. Doona, K. Kustin and F. E. Feeherry 198 Freeze-drying of pharmaceutical and food products T.-C. Hua, B.-L. Liu and H. Zhang 199 Oxidation in foods and beverages and antioxidant applications Volume 1: understanding mechanisms of oxidation and antioxidant activity Edited by E. A. Decker, R. J. Elias and D. J. McClements 200 Oxidation in foods and beverages and antioxidant applications Volume 2: management in different industry sectors Edited by E. A. Decker, R. J. Elias and D. J. McClements 201 Protective cultures, antimicrobial metabolites and bacteriophages for food and beverage biopreservation Edited by C. Lacroix 202 Separation, extraction and concentration processes in the food, beverage and nutraceutical industries Edited by S. S. H. Rizvi 203 Determining mycotoxins and mycotoxigenic fungi in food and feed Edited by S. De Saeger 204 Developing children’s food products Edited by D. Kilcast and F. Angus 205 Functional foods: concept to product Second edition Edited by M. Saarela 206 Postharvest biology and technology of tropical and subtropical fruits Volume 1: fundamental issues Edited by E. M. Yahia 207 Postharvest biology and technology of tropical and subtropical fruits Volume 2: açai to citrus Edited by E. M. Yahia 208 Postharvest biology and technology of tropical and subtropical fruits Volume 3: cocona to mango Edited by E. M. Yahia

© Woodhead Publishing Limited, 2013

Woodhead Publishing Series in Food Science, Technology and Nutrition xxi 209 Postharvest biology and technology of tropical and subtropical fruits Volume 4: mangosteen to white sapote Edited by E. M. Yahia 210 Food and beverage stability and shelf life Edited by D. Kilcast and P. Subramaniam 211 Processed Meats: improving safety, nutrition and quality Edited by J. P. Kerry and J. F. Kerry 212 Food chain integrity: a holistic approach to food traceability, safety, quality and authenticity Edited by J. Hoorfar, K. Jordan, F. Butler and R. Prugger 213 Improving the safety and quality of eggs and egg products Volume 1 Edited by Y. Nys, M. Bain and F. Van Immerseel 214 Improving the safety and quality of eggs and egg products Volume 2 Edited by F. Van Immerseel, Y. Nys and M. Bain 215 Animal feed contamination: effects on livestock and food safety Edited by J. Fink-Gremmels 216 Hygienic design of food factories Edited by J. Holah and H. L. M. Lelieveld 217 Manley’s technology of biscuits, crackers and cookies Fourth edition Edited by D. Manley 218 Nanotechnology in the food, beverage and nutraceutical industries Edited by Q. Huang 219 Rice quality: a guide to rice properties and analysis K. R. Bhattacharya 220 Advances in meat, poultry and seafood packaging Edited by J. P. Kerry 221 Reducing saturated fats in foods Edited by G. Talbot 222 Handbook of food proteins Edited by G. O. Phillips and P. A. Williams 223 Lifetime nutritional influences on cognition, behaviour and psychiatric illness Edited by D. Benton 224 Food machinery for the production of cereal foods, snack foods and confectionery L.-M. Cheng 225 Alcoholic beverages: sensory evaluation and consumer research Edited by J. Piggott 226 Extrusion problems solved: food, pet food and feed M. N. Riaz and G. J. Rokey 227 Handbook of herbs and spices Second edition Volume 1 Edited by K. V. Peter 228 Handbook of herbs and spices Second edition Volume 2 Edited by K. V. Peter 229 Breadmaking: improving quality Second edition Edited by S. P. Cauvain 230 Emerging food packaging technologies: principles and practice Edited by K. L. Yam and D. S. Lee 231 Infectious disease in aquaculture: prevention and control Edited by B. Austin 232 Diet, immunity and inflammation Edited by P. C. Calder and P. Yaqoob 233 Natural food additives, ingredients and flavourings Edited by D. Baines and R. Seal 234 Microbial decontamination in the food industry: novel methods and applications Edited by A. Demirci and M.O. Ngadi 235 Chemical contaminants and residues in foods Edited by D. Schrenk 236 Robotics and automation in the food industry: current and future technologies Edited by D. G. Caldwell 237 Fibre-rich and wholegrain foods: improving quality Edited by J. A. Delcour and K. Poutanen 238 Computer vision technology in the food and beverage industries Edited by D.-W. Sun 239 Encapsulation technologies and delivery systems for food ingredients and nutraceuticals Edited by N. Garti and D. J. McClements 240 Case studies in food safety and authenticity Edited by J. Hoorfar 241 Heat treatment for insect control: developments and applications D. Hammond 242 Advances in aquaculture hatchery technology Edited by G. Allan and G. Burnell 243 Open innovation in the food and beverage industry Edited by M. Garcia Martinez

© Woodhead Publishing Limited, 2013

xxii Woodhead Publishing Series in Food Science, Technology and Nutrition 244 Trends in packaging of food, beverages and other fast-moving consumer goods (FMCG) Edited by Neil Farmer 245 New analytical approaches for verifying the origin of food Edited by P. Brereton 246 Microbial production of food ingredients, enzymes and nutraceuticals Edited by B. McNeil, D. Archer, I. Giavasis and L. Harvey 247 Persistent organic pollutants and toxic metals in foods Edited by M. Rose and A. Fernandes 248 Cereal grains for the food and beverage industries E. Arendt and E. Zannini 249 Viruses in food and water: risks, surveillance and control Edited by N. Cook 250 Improving the safety and quality of nuts Edited by L. J. Harris 251 Metabolomics in food and nutrition Edited by B. Weimer and C. Slupsky 252 Food enrichment with omega-3 fatty acids Edited by C. Jacobsen, N. Skall Nielsen, A. Frisenfeldt Horn and A.-D. Moltke Sørensen 253 Instrumental assessment of food sensory quality: a practical guide Edited by D. Kilcast 254 Food microstructures: microscopy, measurement and modelling Edited by V. J. Morris and K. Groves 255 Handbook of food powders: processes and properties Edited by B. R. Bhandari, N. Bansal, M. Zhang and P. Schuck 256 Functional ingredients from algae for foods and nutraceuticals Edited by H. Domínguez 257 Satiation, satiety and the control of food intake: theory and practice Edited by J. E. Blundell and F. Bellisle 258 Hygiene in food processing: principles and practice Second edition Edited by H. L. M. Lelieveld, J. Holah and D. Napper

© Woodhead Publishing Limited, 2013

1 Automatic process control for the food industry: an introduction Y. Huang,1 United States Department of Agriculture, USA

DOI: 10.1533/9780857095763.1.3 Abstract: In order to ensure food security in food-manufacturing operations, automatic process control is desired. With the operation of automatic process-control systems the deviation of the controlled variables from standards can be consistently monitored, adjusted, and minimized to improve the process operations on a regular basis. Proportionalintegral-derivative (PID) control has been widely used in the food industry. Model-based control has been developed to improve the performance of control systems in the food industry. This chapter overviews the concepts, methods, and systems of automatic process control for the food industry, and projects the future of automatic process control. Key words: food industry, automation, process control, PID, model-based control.

1.1

Introduction

The food industry includes collective businesses and manufacturers that together supply food products for people to consume. Food security in quality and safety are the primary concerns for the food industry. In order to assure the quality and safety of food security, process control is needed to improve food-manufacturing operations. Process control is realized by using the difference between the measured values of the controlled variable(s) and their desired values to regulate the process output to meet performance requirements. Process control can be implemented manually based on the judgment of human operators. Although highly trained human operators are intelligent and able to perceive the deviation of the controlled variable(s) from the standards when problems occur, their judgments may not be 1

Disclaimer. Mention of a trade name, proprietary product, or specific equipment does not constitute a guarantee or warranty by the US Department of Agriculture and does not imply approval of the product to the exclusion of others that may be available.

Published by Woodhead Publishing Limited, 2013

4

Robotics and automation in the food industry

consistent due to fatigue or other unavoidable mental and physical stresses, which may result in inconsistency of food products. With the development of control theory, electronic technology, and computer engineering, automatic process control becomes possible. With the operation of the automatic process-control systems, deviation of the controlled variable(s) from the standards can be consistently monitored, and the difference between the measured and desired values can be consistently used to adjust and improve process operations on a regular basis. In general, automatic process control (called process control hereafter) goes through a procedure as follows: 1. Specify desired values of the controlled variable(s). 2. Measure samples for actual values of the controlled variable(s). 3. Calculate the difference between the actual and desired values of the controlled variable(s). 4. Input the difference to operate the pre-designed controller to adjust the controlled variable(s) to reduce the difference. Design of a controller is important for successful process control. PID control has been widely used in the food industry. To improve the performance of the control systems, model-based control has been developed (Haley and Mulvaney, 1995). This article will overview the concepts, methods, and systems of process control for the food industry, and project the future of process control in that industry.

1.2

Process control systems and structure in the food industry

In a process control system, computers are key to driving and managing the operation of the system. A computer for process control includes hardware, such as the computer itself, peripherals, instrumentation, input–output equipment, and system and application software. Process control engineers, food scientists, and engineers are responsible for developing the application software-based control algorithms. Two types of process control systems exist: open-loop and closed-loop. In openloop systems, the system output is controlled directly only by the input signal, with the output having no effect on the system input. The system cannot compensate for any unexpected conditions in the system output. Closed-loop systems monitor the system output, and feed the output measurement back to the control computer, which continuously minimizes the difference between the measured output and desired output, by adjusting the controller input. Feedback on how the system is actually performing allows the controller to dynamically compensate for disturbances to the system. Compared to closed-loop control systems, open-loop control systems are less commonly used because they are less accurate. With the benefit of feedback, the closed-loop control systems have been widely used in the food industry. The measurement of the system output from food-manufacturing operations is fed back to be compared to the desired value(s) of the variable and to adjust the system output to minimize the error between the measured and desired outputs.

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry

1.3

5

Process control methods in the food industry

In the food industry, various control schemes have been designed and used (Haley and Mulvaney, 1995). Conventional PID controllers have been applied to different processes. However, PID controllers cannot work well consistently, because they are mostly suitable for processes with low-order linear dynamics. In order to improve the performance of process control systems for processes that have high-order non-linear dynamics with time-variant parameters, more advanced control schemes have been designed and applied. Model-based controllers are designed based on the prediction of process models (analytical and empirical) and process control requirements. Artificial neural networks (ANNs) and fuzzy logic (FL) are two of the most common approaches to establishing the relationship between the input and output of the system to estimate the process output instead of measuring it directly. Neuro-fuzzy controllers take the advantages of ANNs and FL to establish the relationship between the input and output of the system to infer the process output that cannot be measured directly. FL and ANNs are two soft computing techniques (Huang et al., 2010). They have been used in food science and technology (Eerikainen et al., 1993). Other soft computing techniques, such as genetic algorithms (GAs) and support vector machines (SVMs), also have the potential to be used in designing process controllers.

1.3.1 Proportional-integral-derivative controller PID controllers have been developed for the classic closed-loop feedback control scheme (Ang et al., 2005). This development is based on the study of the dynamics of second-order linear systems. In order to direct the system to perform as desired a PID controller can be added at the input to form a feedback closed-loop (Fig. 1.1). A PID controller can be expressed by the following equation: u (t ) = K p e (t (t )

t

K i ∫ e ( τ ) dτ K d 0

d e (t ) dt

[1.1]

where u(t) is the output of the PID controller and the input to the process at time instant t, e(t) = ys(t) − y(t) is the difference between the system output y(t) and the

y s(t ) +

e(t)

u (t ) PID controller

–

Process (second-order system)

y (t)

Fig. 1.1 A feedback closed-loop control scheme with PID controller.

Published by Woodhead Publishing Limited, 2013

6

Robotics and automation in the food industry

Δ

Output response of control system

Static value

y s (t )

Static error

δ %=(Δ/y s) x 100% δ % is overshoot Δ = max (y(t ) – y s(t))

y (t )

t

Rising time Regulating time

Fig. 1.2 Typical output response of control system under step input.

desired value ys(t) at time instant t, Kp is the proportional gain constant, Ki is the integral gain constant, and Kd is the derivative gain constant. Figure 1.2 shows the definition of controller tuning parameters in response to a step input. The proportional control mode simply has the effect of reducing the rise time, and reduces but never eliminates the steady-state error. The integral control mode has the effect of eliminating the steady-state error, but it may increase the overshoot and time to make the transient response worse. The derivative control mode has the effect of increasing the stability of the system, reducing the overshoot, and improving the transient response. Typically, PID controllers are used with all the three modes together using Kp, Ki, and Kd to weigh the relative importance of the proportional, integral, and derivative controls. Some applications may only require use of one or two modes to provide the appropriate system control. This is achieved by setting one or two of Kp, Ki, and Kd to zero. A PID controller will be called a PI, PD, P, I, or D controller in the absence of other control actions. PID controllers have been used in various areas for process control, including the food industry. However, the results vary. At times they can perform satisfactorily, even without any tuning, yet on other occasions they can perform poorly, regardless of how the controllers are tuned. In particular, system non-linear characteristics contribute negatively to the performance of a conventional PID controller during transient conditions. These non-linearities can result in increased oscillations, overshoots, and long settling times. To reduce these negative effects, more advanced non-linear control is needed to positively enhance the control performance of the closed-loop system. In order to develop a more advanced controller, a process model is needed to achieve better performance in process control.

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry

7

1.3.2 Advanced process controller PID control is a controller that does not use a process model for controller design. In order to achieve better performance in process control, especially when PID control performs poorly, a process model is useful. This is also called model-based control. In model-based control, process models are required. The models can be analytically based on first principles or empirically based on experimental data. Analytical models are built by considering the particular process as a white box with all necessary information available. In food process control, if all the physical and chemical processes are known in a unit operation, the analytical model can be built for analysis and control of the operation. Ozilgen (1998) provided examples of analytical modeling in food processing, such as transport phenomena modeling and kinetic modeling. However, the physical and chemical information on other processes is unknown. Therefore, it is difficult to establish an analytical model for many processes. In these cases, empirical models can be built using experimental data. Empirical models are black-box models. A black-box process has no prior information available. Without any knowledge of its internal working mechanisms, a black-box process can be studied solely in terms of its input, output and transfer characteristics. System identification provides a mathematical approach to black-box process modeling (Ljung, 1999). In model-based control, the empirical process models can be built online using system identification. The online identified models can be used to tune the parameters of the controllers. Process modeling Generically a manufacturing process can be assumed to be governed by the following discrete-time auto-regressive moving average with exogenous input (ARMAX) relationship: y (k ) = f ( y (k

)

y (k

p) u ( k −

)

u (k − q)

(k − )

(k

r)

)

[1.2] where f is the unknown functional relationship between process input and output, y(k) is the measured process output at instant k, u(k) is the process input at instant k, ε(k) is the noise in the measured output at instant k, which can be assumed as Gaussian white-noise with known variance σε2, p is the order of past output, q is the order of past input, r is the order of past output noise, and Θ is the set of parameters in the functional relationship. In practice, a lot of dynamic processes can be adequately described by the linear ARMAX model as long as the variations in the operating conditions are small enough. However, in many cases, the true relationship between process input and output over the range of interest is non-linear. The linear model could be used to estimate the non-linear process dynamics piecewise by assuming that the variations of the operating conditions are small enough locally. Model [1.2] could be used directly to map between y(k) and u(k) with ε(k). In modeling linear processes, the ARMAX model can model deterministic and stochastic parts of the system independently (Ljung, 1999). However, the method

Published by Woodhead Publishing Limited, 2013

8

Robotics and automation in the food industry

of linear regression cannot be used for model parameters estimation because of the moving average part of the model with ε(k). In this way, the linear ARMAX model can be simplified into the linear autoregressive with exogenous input (ARX) model. Similarly, when modeling non-linear processes, the generic nonlinear ARMAX model as described in Equation [1.2] can be simplified into a non-linear ARX model in many cases: y (k ) = f ( y (k

)

y (k

p) u (k −

)

u (k − q)

)

(k )

[1.3]

This model has been developed for snack-food frying process control (Huang et al., 2001). In this process, the sensors of the product-quality indices, such as product color and moisture, are located at the end of the production line. They are away from the actuators that regulate the factors such as inlet oil temperature and submerged exposure time that have impacts on the product-quality indices. Therefore, the model in [1.3] must be modified for the process to deal with the time lag between the process input and output: y (k ) = f ( y (k

)

y (k

p) u (k − d

)

u (k d − q)

)

(k )

[1.4]

where d is the time lag between process input and output. Model-based control Model-based control basically uses the prediction of the process model to invert the process model to design the controller. There are two types of prediction from the process model, one-step-ahead prediction and multi-step-ahead. One-step-ahead prediction predicts the process output from the process model one step ahead of the current instant k. Multi-step-ahead prediction predicts the process output from the process model l-steps-ahead (l > 1) of the current instant k. From the non-linear ARX model, one-step-ahead prediction is as follows: yˆ ( k +

)=

(

fˆ y ( k ) ,

, y (k

p + ), u (k ),

, u (k − q + ), ˆ

)

[1.5]

In the case of linear modeling, the one-step-ahead prediction is a combination of the current and past y and u. For the design of a non-linear controller, based on ARX one-step-ahead prediction as shown in Equation [1.5], if there is a unique solution for u(k) from the inverse of f +, then the control law can be represented as

(

uˆ ( k ) = g y s ( k ) , yˆ ( k

), y (k ),

, y (k

p + ), u (k − ),

ˆ , u (k − q + ), Θ ,u

) [1.6]

where g is the inverse of f for u(k). For linear models, the inverse is straightforward. However, for non-linear models, the control law is obtained by inverting the process model at each sampling point numerically:

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry y s(k) +

v(k) F

y (k)

u(k) IMC controller

9

Process

– ∧ y (k)

+

Process model –

Fig. 1.3

IMC control loop where F is a filter to improve controller robustness.

uˆ ( k ) = uˆ

1

( k ) + Δ n −1uˆ ( k )

[1.7]

where ûn(t) is the value of û(k) after the nth iteration, and Δn−1u+(k) is the updating increment of ûn(t) at the (n − 1)th iteration, which is determined by the selected numerical method and control law. This control approach constructs a process model-based internal model control (IMC) scheme (Fig. 1.3). IMC is a control method that is used to create a modelbased controller by combining the process model with the controller system performance specification. IMC was first proposed for linear systems (Garcia and Morari, 1982). It later was extended to general non-linear systems (Economou and Morari, 1986). IMC has the properties of dual stability, perfect control and zero offset (Garcia and Morari, 1982). In IMC, a function can be defined to be used to compute the control action for a unique solution of u(k): U ( u ( k ))

v ( k ) yˆ ( k +

)

[1.8]

where v(k) is the tracking signal of the process model prediction. In the IMC loop, in the frequency domain, the relationship between v and e can be derived as:

(

v ( z ) = F ( z ) ys ( z ) − y ( z ) + y ( z ) = F ( z ) ( e ( z ) + y ( z ))

)

[1.9]

If the process model is perfect and u can be solved from the process model in the form of one-step-ahead prediction, then v ( z ) = F ( z ) yˆ ( z ) z

[1.10]

The combination of Equations [1.9] and [1.10] produces the relationship as

Published by Woodhead Publishing Limited, 2013

10 Robotics and automation in the food industry v (z) =

F (z) 1 − z −1

e (z)

[1.11]

Because e is the error signal of the classic feedback control loop, and v is the error signal of the IMC control loop, Equation [1.11] illustrates that an IMC controller is equivalent to a feedback controller on the basis of the process model prediction. The other way to design a model-based controller is to optimize online a designated objective function based on multi-step-ahead prediction from the process model to achieve the control action, û(k), at each time instant. û(k + l − 1) (l = 1, 2, ... , L2) constitutes the solution of the optimization of the following objective function: J ( u ( k ))

L1

∑ ⎡⎣ ys ( k l ) − yˆ ( k + l )⎤⎦ l =1

2

L2

∑λ l =1

1

⎡⎣uˆ ( k l −

) − uˆ ( k + l − )⎤⎦

2

[1.12]

where L1 and L2 are the horizons over which the tracking error and control increments are considered, and λl is the weight of the lth control increment. In predictive control (PDC), the optimization may be subject to some constraints. Similar to IMC, for the non-linear process model, the control law in each time instant needs to be calculated numerically. This control approach constructs a process model-based PDC scheme (Fig. 1.4). The basic idea of PDC originated from dynamic matrix control (DMC) (Cutler and Ramaker, 1979). It was subsequently extended to generalized predictive control (GPC) (Clarke et al., 1987) based on the linear ARMAX model. It has been proven that PDC has desirable stability properties for non-linear systems (Keerthi and Gilbert, 1986; Mayne and Michalska, 1990). Neural network-based control For complex food processes, ANNs could provide process model predictions for controller design. A typical process for ANN application is the snack-food frying process. This process is one of the unit operations in the snack-food manufacturing industry. The manufacturing of snack foods involves a number of unit operations in snack-food manufacturing plants, such as extrusion, frying, baking, and drying. Frying is a process in which a snack is cooked by floating or being immersed in hot oil. There are two types of snack-food frying: batch and continuous. Batch frying is typically used in small-scale operations, such as restaurants. Continuous frying is used at large-scale operations such as snack-food plants. In the continuous frying process, continuous input of extruded snack-food material occurs at one end of the fryer, pushed through by a submerger with oil flow, and then extracted at the other end of the fryer. The primary concern in frying is the low-cost production of nutritious snacks at consistent quality and with minimum waste. In order to ensure consistency in snack product consistency, automatic control is desired. The snack-food frying process is a complex process. It has complicated interactions between the

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry u(k)

11

y (k) Process

∧ y (k)

Process model prediction

+

y s(k)

Process optimization

∧ u (k)

–

Fig. 1.4

PDC control loop.

product-quality indication variables, such as snack moisture, color, and oil content, and those factors that can be changed independently to affect the product-quality indication variables, such as frying inlet oil temperature and snack-material residence time in the fryer, which is related to the speeds of the submerger conveyor and the takeout conveyor (Fig. 1.5).To control this complex process, it is necessary to model it first. Due to the complexity, an analytical model is difficult to develop. A multi-input-multi-output (MIMO) model can be built using the method of system identification. In the snack-food frying process, the productquality index sensors are located at the end of the production line. These sensors are at a distance from the actuators regulating inlet oil temperature and conveyorspeed. Therefore, there exist significant time-lags between the product-quality indices (process output) and their impact factors (process input), which result in global non-linearity of process output in response to process input. A multivariate non-linear model is needed to handle the non-linearity caused by the time-lags. ANNs have the ability to map the generic non-linear relationship between the process input and output without any prior assumptions on the model form. ANNs are chosen to provide model predictions for controller design for controlling the snack-food frying process. Two kinds of ANNs are useful for the snack-food frying process modeling and prediction. One is the feed-forward network (Fig. 1.6). This network maps the process function relationship and provides one-step-ahead prediction for IMC controller design: yˆ ( k +

) = fˆ ( y ( k ) ,

, y (k

p + ), u (k − d ),

, u (k d − q

Published by Woodhead Publishing Limited, 2013

) ,W )

[1.13]

12 Robotics and automation in the food industry Damper Color and moisture sensors

Material inlet belt

Takeout conveyor

Ambient air cooler

Submerger conveyor Cooling conveyor Oil level Weightbelt Oil inlets

Oil outlet

Recirculation pump and heat exchanger

Fig. 1.5

Schematic diagram of a continuous snack-food frying process. (Source: Huang et al., 2001, p. 9.)

y(k – 1)

y(k – 2) y(k – 3) ∧

y(k – 4)

y(k)

u(k – 19) u(k – 20)

Output layer Hidden layer

u(k – 21) Input layer

Fig. 1.6 Architecture of a feed-forward ANN with one hidden layer to map the relationship as Equation [1.13].

where W is the set of network weights and bias terms. Figure 1.7 illustrates the IMC control operation based on the one-step-ahead prediction from the feed-forward ANN process model. In the control loop, the control action is computed with the inverse of the function U(u(k)) = v(k) − ŷ

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry y s(t)

+

∧

u (t)

v(t)

ANN process model inverse

F –

13

y(t) Process

– ∧

ANN process model prediction

y(t+d +1)

+ ∧

y (t)

Z -d-1 –

Fig. 1.7

IMC control operation based on the one-step-ahead prediction from the feed-forward ANN process model.

(k + d + 1), an extended version of Equation [1.8], while the future process outputs need to be estimated through the feed-forward ANN model. The inverse of the ANN process model with one-step-ahead prediction is the primary step in setting up the IMC control loop. This inverse computes the control actions iteratively at each time instance following Equation [1.7]. In order to compute the inverse, an objective function based on U(u(t)) can be defined to be minimized: J ( u ( k ))

U 2 ( u ( k ))

[1.14]

Using the gradient descent method, the control actions can be updated as: Δ n −1uˆ ( k ) = − γ

∂ J ( u ( k )) ∂u ( k )

[1.15] u ( k ) = uˆ n−1 (k )

where γ is a tuning factor, which can be tuned in the control loop. This updating equation guarantees that each update of u(k) makes the objective function J move in the direction of the largest gradient decrease. Figure 1.8 shows the IMC responses of the snack-food frying process output based on controller tuning with γ for gradient descent updating to optimize J(u(k)). The controller tuning was based on the error functions of ISE (integral of square error), IAE (integral of absolute error) and ITAE (integral of absolute error multiplied by time). γ = 0.005 was used for optimal IAE and ITAE and γ = 0.006 was used for optimal ISE. The output responses illustrate that γ = 0.005 provides smooth control compared to γ = 0.006. The second technique is the feedback recurrent network (Fig. 1.9). This network maps the process function relationship as follows: yˆ ( k +

) = fˆ ( yˆ ( k ) ,

, yˆ ( k

p + ), u (k − d ),

, u (k d − q

Published by Woodhead Publishing Limited, 2013

) ,W )

[1.16]

14 Robotics and automation in the food industry

57.6

2.21

57.5

2.2

57.4

2.19 Moisture (%)

(b) 2.22

Color (—)

(a) 57.7

57.3 57.2 0.005

57.1

2.18 2.17

0.005

2.16 2.15

57.0 56.9

2.14

0.006

56.8 0

40 20 Time (seconds)

60

2.13

0.006 0

40 20 Time (seconds)

60

Fig. 1.8 (a) Response of color to IMC controller tuning with γ = 0.005 and 0.006 and set-point of 57. (b) Response of moisture content to IMC controller tuning with γ = 0.005 and 0.006 and set-point of 2.15%. Sampling every 5 s. (Source: Huang et al., 1998, p. 1523.)

Z–1

∧

y(k – 1) Z–1

∧

y(k – 2)

∧ Z–1 y(k – 3)

∧

Z–1 y(k – 4)

∧

y(k) u(k – 19) u(k – 20)

Output layer

u(k – 21)

Hidden layer Input layer

Fig. 1.9 Architecture of a feedback recurrent ANN with one hidden layer to map the relationship as Equation [1.16].

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry

15

This feedback network provides multi-step-ahead prediction for PDC controller design as the objective function described below by considering process input/ output time-lags using Equation [1.12]: J ( u ( k ))

L1

L2

∑ ⎡⎣ y ( k l ) − yˆ ( k + l )⎤⎦ ∑ λ 2

s

l =1

l =1

1

⎡⎣uˆ ( k l − d

)

uˆ ( k l − d − 2 )⎤⎦

2

[1.17] With this objective function, numerical optimization is performed online in the PDC loop based on multi-step-ahead prediction from the feedback network for optimal control actions. If there are constraints on process inputs or outputs, these constraints can be incorporated into the optimization by: J ( u ( k ))

m

J ( u ( k )) + ∑ ri ψ i2 ( u ( k ))

[1.18]

i =1

where m is number of constraints, ri is the ith penalty constant, and ψi(k) is the ith process constraints. The second term in the objective function is a penalty function. It is considered as a certain cost needed in the objective function due to constraint violation in the control action. Neuro-fuzzy control FL-based control has emerged for complex and/or ill-defined process control. Control engineers have been studying a self-learning fuzzy controller since Procyk and Mamdani (1979) developed it. In the food processing industry, FL has been used for the process controller design of flat-bread baking (Eerikainen et al., 1986), baker’s yeast-fed batch bioprocesses (Mahjoub et al., 1994), smokehouse batch-cooking process (Davidson and Smith, 1995), a continuous crossflow grain dryer (Zhang and Litchfield, 1993), whole-milk powder processing (Koc et al., 2002), food-frying process (Rywotycki, 2003), and a processing plant for making sweets (Venayagamoorthy et al., 2003). Process-control practice has shown that a human operator’s intervention is sometimes necessary in order to reduce any irregularities, such as large overshoots, that may occur in control operation. FL is able to transcribe the human experience into a set of linguistically defined rules that are used to implement the actions of the controller. Due to the linguistic characteristics, fuzzy process control is more tolerant than conventional PID control. FL offers a system model based on membership functions and a rule base, but requires an explicit stating of the IF/THEN rules. For complex process control, a predictive model still is needed to characterize the non-linearity in process dynamics. This results in the integration of ANNs with FL for process control, to give socalled neuro-fuzzy control (Khan, 1993; Nauck et al., 1993; Lin and Song, 1994). In food processing neuro-fuzzy control has been developed and applied. Linko et al. (1992) constructed a feed-forward ANN model for simulation and fuzzy control for a flat-bread extrusion process. Kim and Cho (1997) applied neural network modeling in FL control simulation for a bread- baking process. An ANN Published by Woodhead Publishing Limited, 2013

16 Robotics and automation in the food industry predictive model was built to model the non-linearity in the oven system caused by a long time lag in response to the control signal. For snack-food frying process control, Choi et al. (1996) designed a self-learning fuzzy predictive controller with a feedback recurrent network for multi-step-ahead prediction. The design of the neuro-fuzzy controller is composed of a comparator, a fuzzy controller, and an ANN estimator.

1.4

Future trends

In the next decade with further development of computer power, electronics, network technology, and computing technology, process-control technology will advance to provide stable, robust, and reliable controllers for various food processes using established control theory and methods.

1.4.1 Adaptive-network-based fuzzy inference system Neuro-fuzzy controllers will continue to be developed for optimal-food process control. In this aspect, food-control engineers should pay attention to ANFIS (AdaptiveNetwork-based Fuzzy Inference System). Jang (1993) introduced the framework of ANFIS. Jang and Sun (1995) reviewed fundamental and advanced developments in neuro-fuzzy systems for modeling and control. They introduced design methods for ANFIS in modeling and control applications. ANFIS is a fuzzy inference system implemented in the framework of adaptive neural networks. By using a hybrid learning procedure, ANFIS can construct an input–output mapping based on both human knowledge (in the form of fuzzy ‘If–Then’ rules) and stipulated input–output data pairs. It has been developed and used to solve problems in agricultural and biological engineering (Huang, 2009; Huang, et al., 2010). However, no application of ANFIS for food process control has yet been published.

1.4.2 Wireless sensors and sensor networks Wirelessly networked and embedded sensors have been rapidly developed in recent years (Xia, 2009). Wireless sensor networks (WSN) deploy a large number of embedded devices, each equipped with one or more sensors, a processor, memory storage, and a radio. These devices are typically low-cost, low-power and small with limited sensing, data processing and wireless communication capabilities. WSNs with self-organizing, self-configuring, self-diagnosing, and selfhealing abilities could be applied to solve problems that traditional technologies could not. WSN is a promising technology that can provide new economic opportunities for the agriculture and food industries (Wang et al., 2006). In process control, the premise is that these wirelessly networked sensors can be deployed in manufacturing plants, and will autonomously report various online measurements in real time to a centralized decision support system (DSS). The DSS will then interpret the data and provide the suggested actions and control strategies.

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry

17

1.4.3 Internet-enabled systems The author developed a Web-based integrated system for data manipulation and management for irradiated particle transport simulation in foods (Huang et al., 2008). For the simulation, CT (Computerized Tomography) based 3D geometric models of foods were built to provide input data to a general Monte Carlo N-Particle (MCNP) code (Los Alamos National Laboratory, Los Alamos, NM, USA). The Web-based interface provides the online capability to formulate input data for the MCNP and to visualize and analyze output data generated by it. In addition, a database storing data such as food nutrition composition and quality was integrated to support food irradiation research. Online services are provided for D-value look-up, nutrition facts, and quality evaluation of foods. This system enables flexible and reliable operations through the Internet to food irradiation engineers and researchers, by providing data preparation, analysis, and interpretation during the process of irradiation simulation of complex shaped foods. This Web-enabled technique can be adopted for food process control. Through the Internet process, modeling, controller design, and process operations can be performed by the Web-based system. Additionally, in the WSN setting, the nodes in the network can communicate with the Internet as a gateway to provide the capability of communicating with other computers via other networks.

1.4.4 Soft computing-based systems As discussed above, FLs and ANNs are two major soft computing techniques that have been used in food process control. GAs, as another soft computing technique, have been used for studies of food quality and safety (Cogdill et al., 2004; Pearson and Wicklow, 2006; Li et al., 2008). However, no publication has been seen on the application of GAs for food process control (Huang et al., 2010). GAs are an optimization and heuristic search technique that use methods inspired by evolutionary biology, such as inheritance, mutation, selection, and crossover (also called recombination). GAs work simultaneously on a set (population) of potential solutions (individuals) to the problem. The algorithms start with a set of solutions (representing chromosomes) called a subpopulation. The fitness with which solutions meet some performance criterion is evaluated and used to select ‘surviving’ individuals that will ‘reproduce’ a new, better subpopulation. Then, the individuals will undergo alterations similar to natural genetic mutation and crossover. The selection scheme makes progress toward high performance solutions. A careful selection of GA structure and parameters can ensure a good chance of reaching the globally optimal solution after a reasonable number of iterations. GAs are computationally simple, yet powerful enough to provide a robust search for difficult combinational search problems in complex spaces, without being stuck in local extremes (Goldberg, 1989). Therefore, GAs are powerful alternative tools to traditional optimization methods and will find applications where optimization is needed in food process control. SVMs have been introduced as a new set of supervised generalized linear classifiers (Vapnik, 1995; Burges, 1998; Cristianini and Taylor, 2000). SVMs are

Published by Woodhead Publishing Limited, 2013

18 Robotics and automation in the food industry closely related to ANNs. In fact, a SVM model using sigmoid kernel function is equivalent to a two-layer perceptron neural network. Using a kernel function, SVMs are alternative training methods for polynomial, radial basis function, and multilayer perceptron classifiers, in which the weights of the network are found by solving a quadratic programming problem with linear constraints, rather than by solving a non-convex, unconstrained minimization problem as in standard ANN training. SVMs have often higher classification accuracies than multilayer perceptron ANNs. SVMs have attracted great interest recently in agricultural and food engineering (Huang et al., 2010). In food process control, SVMs will produce comparable results to ANNs, and even improve upon them.

1.5

References

Ang, K.H., chong, G.C.Y. and Li, Y., 2005. PID control system analysis, design, and technology. IEEE Transactions on Control Systems Technology, 13(4), pp. 559–576. Burges, C.J.C., 1998. A tutorial on support vector machines for pattern recognition. Data Mining Knowledge Discovery, 2, pp. 121–167. Choi, Y. S., Whittaker, A.D. and Bullock, D.C., 1996. Predictive neuro-fuzzy controller for multivariable process control. Transactions of the ASAE, 39(4), pp. 1535–1541. Clarke, D.W., Mohtadi, C. and Tuffs, P.S., 1987. Generalized predictive control-Part I. The basic algorithm. Automatica, 23(2), pp. 137–148. Cogdill, R.P., Hurburgh, C.R. and Rippke, Jr., G.R., 2004. Single-kernel maize analysis by near-infrared hyperspectral imaging. Transactions of the ASAE, 47(1), pp. 311–320. Cristianini, N. and Taylor, J.S., 2000. An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods. New York, NY: Cambridge University Press. Cutler, C. and Ramaker, B., 1979. Dynamic matrix control – a computer control algorithm. 86th National Meeting of the American Institute of Chemical Engineering. Houston, TX, April 1979. Davidson, V.J. and Smith, K.D., 1995. A fuzzy controller for batch cooking process. Journal of Food Engineering, 24(1), pp. 15–24. Economou, C.G. and Morari, M., 1986. Internal model control. 5. Extension to nonlinear systems. Industrial & Engineering Chemistry Process Design and Development, 25, pp. 403–411. Eerikainen, T., Linko, S. and Linko, P., 1986. The potential of fuzzy logic in optimization and control: Fuzzy reasoning in extrusion cooker control. In: M. Renard and J. J. Bimbenet, eds. Automatic Control and Optimization of Food Processes. London: Elsevier Applied Science, pp. 183–200. Eerikainen, T., Linko, P., Linko, S., Siimes, T. and Zhu, Y. H., 1993. Fuzzy logic and neural network applications in food science and technology. Trends in Food Science & Technology, 4, pp. 237–242. Garcia, C.E. and Morari, M., 1982. Internal Model Control – 1. A unifying review and some new results. Industrial & Engineering Chemistry Process Design and Development, 21, pp. 308–323. Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization, and Machine Learning. Reading, MA: Addison-Wesley. Haley, T.A. and Mulvaney, S.J., 1995. Advanced process control techniques for the food industry. Trends in Food Science & Technology, 6, pp. 103–110. Huang, Y., 2009. Advances in artificial neural networks – methodological development and application. Algorithms, 2(3), pp. 973–1007.

Published by Woodhead Publishing Limited, 2013

Automatic process control for the food industry

19

Huang, Y., Kim, J., Moreira, R.G. and Castell-Perez, M.E., 2008. A Web-based information system for MCNP simulation of irradiation of complex shaped foods. Applied Engineering in Agriculture, 24(2), pp. 227–231. Huang, Y., Lan, Y., Thomson, S.J., Fang, A., Hoffmann, W.C. and Lacey, R.E., 2010. Development of soft computing and application in agricultural and biological engineering. Computers and Electronics in Agriculture, 71, pp. 107–127. Huang, Y., Whittaker, A.D. and Lacey, R.E., 1998. Internal model control for a continuous snack food frying process using neural networks. Transactions of the ASAE, 41(5), pp. 1519–1525. Huang, Y., Whittaker, A.D. and Lacey, R.E. 2001. Automation for Food Engineering: Food Quality Quantization and Process Control. Boca Raton, FL: CRC Press LLC. Jang, R.J.S., 1993. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man and Cybernetics, 23, pp. 665–685. Jang, R.J.S. and Sun, C.T., 1995. Neuro–fuzzy modelling and control. Proceedings of the IEEE, 83(3), pp. 78–406. Keerthi, S.S. and Gilbert, E.G., 1986. Moving-horizon approximations for a general class of optimal nonlinear infinite-horizon discrete-time systems. Proceedings of the 20th Annual Conference on Information Sciences and Systems, Princeton, NJ: University Press, pp. 301–306. Khan, E., 1993. An elegant combination of fuzzy logic and neural nets. Proceedings of Fuzzy Logic’93, pp. A223-1-A223-7. Kim, S., and Cho, S.I., 1997. Neural network modeling and fuzzy control simulation for bread–baking process. Transactions of the ASAE, 40(3), pp. 671–676. Koc, A.B., Heinemann, P.H., Ziegler, G.R. and Roush, W. B., 2002. Fuzzy logic control of whole milk power processing. Transactions of the ASAE, 45(1), pp. 153–163. Li, C., Heinemann, P.H. and Reed, P.M., 2008. Genetic algorithms (GAs) and evolutionary strategy to optimize electronic nose sensor selection. Transactions of the ASABE, 51(1), pp. 321–330. Lin, J. and Song, S., 1994. A novel fuzzy neural network for the control of complex systems. Proceedings of the 1994 IEEE International Conference on Neural Networks, pp. 1668–1673. Linko, P., Zhu, Y.H. and Linko, S., 1992. Application of neural network modeling in fuzzy extrusion control. Food and Bioproducts Processing: Transactions of the Institution of Chemical Engineers., 70, pp. 131–137. Ljung, L., 1999. System Identification: Theory for the User, 2nd ed. Upper Saddle River, NJ: Prentice Hall. Mahjoub, M., Mostrati, R., Lamotte, M., Fonteix, C. and Marc, I., 1994. Fuzzy control of baker’s yeast fed–batch bioprocess: A robustness study. Food Research International, 27(2), pp. 145–153. Mayne, D.Q. and Michalska, H., 1990. Receding horizon control of nonlinear systems. IEEE Transaction on Automatic Control, 35, pp. 814–824. Nauck, D., Klawonn, F. and Kruse, R., 1993. Combining neural networks and fuzzy controllers. In: E. P. Klement and W. Slany, eds. Fuzzy Logic in Artificial Intelligence (FLAI93). Berlin: Springer-Verlag, pp. 35–46. Ozilgen, M., 1998. Food Process Model and Control. Amsterdam: Gordon and Breach Science Publishers. Pearson, T.C. and Wicklow, D.T., 2006. Detection of corn kernels infected by fungi. Transactions of the ASABE, 49(4), pp. 1235–1245. Procyk, T.J. and Mamdani, E.H., 1979. A linguistic self-organizing process controller. Automatica, 15, pp. 15–30. Rywotycki, R., 2003. Food frying process control system. Journal of Food Engineering, 59, pp. 339–342. Vapnik, V., 1995. The Nature of Statistical Learning Theory. Berlin: Springer-Verlag.

Published by Woodhead Publishing Limited, 2013

20 Robotics and automation in the food industry Venayagamoorthy, G.K., Naidoo, D. and Govender, P., 2003. An industrial food processing plant automation using a hybrid of PI and fuzzy logic control. Proceedings of the IEEE International Conference on Fuzzy Systems, pp. 1059–1062. Wang, N., Zhang, N. and Wang, M., 2006. Wireless sensors in agriculture and food industry – recent development and future perspective. Computers and Electronics in Agriculture, 50, pp. 1–14. Xia, F., 2009. Wireless sensor technologies and applications. Sensors, 9(11), pp. 8824–8830. Zhang, Q. and Litchfield, J.B., 1993. Fuzzy logic control for a continuous crossflow grain dryer. Journal of Food Process Engineering, 16(1), pp. 59–77.

Published by Woodhead Publishing Limited, 2013

2 Robotics in the food industry: an introduction J. O. Gray, The University of Manchester, UK and S. T. Davis, University of Salford, UK

DOI: 10.1533/9780857095763.1.21 Abstract: This chapter briefly reviews the European food and beverage market and outlines current trends including customer and regulatory demands, commodity costs, demographic and environmental issues, which will shape its future development. Current manufacturing procedures are discussed together with the drivers for change with emphasis on the need to adopt state-of-the-art engineering technology. The challenges and opportunities presented by the adoption of automation are emphasised and some food-factory-of-the-future concepts are outlined. Keywords: current market, drivers for change, technical challenges, the role of autommation, factory of the future.

2.1

Introduction

The use of robotics has transformed manufacturing in almost every industrial sector, and by increasing efficiency and enforcing consistency has resulted in the provision of the vast range of reliable products at affordable prices that underpins all modern economies. Until recently, food manufacturing has been an exception to this trend and many reasons can be cited for this including the idea that because of their natural variability in consistency and shade they are not amenable to automated processing procedures. However, advances in technology and demographic and market forces are now rapidly changing the European landscape in the food manufacturing sector and an outline of their factors and the opportunities created is briefly reviewed below. 2.1.1 The industry The food manufacturing industry is one of the largest, if not the largest, manufacturing sectors in most European countries and is a vital strategic element in all national economies. Overall, food manufacturing represents some 13% of all manufacturing in the European Union, contributing about €900 billion to the

© Woodhead Publishing Limited, 2013

22 Robotics and automation in the food industry economy and engaging some 4 million employees in its activities. It is a diverse and somewhat fragmented industry involving about 300 000 companies, the vast majority of which can be classified as small-to-medium sized enterprises (SME). Manufacturing operations are mainly manual in nature, particularly in the SME sector and, while islands of excellence in state-of-the-art automation are increasingly common, the technical infrastructure of the sector is generally weak and expenditure on engineering research and development is lower than in most other manufacturing sectors. In this chapter, we will briefly review the current European market (which is a typically diverse market) and how it influences manufacturing procedures. Drivers for change will be outlined, together with the economic and technical challenges they present, and some possible future manufacturing concepts will be introduced that may help to address some of the socio-economic market demands envisaged for the next decade. 2.1.2 The market Both the industry and the retail market are very diverse, but as shown in Fig. 2.1 market share in Europe tends to be focussed on a comparatively small set of major retailers and it is inevitable that they have a significant influence on the market shape, product range and pricing philosophies. There are however some overarching trends, including:

90 80

% of market

70 60 50 40 30 20 10

1st

2nd

UK

Sweden

Spain

Portugal

Poland

Netherlands

Italy

Ireland

Hungary

Greece

Germany

France

Finland

Denmark

Czech Rep.

Belgium

Austria

0

3rd

Fig. 2.1 European retail market share: Current market share of the three largest retailers in various member states. (Source: CAA calculations based on data provided by CIAA members.)

© Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 23 • a downward pressure on costs matched by an upward pressure on commodity prices; • an increasing capacity in non-European countries that directly compete with European products, through use of global logistics and economic services; • a demand for a greater variety of products with shorter run times, which inserts volatility into the market; • a requirement for a faster response in the supply chain; • more fresh food with a concomitant shorter time to market; • an increasing public awareness of food content and hygienic standards and thus a requirement for manufacturers to meet the highest Hazard Analysis and Critical Control Points (HACCP) Standards.* The market will also be influenced by a gradual change in European demography and the desire for a ‘green’ market image, which will reflect on manufacturing procedures, packaging and logistics to emphasise the minimising of waste and energy. In an increasingly competitive market retailers will require a consistency of product quality at lowest cost, high shelf life, and a flexible, rapid response to customer demand, to maintain and develop market share.

2.2

Current manufacturing procedures

The food manufacturing industry has its origins in the extended kitchen concept where there was a focus on food design and preparation, with the factory being an amplified version of domestic procedures of manual preparation and supervision. As production increased to meet demand the larger companies saw the benefit of using automation, particularly in the end-of-line packaging and palletising. The availability of highly effective machines, such as the Delta class of robot, has allowed automation to move upstream and undertake rapid pick-and-place operations with food products on the production lines. Numerous impressive systems are currently installed, generally focussed on the high-volume, long-life, single-product lines. Smaller companies, which constitute over 90% of the European food manufacturing base, have been slow to emulate the larger corporations and there are a number of reasons why this might be so: • Until recently there has been an adequate supply of available labour that could be deployed on a contract basis and that could quickly be deployed to meet rapid changes in demand. Manipulative tasks are generally simple repetitive operations, so training requirements and associated costs were correspondingly minimal. • The marketplace is particularly volatile in nature with short term orders being the norm rather than long-term contracts. This has discouraged capital investment in automation, which is often wrongly considered inflexible in *

This is a management system in which food safety is addressed through the analysis and control of biological, chemical and physical hazards from raw material production, procurement and handling, distribution and consumption of the finished product.

© Woodhead Publishing Limited, 2013

24 Robotics and automation in the food industry operation and thus unsuitable for dealing with short runs of products, which are increasingly a feature of the market pattern. • It was widely thought that automation was quite unsuitable for the assembly and placing of soft, variable-dimensional, fragile, slippery/sticky natural products, a view supported by the ubiquitous images of robot operations on large-scale car production lines. • The professional engineering expertise available within existing SME management structures is generally weak in this sector. This not only impacts on the uptake of the technology at line level, but also deprives companies of informed internal guidance on capital investment decisions at Board level, and thus increases the perceived risk of adopting advanced manufacturing technology. The above considerations have influenced the evolution of the industry, particularly in the predominant SME sector, with the result that there is still a strong reliance on manual manufacturing procedures. Although the impression given is sometimes that of a nineteenth century rather than twenty-first century enterprise, the dynamic and vibrant management structure that characterises the sector has so far ensured that the demands of the market place have been essentially satisfied. However over the last decade or so, a number of market and other trends have emerged that seriously challenge the traditional manual intensive manufacturing approach. • For a number of demographic and other reasons, the availability of suitable low-cost casual labour within the European Union has declined, which directly impacts on the flexibility of the labour-intensive approach. • Existing and imminent EC legislation on employment law and health and safety directives are putting upward pressure on labour costs and rendering alternative solutions more attractive. • The increasing demand for assured hygienic procedures, consistency of product quality and line security, are providing persuasive arguments for separating human operators from the production process. • Increased commodity costs, the need for reductions in energy and water usage, potential zero access to landfill sites, and overall carbon footprint issues, are amongst a range of factors now prompting a radical review of manufacturing processes within the sector. • Modern requirements for product traceability and the obvious commercial advantages of using a completely integrated manufacturing, distribution, automated warehousing and marketing systems are simply incompatible with traditional manual manufacturing procedures. Thus although some of the drivers for change are set by legislative and demographic factors, the key drivers are being set by the market itself. Customer satisfaction and desired repeat sales can only be guaranteed by products of consistent high quality, which are absolutely safe to consume, are attractively packaged, and lie within acceptable cost boundaries. It is beneficial if the retailer can claim that

© Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 25 his products are sourced from manufacturers who meet the highest standards in hygienic production, and embrace a green agenda, implying high efficiency, low environmental impact, ethical sourcing of material, and use skilled professional personnel. Retailers now require a totally integrated supply chain with traceability through manufacturing, logistics, storage, shelf placement to point of sale and beyond. This provides a powerful responsive tool that can strip costs out of the process, and can only be effective if embedded within an advanced manufacturing environment.

2.3

Automation in the food sector

The above drivers have led to a gradual uptake of automation within the sector, particularly within larger manufacturers. Currently (British Robot Association, Annual Report 2011) the sector represents the largest market for robotics within the UK, which is perhaps indicative of a general trend within the industry. Because of the diversity and nature of the product line, a number of manufacturing strategies have been adopted. In high-volume, long-life, branded single-product lines, a so-called hard automation solution is appropriate, being implemented by robots or combinations of simpler electro-mechanical or other devices. If the product is regular in shape and well located on the line, then simple electro-mechanical solutions will suffice. Poor line localisation will usually involve visual-servoing robotic procedures, particularly if the product is also irregular in shape. It has been demonstrated that a combination of robotics and electro-mechanical processes can address the automation of most food products of either regular or irregular shape. Initially in the sector, standard industrial robots were used for end-of-line tasks including packaging and palletising. These were gradually moved up the line to undertake simple product placement tasks and demonstrated their abilities in a range of applications one of which is shown in Fig. 2.2. However, there was a demand for faster and more agile machines suitable for rapid pick-and-place operations on individual food products. This challenge was met by the introduction of the Delta family of robots now available from a range of manufacturers. These robots are optimised for fast operation (typically 100– 120 picks per minute) with lightweight payloads (typically 1–2 kg) and have been particularly successful in processing food products. A typical application involving placing products is shown in Fig. 2.3. Installations can be cascaded to generate a workspace devoid of human operators as shown in Fig. 2.4, a configuration that is indicative of future manufacturing trends. A key advantage of robotics over ‘hard-wired’ electro-mechanical systems is the ability to reprogram their operations for different tasks using simple, easily implemented procedures, thus meeting the need for a flexible production capability within the industry. Another very positive trend is the constant reduction in the cost of robots versus the inevitable rise in labour costs, over the last decade (as shown in Fig. 2.5), which renders this approach increasingly attractive.

© Woodhead Publishing Limited, 2013

26 Robotics and automation in the food industry

Fig. 2.2

Fig. 2.3

Batch processing on a bakery assembly line.

High-speed processing on a high-volume pancake line.

• In the period from 1990 to 2007, the average hourly wage of a typical worker increased by 105%. • In the same period, the real price of robots decreased by 48%.

2.4

Specifications for a food sector robot

Robotic systems have been developed over decades for successful application in a vast range of industrial sectors and they underpin the advances in worldwide

© Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 27

Fig. 2.4 A series of Delta robots performing ‘pick-and-place’ operations in the large-scale production of brittle pretzel products. 220 200 Labour compensation

Index 1990 = 100

180 160 140 120 100 Robot prices

80 60 40

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007

Fig. 2.5 Trends in capital costs of robots versus European labour costs. (Source: International Federation of Robotics, Annual Review World Robotics Magazine 2008.)

economic development. A variety of robotic systems, developed for such sectors, are readily available from international companies, and are often offered as manufacturing solutions for the food industry. However, as indicated, the food industry is a special market and it is useful to outline some guidelines for the design of equipment for the sector. The use of robotics in the food industry was reviewed by Wallin (1995, 1997) and Purnell (1998) and more recently by Moreno-Massey et al. (2010), who outlined a series

© Woodhead Publishing Limited, 2013

28 Robotics and automation in the food industry of desirable design features that should be considered in the development of a food-grade robot. Any design should meet the guidelines for food-handling equipment given, for example by and Lelieveld et al. (2003, 2007) and be ideally of stainless-steel construction with ingress protection rating to IP67, and all parts visible and accessible for inspection and manual cleaning. The standards of hygienic design required will depend on the application, being higher, for example, in processing poultry and fish products than in processing dry-food products, such as biscuits or bread. A general purpose robot, which could potentially be used to process any type of product, should meet the highest standards of hygienic design. For use on standard-width, high-volume production lines there is a requirement for relatively fast transfer speeds of between 80 and 120 picks per minute with a payload of 1–2 kg in an operating space of about 1 × 2 m. Safety and power factors suggest a lightweight design deployable, if necessary, with minimal guarding in close proximity to human operators in a mixed-mode production scenario. Ideally it should occupy a minimal volume, be moveable without significant recalibration, and placement should not impinge on the integrity of the factory floor. The technical skills required for efficient operation should match the application skill base of the industry and since basic IT skills are generally limited in the sector, programming and reprogramming tasks should be made simple and use intuitive, human-based methods that minimise technical stress. The robotics industry has made significant progress in addressing this issue by developing graphical and teach-by-showing procedures that minimise the IT skills required to operate. Future developments are possible using icon based programming touch screens, gesture and speech recognition, or fully instrumented gloves that directly control the robot’s operation. Cost will always be a factor influencing the take up of automation, particularly in the SME sector. As has been shown above, the real cost of robots has fallen over the years, and the trend will no doubt continue to the advantage of the manufacturing sector. Another positive factor that can be exploited is the reduced accuracy requirements in handling food products, as opposed to the precision demanded, for example, in the automobile industry. This design relaxation can be translated into a requirement for less stiff and thus lighter arms, with smaller drive mechanisms, needing less inertia and possibly resulting in lower costs. 2.4.1 Gripper technology Food products by their nature tend to be irregular in shape, sticky/slippery, soft or fragile. It was generally believed that safe and efficient working could only be provided by the intervention of the human hand, and such a perception has remained unchallenged until recent years. Clearly any automated procedure will require the deployment of a suitable end effecter or gripper and, given the variability of food products, the design will generally be unique to the target product but must meet generic hygienic requirements. Gripper design is thus more application focussed in this sector than elsewhere in manufacturing. Given this application-oriented

© Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 29 approach a number of automated solutions have been developed in recent years that have deployed grippers in processes that hitherto have been regarded as the exclusive domain of human operators. The success here, particularly with difficult food products, appears to indicate that gripper design should pose no serious problem in the robotic applications to food products. A few examples of these solutions are outlined briefly below (Chua et al., 2003). An early problem posed in the automation of a sandwich production line was the accurate placement of thin cucumber and tomato slices. A hygienic non-contact gripper solution based on the Bernoulli principle was developed by Davis et al. (2008) and shown in Fig. 2.6 that met the full operational speed required of the production line. The concept is now widely deployed in the sector, for example, in the packing of brittle popadom products, and the extension to other applications such as fresh vegetables has been explored by Petterson et al. (2011). The accurate and fast placement of thin, fragile pasta slices in the assembly of lasagne portions has normally required the application of skilled human dexterity. A simple mechanism developed by Moreno-Massey and Caldwell (2007) and shown in Fig. 2.7, easily emulates the human operators both in dexterity and speed. The principle is applicable to the assembly of similar food products. Certain food products require the folding of thin, fragile shapes, a task for which human fingers are well adapted. Gripper designs developed by researchers such as Yao and Dai (2008), demonstrate how this human dexterity can be emulated in accuracy and speed. An example of such a gripper generating a complex origami shape in light paper material is shown in Fig. 2.8; the extension of this

Fig. 2.6

Picking and placing of thin cucumber and tomato slices on a sandwich production line, using a Bernoulli non-contact gripper.

© Woodhead Publishing Limited, 2013

30 Robotics and automation in the food industry

Fig. 2.7 A cylindrical gripper for the acquisition and precision placing of thin pasta slices, on a Lasagne Ready Meal production line.

Fig. 2.8 The generation of a complex ‘origami’ shape in thin cardboard using a dexterous gripper design.

technique to fine bakery and confectionery products is obvious and explored more fully below. It can thus be summarised that gripper design is now a well-developed technology, can meet the challenges presented by the food industry, and need not set limits to the range of automation solution possible in food manufacture.

© Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 31 2.4.2 Sensor technology Manual operation has the advantage of providing a sequence of human actions, not just to assemble the products but also to obtain visual feedback for quality assurance and, if necessary, for corrective procedures. Removing human operators from the production line places an emphasis on automatic inspection systems, which must monitor the quality of the throughput at every stage in the process. Hygiene, safety and reliability are key concepts here, requiring the deployment of appropriate sensors, which must be of a hygienic design, non-contact, low-power, non-hazardous and proof against frequent, rigorous cleaning regimes. Current procedures involve the use of ubiquitous vision systems with metal detector and X-ray inspection systems for contaminant detection and removal. However, in fully automated systems there is a need to extend the range of available sensor technology to monitor constantly, for example, the quality of input raw materials and internal quality of each food product as well as the external surface characteristics. Some promising developments have been reported by Wu (2011), based on the use of low-power microwave technology, which have resulted in the design of sensors capable of the high-speed imaging of the internal structure of bread, fruit and other food products. An example of the internal image of an egg is shown in Fig. 2.9 and compared with the actual egg cross section obtained by dissection. The imaging process is non-contact, power dissipation is claimed to be a fraction of that of a mobile telephone, and speeds of up to 100 images per second are possible. Clearly this is the type of sensor development that must proceed in parallel

Joint 5 motor Joint 3 motor Joint 2 motor

Motor controllers

Fig. 2.9 The ‘Grail’ food-grade prototype robot.

© Woodhead Publishing Limited, 2013

Joint 1 motor

32 Robotics and automation in the food industry with the extension of fully automated systems to ensure quality of output and customer confidence in the production procedures. Another ideal future requirement for automated production is to trace each product from instant of packing through transport, storage, to point of sale and beyond. Current radio-frequency identification (RFID) technology is certainly technically capable of performing this function, but the existing cost model and equipment design do not yet appear appropriate for widespread deployment in the sector. The impact of successful deployment of this technology on quality control is clear, but it will also represent a powerful marketing tool by integrating traceability into a complete production-logistics control scheme that could optimise response to rapid changes in the market demand. Currently a Department of Food and Rural Affairs (Defra) funded, feasibility study into the application of this technology (Wu, 2010) is being undertaken at the University of Manchester, which if successful could accelerate the uptake of this important sensor technology within the industry.

2.5

Future trends

Market forces will continue to encourage the uptake of automation in food production with a range of manufacturing techniques. Standard industrial robots will be pervasive in the packaging and palletising regimes, while Delta-type robots will continue to dominate in the high-volume, assembly/pick-and-place role, particularly where product placement is poor requiring corrective visual sequencing and where flexible reprogramming may be a required option. For high-volume, single-product lines with good localisation, hard-wired electro-mechanical systems may provide more economical solutions. The challenge is to obtain affordable automation solutions in low- to medium-volume mixed-product lines with generally poor product localisation. The application domain would be predominantly in the SME sector, where there will be a demand for a mixture of manual and automated manufacturing solutions. Ideally this requirement could be met by a lightweight, low-inertia, bench/platform mounted sensor driven robot that meets the hygienic requirements of the food industry. A possible candidate for this role is the ‘Grail’ robot, being currently developed at the University of Sheffield, with the support of ‘Defra’ Link Research Funding (Moreno-Massey et al., 2010) and it is of interest to review its design concepts. This machine, which is currently undergoing industrial trials, is powered by brushless DC motors with simple belt-driven transmission mechanisms, configured to minimise arm inertia. All joints are back-driveable and the casing is constructed from lightweight stainless-steel panels, with special attention given to the design of joint seals to ensure that the robot meets the required hygienic operating standards and security against water ingress during wash down. An optimised trajectory planning algorithm has been developed to maximise pickand-place performance speeds and provide asymmetric acceleration profiles to minimise required peak joint torques. The general outlines of the robot are shown in Fig. 2.10, and the technical design specifications are listed below: © Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 33

Fig. 2.10

2.5.1 • • • • • • • • • • •

‘Food Factory in a Pipe’ concept.

Design specifications

Main application pick-and-place Material 304 or 316 stainless steel Degrees of freedom 4 Payload 1 kg Reach 720 mm Repeatability ± 2 mm Weight 20 kg Mounting floor/ceiling Ingress protection rating IP 67 Operating life 40 000 h Pick-and-place speed (25 × 300 × 25 mm cycle) – 90–100 cycles/min (0.5 kg) – 75–85 cycles/min (1 kg)

The combination of a lightweight structure, low-inertia drive geometry and constrained elbow joint velocities and torques contribute to the overall safety of a device that could work in close proximity to human operators with a minimum of barrier protection. Future developments will explore the incorporation of variably compliant joints to enhance the safety of the operation (Tsagarakis et al., 2009). An innovative teach-by-showing strategy is proposed for initial programming and rapid programme changes will be effected by a touch screen graphical selection panel that provides access to a menu of possible actions and requires no IT skills to implement (Gray and Pegman, 2011; Kormushev et al., 2010). This is clearly a feasibility study that will explore the possibility of developing, low-cost, relatively safe, highly functional robotic devices optimised for low to medium runs of variable geometry food products, within a working environment where the IT skill base is limited. If the concept is successful the development of such devices will be a useful addition to the growing array of highly effective automation equipment now available to the food industry, to meet the very diverse nature of its manufacturing profile.

© Woodhead Publishing Limited, 2013

34 Robotics and automation in the food industry Most food manufacturing processes are undertaken in standard factory units with high ceilings to accommodate a wide range of manufacturing equipment. However, the vast majority of food assembly processes are usually confined to a working volume that ranges from 2 to 3 m in height, resulting in a relatively large volume of unused factory space which must be cooled, filtered and hygienically treated. The concept of containing the manufacturing process within a relatively small, sealed environment was suggested by Gray (2004) and the idea explored in more detail with the support of Defra research funding (Gray et al., 2005). The so called ‘food factory in a pipe’ concept would consist of a series of sealed, standard dimensional, linked production modules of (for convenience of manufacture) circular cross section, that are lightweight in construction, easily moved, and locked in position with standard mechanical sealing and communication interfaces. A schematic of the approach is illustrated in Fig. 2.10. Variable configurations are possible; the system is scaleable and adjustable to meet varying production modes. The concept is considered to be most suitable for medium- to long-term production runs with the approach minimising internal environmental energy and water requirements and enhancing hygienic operation, through the complete removal of operators from contact with the product – a feature that allows the use of inert environments, possibly with powerful decontaminating procedures that could enhance the shelf life of the product. Clearly there are engineering challenges to be met in the practical implementation of such a scheme, such as the development of automatic quality-control systems, and a facility for intervention in the case of system failure. However, such issues have already been addressed in other manufacturing sectors where operators must be physically separated from the process for safety or hygienic reasons. Such a concept appears to address a number of manufacturing and environmental issues that are currently impacting on the sector and which will influence its future evolution. Full automation will produce outputs of consistent quality, with less waste and minimum risk of contamination through human contact. Confining manufacture within a sealed environment will result in less effluent disposal, greater hygienic potential, and an overall smaller carbon foot print. The required computer control system will be compatible with advanced traceability technology, logistics, and automated warehousing procedures, allowing management to adopt a fast and flexible response to market demand, reduce waste and maximise the return on investment.

2.6

Conclusion

Food manufacturing is, by its nature, quite diverse and its procedures will no doubt remain so, but the bias against automation has now largely disappeared. Manual operation will remain, particularly to meet local or bespoke markets, but will be interspersed with islands of appropriate automation. However, in large-scale production, automation will dominate, with the emergence of fully automated plant, geared for flexible production in an energy-efficient aseptic environment and linked to a totally traceable logistics chain to the point of sale and perhaps beyond. The overall shape and nature of the industry will be determined by market forces, which in turn will be influenced by demographic, ecological and economic © Woodhead Publishing Limited, 2013

Robotics in the food industry: an introduction 35 pressures. The automation industry has an important role to play in providing the technology that will ensure that the sector continues to meet the demand for readily available, high quality and affordable food products. Perhaps the greatest challenge the industry faces in the uptake of automation technology is its limited access to professional engineering and IT skills that are required to underpin the adoption of advanced manufacturing techniques. Thus educational establishments, training centres and professional engineering institutions have an equally important role in creating a genuine twenty-first century European food manufacturing industry.

2.7

References

CHUA P.Y., ILLNER T. and CALDWELL D.G. (2003) Robotic Manipulation of Food Products A Review. Industrial Robotics: An International Journal. 30(4), 345–354. Davis S., Gray J.O. and Caldwell D.G. (2008) An end effector based on the Bernoulli principle for the handling of fruit and vegetables. International Journal of Robotics and C.I.M. 24(2). EHEDG (2007) Materials for the construction of equipment in contact with food. Trends in Food Science and Technology 18(S1), S40–S50. Gray J.O. (2004) Modular reconfigurable shelled automation in food manufacture. Proceedings of E.C. Food Factory of the Future Conference Series, Laval, France, October, pp. 53–58. Gray J.O. and Pegmann G. (2011) Festo Advanced Handling Conference. Northampton, UK, March ([email protected]). Gray J.O., Lloyd B. and De Vic A. (2005) Modular reconfigurable shelled robotics automation. In Food Manufacture Final Report: Defra Link Project FT0619, January. Kormushev P., Calinon S. and Caldwell D.G. (2010), Robot Motor Skill Coordination with EM-based Reinforcement Learning, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2010) Taipei, Taiwan. LELIEVELD H. and KEENER L. (2007) Global Harmonisation of Food Regulations and Legislation. EHEDG Yearbook 2007 Trends in Food Science and Technology 18(S1), 15–19. Lelieveld H.L.M., Mostert M.A., Hola J. and White B. (eds) (2003) Hygiene in Food Processing. Woodhead Publishing, Cambridge, UK. MORENO-MASSEY R.J. and CALDWELL D.J. (2007) Design of an automated system for limp flexible sheet Lasagne Pasta. Proceedings of the IEEE International Conference on Robotics and Automation, pp. 1226–1231. MASSEY R., GRAY J., DODD T. and CALDWELL D. (2010) Guidelines for the design of low cost robots for the food industry. Industrial Robot 37(6), 509–517. Petterson J.A., Ohlsson T., Gray J.O., Davis S. and Caldwell D.G. (2010) Bernoulli principle gripper for handling planar and three dimensional products. Industrial Robot 37(6), 518–526. Tsagarakis N.G., Laffranchi M., Vanderborght B. and Caldwell D.G. (2009) A Compact Soft Actuator Unit for Small Scale Human Friendly Robots, ICRA 2009, Kobe, Japan, pp. 4356–4362. Purnell G. (1998) Robotic equipment in the meat industry. Meat Science 49(S1), 297–307. Wallin P.J. (1995) Review of the use of robotics and opportunities in the food and drinks industry. Industrial Robot 22, 9–11. Wallin P.J. (1997) Robotics in the food industry, an update. Trends in Food Science and Technology 8(8), 193–8. Wu Z. (2010) Defra Food Link Project AFM 289/FT1576. Wu Z. (2011) Private communication. University of Manchester, UK. Yao W. and DAI J. (2008) Dexterous manipulation of origami cartons with Robotic fingers, bases on interactive configuration space. Trans ASME Journal of Mechanical Design 130(2), 022303. © Woodhead Publishing Limited, 2013

3 Sensors for automated food process control: an introduction P. G. Berrie, Endress+Hauser Process Solutions AG, Switzerland

DOI: 10.1533/9780857095763.1.36 Abstract: The measurement of process variables provides not only the means for monitoring and controlling a process, and hence for providing constant quality unaffected by the operator; it also can be key to reducing capital tied up in inventory and to using energy more efficiently. Parallel to this, the management of assets both at process and enterprise level is gaining in importance. This chapter reviews the requirements of the food industry on field instrumentation, the instruments available to measure process variables, and the integration of these instruments into automation systems. A few practical examples and an outlook on future developments complete the chapter. Key words: process instrumentation, process control, plant asset management, device integration.

3.1

Introduction

Food, like water, is an essential requirement of life. For many in the world who live where food is scarce, necessity overrules quality in the daily struggle for life. In the developed world, however, quality is a huge factor in determining what we buy. Quality, of course, lies in the eyes and taste buds of the consumer: for the greens among us in ‘organically grown’ fruit and vegetables, for the dietician in nutritional value, and for the greedy in the size of the portion. For the manufacturer, other criteria must be fulfilled to achieve basic quality. What no-one wants is snails in their lettuce and flies in their soup, or as far as processed food is concerned, hair, glass, metal, plasters, chemicals and bacteria in their tins or packages. And of course, akin to a Warhol painting, today’s tin of soup should taste like yesterday’s tin of soup, and tomorrow’s like today’s. It is not only the variety of processed food on the market available today that impresses but also the consistent quality with which it is produced.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 37 A second factor in food choice is, of course, cost – not just for the consumer who likes to buy his food as reasonably as possible, but also for the supermarket where he buys. Faced with fierce competition from rival companies, price-cutting dictates the market. This in turn forces the food manufacturer to produce as cheaply and efficiently as possible, but without a loss of quality. Cost reduction is traditionally associated with an increase in automation and reduction of manpower within a production facility. Consistent quality has much to do with controlling processes. ‘First measure, then control’ is an old adage, but depending where it is applied along the supply chain, measurement brings additional benefits. Food production facilities share a common infrastructure: inbound raw materials are stored ready for the production process; outbound products are stored before delivery. Here measurement of stock level, and concepts such as vendormanaged inventory, can lead to significant savings through reduction of stock to the minimum required for reliable production. The production process itself is often controlled by recipes to make products that frequently change. Then, the finished product must be bottled or packed, a procedure that requires an entirely different kind of control from that of the making of the product. Finally there are the so-called utilities: electricity, air, heat, cooling and water. Efficient use of energy has become a byword in manufacturing, and key to optimal use is measuring what is consumed.

3.2

Special considerations for food instrumentation

The processes involved in producing food find many parallels in other industries. Where they differ is that food is produced for eating and drinking by humans or animals: as a result, the product must be safe for consumption. For process instrumentation this leads to a number of special considerations with regard to its design and use.

3.2.1 Regulatory agencies The food industry differs from many others in the extent to which the entire supply chain, from production to distribution, is monitored by government agencies. The industry itself has also set up its own watchdogs to assess advances in production engineering and make recommendations regarding good manufacturing practice. When choosing process instrumentation, therefore, it pays to check whether a corresponding approval or authorisation has been granted. Table 3.1 lists several bodies that are of particular importance to safe food manufacture. The prime aim is to ensure the safety of a product, thus protecting consumers from exposure to, for example, chemical contaminants and toxins, bacteria or foreign bodies. Once identified as such, a defective product has to be removed from the market immediately, demanding high levels of traceability at all stages of production. As far as instrumentation is concerned, the effect of regulation is twofold. On the one hand, the regulations dictate how and of what materials a

© Woodhead Publishing Limited, 2013

38 Robotics and automation in the food industry Table 3.1 Regulatory and standardisation authorities Body

Field of activity

3-A sanitary standards

American standards service with a long history of producing hygienic standards for the dairy industry. Also tests and authorises process instrumentation regarding its fitness for use in food production. 3-A standards are binding in the USA, but nowadays have less impact worldwide. European Hygienic Formed in 1989 with the primary aim of the safe manufacture Equipment Design of food products, is supported by research institutes, Group equipment manufacturers and end users. Issues guidelines on hygiene, tests and certifies equipment and publicises state-of-the-art technologies. An EHEDG approval means that the equipment has been successfully tested with respect to its suitability for food applications. International Dairy Responsible for standards and recommendations in the dairy Federation industry which are often published as ISO standards. US Food and Drug US government agency charged with ensuring the safety of Administration food production. Issues licences for products, whereby the processes, constituents, materials and constructional details are subject to examination. The use of an FDA-approved material is a guarantee that the component concerned will not be attacked by the food product. EU and national Play a similar role to the FDA in the European Union or the governmental country to which they belong. Lay down the permissible agencies levels of trace organic and inorganic substances in food products. Since this is an area of permanent research, it is possible that regulations vary from country to country. Retail organisations: IFS, Retail associations, certification bodies and quality standards GFSI, BRC, SQF, set up to ensure the quality of food in supermarkets and ISO 22 000 other retail outlets by auditing their suppliers. Although not the primary focus, good manufacturing practice and correctly installed technology are factors in certification.

measuring device may be constructed, on the other, their measurements help fulfil the regulations by providing the information necessary to control and monitor the process. 3.2.2 Wetted parts A well-designed instrument for use in the food industry must take three factors into consideration: how does the device behave in normal operation, what happens if a component part becomes mechanically defective and how can it be cleaned and sterilised. The so-called wetted parts, that is, those parts of the instrument in direct contact with the material being measured, are the most critical in both respects. Even non-contact devices must be considered to have wetted parts when they intrude into a vessel. Although they are not directly in contact with the medium, they may have crevices in which food can gather and rot, or be exposed to high temperatures or vapours that release unwanted substances.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 39 Table 3.2 Various stainless steels suitable for the food industry Material

AISI*

Properties

Uses

1. 4301

304

Pipework, vessels, instrument housings

1. 4404

316 L

1. 4435

316 L

1. 4571

316 Ti

1.4362 1.4462

2205 2304

Good resistance against organic acids, salt and alkalis at moderate temperatures Increased resistance against non-oxidising acids such as acetic acid, tartaric acid, phosphoric acid. Increased resistance against pitting and intercrystalline corrosion Better corrosion resistance than Type 1.4404 Increased corrosion resistance against particular acids and salt water Resistance against pitting corrosion High strength duplex steels with good resistance to stress corrosion cracking in salt solutions at elevated temperatures Used with, e.g. hot brine (>50°C) with solids, stagnant and slow moving salty foods

Pipework, vessels, instrument wetted parts

Pipework, vessels, instrument wetted parts Pipework, vessels, instrument wetted parts Retorts, storage vessels where light construction and strength is important, instrument wetted parts (for the oil industry)

* The AISI steels are equivalents but do not have identical compositions.

Taking the three cardinal sins, chemical contamination, bacteria and foreign bodies as criteria, each wetted part must be checked to see whether it poses a hazard. Chemical contamination may be caused by migration, wear and tear of a moving part, abrasion by flowing material, leakage into the process or simply by choosing the wrong material for the foodstuff being processed. It is important to remember that, even if we do eat them, foods are nothing more than mixtures of organic and inorganic chemicals, which may be acid or alkaline and which may react with the materials with which they are in contact. To avoid chemical contamination through the wrong choice of material, it is important that instrument manufacturers use substances considered safe for food production. To this end, the Food and Drug Administration of the United States of America (FDA) maintains a list of Food Contact Substances (FCS) (FDA, 2010). A similar document for metals is Guidelines on Metals and Alloys used as Food Contact Materials (Council of Europe, 2002), and for stainless steels the Sheffield Corrosion Handbook (Avesta, 1994). Most instruments used in food production are made of steel grades 1.4301/304 (18% Cr, 10% Ni) or 316 (17% Cr, 12% Ni + Mo), although other alloys or titanium may be used. Table 3.2 lists the properties of various stainless steels used in food production.

© Woodhead Publishing Limited, 2013

40 Robotics and automation in the food industry A number of factors may help decrease the risk of bacterial contamination. Regular cleaning and gap-free design reduces the broad risk of infection, but the proper material, design and finishing of the wetted parts is just as important. Dead ends should be avoided since they are harder, if not impossible, to clean. In addition to product wastage, they mean longer phase shifts with more flushing water, longer heating times for cleaning, and a huge heat demand for steam sterilisation. Wetted parts must exhibit a high degree of corrosion resistance, both against foodstuffs with which they are in contact as well as the agents used to clean and sterilise them. In addition, high temperatures, strong vibration and mechanical stress may enhance the risk of electrochemical and intergranular corrosion of stainless steel. The one results in surface pitting, providing an ideal breeding place for bacteria, the other in the depletion of the nickel and chromium at the grain boundaries, which means a component will rust. Design factors that reduce the risk of bacterial contamination are streamlined contours, clean welding, no nooks and crannies, and no obstructions that might cause the product to gather and rot. It is also standard practice that wetted stainless steel parts are electrically or mechanically polished. A surface roughness Ra ≤ 0.8 μm/150 grit is regarded as safe but, for certain process connections, even lower values are normal. Polishing has the double effect of reducing susceptibility to surface pitting and preventing products from sticking. The introduction of foreign bodies happens more by accident than design, but good risk analysis can mitigate the effects, should the accident happen. What could be the causes of failure, and what happens when a component fails? Are flowing gases, liquids or solids causing abrasion or generating high mechanical forces that, combined with high temperature or vibration, are increasing electrochemical attack or mechanical fatigue? Is the back pressure of a flow meter high enough to prevent liquids from boiling and causing cavitation? Has a thermowell been designed to withstand the pressures to which it is subject? Does its position within a pipe subject it to more than the design forces? Are instruments able to withstand the forces and temperatures generated during external cleaning or cleaning-in-place? What happens if the diaphragm of a pressure transmitter is ruptured? Is the fill fluid FDA approved – does its release mean that the entire batch must be thrown away? These and many other criteria should be assessed when choosing a measuring instrument. It is the result of such considerations that pressure transmitters with metal diaphragms, rather than ceramic measuring cells, are used in direct contact with food. Mercury-in-glass thermometers are actually proscribed by FDA Rule 21 CFR Part 113 for low-acid canned-food retorts, but here concerns by the manufacturers about mercury spills have lead the FDA to propose a rule change allowing equivalent electronic devices to be used instead. In food filling applications, there is a definite trend towards replacing older mechanical piston and cup fillers by electromagnetic and Coriolis flowmeters. Here wear and tear, maintenance, as well as measurement accuracy, are the driving factors for refurbishment.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 41 3.2.3 Process connections A process connection is the means by which an instrument is installed in a pipe or vessel that contains the process material to be measured. Ideally, it should offer the product no possibility to become entrapped. One way to do this is to weld the instrument connection in place and then grind and polish the weld. This is, of course, not practical for small pipe diameters and replacement is difficult, should the instrument fail. Nevertheless welding is often encountered for thermowells, where the sensor insert is easily replaced, as well as for some types of flowmeter. The majority of instruments in food processing, however, are installed by means of so-called sanitary couplings. These provide gap-free connection and are designed to be easily dismounted, allowing instruments and pipes to be quickly removed for cleaning. The couplings themselves are normally held together with clamps or external screw unions. Over the years a number of different designs have come onto the market, a selection of which are summarised in Table 3.3. Manufacturers also offer special adapters for mounting instruments on tanks and ports for the insertion of, for example, pH sensors into tanks and pipes. Here it is important that these have a sanitary approval issued by the European Hygienic Equipment Design Group (EHEDG) or similar body.

3.2.4 Instrument housings Depending upon the design of the instrument, a housing may contain only the connecting terminals, for example, as for a temperature probe, or the entire transmitter electronics. In both cases it must provide adequate protection against the ingress of moisture or dust and, when the instrument is used in an explosive risk area, the egress of a spark or flame. The former is governed by the degree of protection provided by the housing, the latter by a suitable type of protection provided by the housing or the instrument output circuit. As far as the ingress of moisture and dust is concerned, there are two major classification systems, IP ratings (IEC, 2001) and NEMA ratings (NEMA, 2008). Tables 3.4 and 3.5 explain the meaning of each rating type. It is usual that manufacturers quote both in their technical specifications. In the IP standard, the degree of protection is indicated by a two-part code. The first number is concerned with the protection from the ingress of solid matter, the second with water: the higher the number, the better the protection against dust, water jets or submersion. In order to withstand the frequent cleaning in a food production facility, housings with ratings of IP 65 or better are required. The NEMA standard comprises 14 type codes that deal with practical requirements on housings suitable for indoor and outdoor use. It also makes a statement about the protection from external influences and conditions such as mechanical impact, corrosion, humidity, mould, pests, dust, etc. A NEMA 4X enclosure is best suited to the requirements of the food industry. Explosion protection is not a major issue in the production of food. If flammable liquids or easily ignitable gases are present, for instance in the distillation

© Woodhead Publishing Limited, 2013

42 Robotics and automation in the food industry Table 3.3 Sanitary couplings for the food industry Type

Use

Description

Dairy coupling to DIN 11851

Pipes and tanks

Low cost coupling with threaded boss and slotted sleeve. It does not allow flush mounting and is no longer considered to be hygienic; however, a special gasket set can be purchased which brings the coupling up to standard. Designed to EHEDG recommendations, as a replacement for the dairy coupling, these unions offer better hygiene thanks to a flush sealing construction. The mechanical coupling is via threaded sleeve, bolts or clamp, the seal being flush with the pipe wall. Can be used for in-line and insertion instruments. In-line housing that allows the flush mounting of a sensor, which is attached to the housing by means of a screw clamp. Three housing types cover a wide range of pipe diameters. A tank adapter is also available. In-line housing of similar construction to the Varivent coupling. The sensor, however, is bolted in position. Low cost, Scandinavian standardised screw coupling to SMS 195346. Its weakness lies in the hygienic adaptation to the process, which does not allow flush mounting, and for this reason it is no longer considered to comply with modern hygienic standards. International Dairy Federation (IDF) screw coupling standardised in ISO 2853 Sanitary coupling with bevel seating produced by the Tri-Clover Company in America. Instruments are quickly mounted and fixed with snap-on clamps. The couplings find widespread use in America.

Aseptic screwed, flange Pipes and tanks (DRD) and clamp unions to DIN 11864-1, DIN 11864-2 and DIN 11864-3

Varivent® coupling

Pipes

APV coupling

Pipes

SMS coupling

Pipes and tanks

IDF coupling

Pipes and tanks

Tri-clamp® coupling

Pipes and tanks

of spirits, then instrumentation must be approved for use in explosion-hazardous areas. Powders can also be a problem, since clouds of dust are easily combustible under certain conditions. For milling, storage, conveyance and bagging operations, therefore, the Dust-Ex equipment should be used. Although suitable types of protection have long been agreed, until recently Europe and North America had their own nomenclature for classifying a hazardous

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 43 Table 3.4 Ingress protection categories according to IEC 60 529 Code

Ingress protection against solids

Code Ingress protection against water

0 1 2 3 4 5 6

Not protected ≥50 mm diameter, e.g. hand ≥12.5 mm diameter, e.g. finger ≥2.5 mm diameter, e.g. tool ≥1 mm diameter, e.g. wire Protected from dust Dust-proof

0 1 2 3 4 5 6 7 8 9

Not protected Vertical dripping Dripping (15° inclination) Water spray Splash water Jet of water Strong jet of water Temporary submersion Total submersion High pressure water, 80°C, 120 bar, 30 cm distance from nozzle

Table 3.5 Degree of protection of enclosures according to NEMA Standard 250 Type

Indoor

1

Yes

2

Yes

Outdoor

3

Yes

3R

Yes

3S

Yes

4

Yes

Yes

4X

Yes

Yes

5

Yes

6

Yes

Yes

6P

Yes

Yes

11

Yes

12

Yes

12K

Yes

13

Yes

Degree of protection Protection against contact with equipment within the housing Protection against a specified quantity of water droplets and dirt Protection against blown dust, rain, sleet, rain and snow, and external ice formation Protection against rain, sleet, rain and snow, and external ice formation Protection against blown dust, rain, and operation of an external mechanism is guaranteed if this is iced up Protection against blown dust, jets splashes and jets of water Protection against corrosion, blown dust, jets splashes and jets of water Protection against dust falling, dirt and lubricating non-corrosive fluids Protection against water penetration during occasional temporary submerging in limited depth Protection against water penetration during prolonged submerging in limited depth Protection against the corrosive effects of fluids and gases of operating material inside the enclosure submerged in oil Protection against dust, dirt and lubricating non-corrosive fluids Protection against dust, dirt and lubricating non-corrosive fluids for enclosures subjected to vibration Protection against dust, water spray, oil and non-corrosive coolants

© Woodhead Publishing Limited, 2013

44 Robotics and automation in the food industry Table 3.6 Types of protection for gas and dust (D) atmospheres Type of protection

Standard

Significance

Flameproof enclosure ‘d’ Protected by enclosure ‘tD’

IEC 60079-1 IEC 61241-1

The enclosure (housing) is designed such that any spark or explosion is retained within it. Frequently met in America where the entire cabling is routed in conduits. Also used in Europe for Dust-Ex protection and system components. Electronics are designed such that high temperatures and sparking cannot occur during normal operation. This type of protection is met e.g. in 4-wire devices and power supplies. Protection is achieved by limiting the current flowing in the device signal circuits. Restrictions apply to installation and cables used. Favoured method for European installations. Device is designed so that, under normal and defined abnormal operating conditions, it cannot ignite a surrounding explosive atmosphere. Used for all Zone 2 electrical equipment. The entire electronics of the device are potted, so that no spark can enter the surrounding atmosphere.

Increased safety ‘e’ IEC 60079-7

Intrinsic safety ‘i’ Intrinsic safety ‘iD’

IEC 60079-11 IEC 61241-11

Type of protection ‘n’

IEC 60079-15

Encapsulated ‘m’ Encapsulated ‘mD’

IEC 60079-18 IEC 61241-18

area. The situation has been eased by the latest IEC standard, which has also been adopted by North America. Table 3.6 lists and explains the types of protection of interest to food processing. Readers wanting a fuller explanation of explosion protection are referred the booklet ‘Basics of Explosion Protection’ (Stahl, 2007).

3.3

Measurement methods

In order to ensure consistent quality and economic production of food, the underlying processes must be controlled. This means in turn that the variables affecting the process, for example, temperature, or the parameters indicating a certain quality level, for example, specific gravity, must be measured. Normally in process control it is sufficient to measure the variables pressure, temperature, flow and level. Variables such as pH, turbidity, viscosity and density give an insight into the quality of a product. There are other important quantities, however, that are not covered in this chapter. Trace moisture and humidity sensors play an important role in storage and conditioning of products such as stored fruit, meat or other raw materials. The reader interested in humidity measurement methods is referred to other literature on the subject (Bentley, 1998).

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 45

Fig. 3.1 Pressure transmitter on a keg-filling line. (Source: Courtesy of Endress+Hauser.)

Readers wanting a more comprehensive overview of the instrumentation available to measure pressure, temperature, flow and level are referred to ‘Instrumentation and sensors for the food industry’ (Berrie, 2001). This chapter restricts itself primarily to electrical instruments that are easily integrated into control and plant asset-management systems.

3.3.1 Pressure In food processing, the main applications for pressure measurements are in piping, across filters, as well as in closed tanks. Figure 3.1 shows a typical application. The toughest conditions are to be found in piping, where instruments must be designed to withstand overload pressures far beyond their normal operating range. Pumping action and opening and closing of valves causes surges in pressure up to the maximum operating pressure of the pump. High pressure peaks are also generated if a valve closes abruptly – a particular problem in the lowpressure region. Pressure instruments must also be immune to vibration and be able to withstand the temperatures and stresses induced by internal and external cleaning. Pressure instruments may measure absolute, gauge and differential pressure: • Absolute pressure devices measure the actual pressure acting on the sensor, displaying approximately 1000 millibar for atmospheric pressure.

© Woodhead Publishing Limited, 2013

46 Robotics and automation in the food industry • Gauge-pressure devices measure the pressure relative to atmospheric pressure, displaying approximately 0 millibar for atmospheric pressure – lower pressures are output as negative values. • Differential pressure devices measure the pressure difference between two tapping points, for example, in a pipeline or a tank. They are usually used to measure level and, in connection with a primary element, flow. There are three basic types of pressure-measuring instrument: manometers, mechanical pressure gauges and electrical pressure transmitters. Of these, electrical transmitters are most suited to automation, although the other types are still encountered in many plants. Electrical pressure transmitters Electrical pressure transmitters output a standardised signal that represents the quantity being measured, for example a 4–20 mA current signal, or a measured value transmitted over a digital communication interface. Most electrical pressure transmitters use a flexible, ceramic or metallic diaphragm as the pressure-sensing element, which forms the front isolating element of a sensing chamber. By using the appropriate process connection, it is possible to produce flush-mounted pressure devices that have no cavities and that are easy to clean. In general there are three measurement principles: resistive, capacitive and inductive. The design of the cells depends on the manufacturer and measurement principle used. Pressure can also be measured by the resonance-wire or vibrating-beam method, whereby the shift in resonance frequency or change in vibration frequency is proportional to the pressure acting on the sensing element. Table 3.7 summarises typical operating conditions for pressure-measuring instruments. No accuracies are included, since they depend on the individual transmitter. Most data sheets quote the measured error, the hysteresis and reproducibility measured at a reference temperature, usually 20°C although sometimes other ‘overall’ accuracies are given. The temperature effect, which indicates by how much the accuracy changes for a given temperature rise, is an important factor in pressure measurement. Depending on the fill fluid, which is required when the sensor is operating with a diaphragm seal (an extension which allows flush mounting) or a remote seal and capillaries, this can be quite large. The overload pressure indicates the ability of a device to withstand pressure peaks without mechanical damage to the sensor (a recalibration may be necessary).

3.3.2 Temperature Monitoring and control of temperature is an important factor in assuring the quality of the end product. For example, a wrong temperature during the manufacture of dairy products seriously affects their shelf life, while the wrong temperature in a retort for producing canned foods may result in bacterial contamination. In pasteurisation, freeze drying and similar processes, temperature control is essential

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

–10°C to +60°C

Manometer

Pressure gauge –25°C to +60°C, (diaphragm, diaphragm gauges capsule, up to 100°C Bourden gauges) Electrical pressure –30°C to +200°C transmitter

Typical temperature range

Type

0–40 bar (ceramic) Up to 3 bar for static 0–600 bar pressures up to 420 bar (metal) (ceramic) Up to 500 bar for static pressures up to 500 bar (metal)

0–100 mbar

–

–

0–25 bar

0–20 bar

Differential pressure/static pressure

0–10 mbar

0–1 mbar

Min. pressure Max. pressure range range

Resistant to vibration and overload. Measuring range of 4–20 mA/HART devices can be turned down by 40:1 or more. Differential pressure transmitters with remote seals and capillaries may have a significant temperature effect

Inaccurate if not operated within the rated temperature range Significant temperature effect. Diaphragm gauges can withstand up to 5× overload (maximum 40 bar)

Remarks

Table 3.7 Typical operating conditions for pressure-measuring devices. Values are guidelines only: for specific values see the manufacturer’s data sheet

48 Robotics and automation in the food industry

Fig. 3.2 Temperature measurement in a CIP pipe. (Source: Courtesy of Endress+Hauser.)

if the quality and safety of the product are to be guaranteed. Any unaccountable rise in temperature in a storage silo may be the first indication that the raw material in store is deteriorating. Temperature is also measured in utility applications, for example in refrigeration, steam heating or cleaning-in-place (CIP) plant, see Fig. 3.2. The temperatures to be recorded range from about −50°C (−58°F) in cold storage to +150°C (+302°F) in sterilisation-in-place (SIP) applications. Only in steam generation will higher temperatures up to +250°C (482°F) be found. Accuracy, depending on application typically ±1°C or better, is very important, and of course, hygiene is essential where there is contact with food. Measurement principles Temperature can be determined by measuring one of three quantities on which it has an effect: force, electricity and radiation. Radiation is the phenomenon used by pyrometers, and since they are mainly used in high-temperature applications, for example, in furnaces, they will not be considered further. Force-type temperature devices include bimetallic strips, which may be used as continuously-reading gauges or switching devices, as well as filled thermal systems, that is, thermometers. Mercury-in-glass thermometers are still proscribed for certain applications by the FDA, see Section 3.2.1, but in these cases, a second temperature sensor can be installed for control. Otherwise their use today is limited to optical back-up of electrical systems.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 49 Electrical temperature devices Electrical temperature devices use the temperature dependency of electrical properties of particular materials to provide a measurement. There are four different types: semiconductors, resistance temperature detectors (RTD), thermocouples and silicon resistors, although only RTDs and thermocouples are of real interest in the production of food. RTDs, also known as resistance thermometers, comprise a thin-film or wire resistor with a standard resistance of 100 Ω, 500 Ω or 1000 Ω. The resistor material may be platinum or nickel, the standardised designations being Pt100, Pt500, Pt1000 or Ni100. The sensors are very stable, have a wide operating range from example −200°C to +850°C (−358°C to +1562°C) depending upon type, and exhibit a well defined relationship between resistance and temperature. The dependencies are specified in a number of international standards (IEC, 2008). RTDs conform to one of two standardised accuracy classes: Class A = 0.15°C + 0.002 (t°C); Class B = 0.30°C + 0.005 (t°C); where (t) is the unsigned numerical value of the temperature in °C. Higher accuracies are also available. RTDs often have a relatively slow time response and produce only a small change in resistance per unit increase in temperature. To limit or avoid errors where sensing element and evaluating electronics are separated by some distance, 3-wire or even 4-wire connections are used. RTDs are also sensitive to vibration and shock, so that care must be taken in their design. A thermocouple (TC) comprises two wires of different metal joined together at their ends. If the two junctions are at different temperatures, a potential difference is created that causes a current to flow around the loop. If one of the junctions is kept at a constant temperature, the magnitude of the current that flows is a measure of the temperature at the other. TCs have a wide operating temperature range: depending upon type and conductor diameter, they can measure up to +1800°C (+3472°F) and are standardised (IEC, 1995). They usually have a faster response than RTDs, but are less accurate (±0.5°C to ±1.5°C for IEC 60584 Class 1, depending on type). In Europe they are usually only used in preference to RTDs for temperatures above 600°C. In America, however, they are being increasingly used for lower temperature ranges where accuracy is not a major factor in measurement. TCs are simple, rugged, and inexpensive and require no external power supply. On the other hand, the signal is non-linear, they exhibit low sensitivity, and have relatively low stability. A reference junction is required and they must be compensated. Moreover, metallurgical changes and ageing sometimes causes a loss of performance. A more serious problem is that the low voltage output is susceptible to electromagnetic interference. Although a complete temperature transmitter with thermowell can be purchased from most manufacturers, many users prefer to buy the various components from different suppliers. Figure 3.3 shows schematically a typical temperature assembly:

© Woodhead Publishing Limited, 2013

50 Robotics and automation in the food industry Housing lid Connection head Cable gland Housing base Sensor insert

Extension/neck

Compression screw

Process connection Thermowell (protective tube)

Fig. 3.3 Components of a temperature-measuring assembly, showing a sensor insert with terminal connections.

• The sensor insert is the heart of the assembly. It comprises a metal sheath containing the sensing element and connection wires. The connection terminals normally cater for 2-wire, 3-wire and 4-wire connection. • The housing protects the sensor insert, the terminals and, where appropriate, the head transmitter from the ingress of water and dust. Many housing designs are standardised. • The thermowell may be separate or an integral part of the temperature assembly and provides the contact with the medium. It is designed for a particular medium and range of flow rates and must be used as described in the appropriate standard or data sheet. • For high-temperature applications, an extension (or neck) is inserted between the housing and thermowell. To reduce mechanical load when thermowells are mounted in pipes, a position or orientation should be chosen that reduces the surface area facing the direction of flow. In addition, a thermowell must be covered with medium at all times when the signal is required. Since the thermowell produces turbulence in the downstream flow, care should be taken that it is not positioned too closely to any other instrumentation in the pipe. In tanks, there is usually space for vertical or horizontal mounting. Care should be taken not to position it too close to moving parts such as agitators, pumps or valves, since the forces generated by the movement of fluid may exceed the mechanical strength of the thermowell or lead to increased abrasion or corrosion.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 51 Table 3.8 Areas of application of electrical flowmeters Measurement principle

Application

Electromagnetic

Electrically conductive liquids (>5 µS/cm) with or without solids, e.g. water, wastewater, sludge, slurries, pastes, acids, alkalis, juices, fruit pulp, etc. Virtually all fluids: cleaning agents and solvents, fuels, vegetable oils, animal fats, alcohol, fruit solutions, vinegar, ketchup, mayonnaise, gases, etc. Mass flow, density and temperature (the primary measured variables) can be used to derive other variables such as volume flow, solid contents, concentrations, and complex density functions. Viscosity can also be measured. Volume flow of liquids, gases and steam. Volume flow of any liquid, regardless of electrical conductivity. Flow measurement of gases, steam and liquids. Mass flow of gases.

Coriolis

Vortex Ultrasonic Differential pressure Thermal mass

3.3.3 Flow Flow is an important control variable that also delivers quantity information to the plant operator. There are many methods of measuring flow, and selection can be difficult. It must take into account not only the physical properties of the medium to be measured but also the required accuracy, line size, environment, purchase costs and installation. With the exception of clamp-on ultrasonic instruments, all flowmeters are installed in-line, so hygiene is also an important factor. The sensing element must be able to withstand the temperatures met in CIP or SIP procedures, which most hygienic versions do, or if the instrument cannot be cleaned in situ, it must be easily dismantled to facilitate cleaning. Fluids with high viscosity or containing entrained solids call for non-obstructive measurement. If fluids solidify at relatively high temperature, as is the case for hydrogenised oils, or are sticky, as are chocolate and sugar syrup, the flowmeter must be heated. Condensation might be a problem when the product is chilled: in this case a remote housing, which can be mounted at some distance from the measuring point, must be selected. Mechanical flowmeters are seldom used in hygienic processes nowadays, but two mechanical flowmeters, positive displacement and turbine, are still encountered in older plant. A description can be found in ‘Instrumentation and sensors for the food industry’ (Berrie, 2001). Table 3.8 indicates the areas of application of each measuring principle covered in this chapter. For the purposes of flow measurement a conductive liquid has a conductivity ≥ 5 µS/cm for externally powered flowmeters, ≥ 50 µS/cm for loop-powered devices. Examples of conductive liquids are well water, fruit juices, yoghurt, milk, beer in addition to acids, alkalis and water-based emulsions. Nonconductive liquids are fuel oils, organic solvents, liquid sugar and demineralised water.

© Woodhead Publishing Limited, 2013

52 Robotics and automation in the food industry

Fig. 3.4 Electromagnetic flowmeters installed in a brewery. (Source: Courtesy of Endress+Hauser.)

Electrical flowmeters Electromagnetic flowmeters use the voltage induced by a conductive liquid flowing through a magnetic field to measure volume flow (Faraday’s law of electromagnetic induction). The induced voltage is proportional to the flow rate and independent of changes in fluid density, viscosity and pressure. As the flowmeter is essentially a straight pipe, not only beer, water and milk can be measured, but also products such as yoghurt, molasses (when conductive) or rice pudding, which could loose their constitution if forced through a restriction. The flowmeter is suitable for a variety of applications, for example flow control, custody transfer and high-speed filling. Figure 3.4 shows two electromagnetic flowmeters installed in a brewery. Electromagnetic flowmeters have a general installation requirement of only 3–5 straight pipe diameters upstream of the measuring point. They operate best in a turbulent flow regime. If used to monitor high viscosity products in a laminar (plug flow) or transitional flow regime, the absolute accuracy of the meter may suffer. Coriolis mass flowmeters use the frequency shift caused by a flow of liquid through two metering tubes oscillating at their natural frequency to measure mass flow. Since the natural frequency is directly related to the mass of fluid in the tubes, it is also possible to measure density. The tube temperature is also monitored to compensate for density and single-tube flowmeters can be used to measure viscosity. This versatility combined with high accuracy makes the Coriolis mass flowmeter a very valuable measurement device. The straight tube design is ideal for the food industry, provided titanium is an acceptable material. Coriolis flowmeters find use in continuous blending, filling and process monitoring applications, particularly when the mixing is controlled by density or the final product is to be sold by weight. Their high accuracy makes them suitable © Woodhead Publishing Limited, 2013

Sensors for automated food process control 53

Fig. 3.5 Coriolis flowmeters installed on a filling machine. (Source: Courtesy of Endress+Hauser.)

for dosing flavours and similar food additives. They measure viscous media and compound products such as soup, chocolate, honey and mayonnaise. There is no requirement for up- and down-stream lengths and the devices operate independently of changing fluid properties. Figure 3.5 shows a bank of Coriolis flowmeters specially designed for dosing applications on a filling machine. Vortex flowmeters measure flow velocity by counting the number of vortices shed per second from a bluff body (Karman vortex street). The vortices are detected by sensing the vibration of the bluff body caused when they are shed. A correctly sized meter provides a linear output, low pressure loss and high accuracy over a wide flow range for low viscosity liquids, steam and gas. When the minimum specified flow rate is reached, however, the device will cut off, that is, the output drops to zero. Modern vortex meters are cost-effective in providing instantaneous flow rate and totalised flow. The device has no moving parts, and inserts directly into the pipeline, making installation and cleaning-in-place straightforward, although it is not suitable for hygienic applications. Increasing numbers are to be found in utility applications, however, particularly for steam-flow measurement. Ultrasonic flowmeters measure flow by either the Doppler method or the time-of-flight (TOF) method. In the Doppler method, the change in frequency that occurs when an ultrasonic pulse is reflected by a bubble or particle provides a measure the velocity of a flowing medium. In the TOF method the difference in the time of propagation between a pulse travelling with and against the flow provides the same information. Doppler flowmeters are seldom found in the food industry. © Woodhead Publishing Limited, 2013

54 Robotics and automation in the food industry TOF ultrasonic flowmeters may be in-line, clamp-on, or the sensors may be inserted into the piping. Accuracy is low in standard applications, but can be increased by using additional sensors to form a multiple path. In this configuration, unsymmetrical and laminar flows can be measured. They are suitable for fluids with very little or no solid content, highly viscous fluids and very low flow rates. Their flexibility makes them useful for trouble-shooting and commissioning. Differential pressure flowmeters measure the pressure drop across a restriction in a pipe, which is proportional to the square root of the fluid velocity (Bernoulli’s law). The restriction is the so-called primary element. A wide variety of primary elements exist, but by far the most common is the orifice plate. The range of a differential pressure flowmeter depends upon the size of the orifice. Turndowns of maximum 4:1 are possible. Wear of the orifice edge will reduce the meter’s absolute accuracy over time, so regular calibration is necessary. The food industry uses differential pressure measurement almost exclusively for flow measurement in service systems, for example, steam or compressed air lines. The overriding factors here are the high operating pressures and temperatures that can be encountered. The technique is popular because of its robustness and simplicity, combined with a wealth of independent data surrounding operation and installation. Thermal mass flowmeters use the heat-convection effect of a moving fluid to measure mass flow. They are normally used for dry-gas measurement, having a wide turndown, up to 100:1, and negligible pressure loss. Mass flow rate is measured directly without need for additional temperature or pressure compensation. In the food industry, they can be used in utility and process gas measurement, including CO2 distribution, nitrogen gas distribution and purge flows, natural gas metering and compressed air. The thermal properties of the gas must be taken into account when specifying the meter. As they can detect extremely low flows, they are also suitable for leak detection. Bulk solid flowmeters There are many applications in the food industry which involve monitoring the transportation of bulk solids to various parts of the plant, for instance coffee beans, salt, potatoes, tea, cereals, fruit and vegetables. The requirement ranges from simple switches to determine that a product is no longer falling from a conveyor belt, to a continuous measurement, for example, the flow of crisps to a flavour drum, to ensure the correct ratio of the flavouring. Microwave barriers, for example, signal either the presence of absence of product between their transmitter and receiver units. A Gunn diode can be used to determine whether a product is moving or not. For continuous measurement, impact plate weighers can be used. Here free-flowing product falls onto a sensing plate that measures its bulk flow. The fall height must be at least 800 mm, and the measuring range is from 30 to 120 kg/h. The alternative is to use a load cell mounted under a conveyer belt. After calibration, such systems are capable of accuracies of 0.25–2%, depending upon design.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 55 3.3.4 Level The primary application of level measurement in the food industry is to monitor contents of storage or process tanks and silos containing raw, intermediate and finished products. In process control, continuous measurements are used to control the filling and emptying of tanks, retorts, reactors, etc. Prevention of overspill and dry running of pumps are typical applications where level switches are used. Table 3.9 lists the level-measuring instruments normally used in food processing. With the exception of ultrasonic transmitters, all are available in hygienic versions. Capacitance level transmitters use the capacitance between the sensing element, a rope or rod, and the tank walls, as a measure of product level. Depending upon design, they may be used for both continuous level measurement and level-limit detection of both conductive liquids and electrically non-conductive materials. For non-conducting liquids, measurement is dependent upon the dielectric constant, and the transmitter must be recalibrated on a change of product. Depending upon the transmitter, it may be possible to self-calibrate by comparing the reading with a point measurement. Bulk solids with grain size up to 30 mm can also be measured. Current applications are restricted to limit detection. The method is suitable for pressures up to 100 bar and in the temperature range –80°C to +400°C (−112°F to +752°F). It is also insensitive to pressure surges caused by, for example, stirrers. Since capacitance transmitters offer a fast response they are particularly suitable for small vessels, for example, for dosing and filling machines. Conductance level switches are used for limit detection in conductive liquids. The sensing element comprises a stainless steel rod that is mounted in the top or the side wall of the vessel. The switching position is determined by the length or the mounting point of the sensor respectively. There is a sudden drop in resistance (or rise in current) when liquid comes into contact with the sensing element signalling that the limit level has been reached. The signal is be used to switch a relay, NPN- or PNP-transistor, or to provide a switching signal to a remote transmitter. The probes are essentially the same as capacitance probes. The operating conditions are also the same, whereby a change in the conductivity of the liquid has no effect on the switching. Hydrostatic level transmitters measure the pressure exerted by a head of liquid, which is proportional to its height. Measurements are made with specially designed differential-pressure or gauge-pressure transmitters. The measurement is dependent on the density, but this can either be entered as a calibration parameter or eliminated by calibrating at the desired empty and full levels. In the case of an open vessel, where atmospheric pressure is acting on the head of liquid, only one measuring point at the bottom of the tank is required, see Fig. 3.6. If the vessel is closed, an additional measurement of head pressure is required. The difference between the two measurements gives the level. For differential pressure transmitters, the level is displayed directly, but the fill fluid in the capillary to the head-pressure measuring point may produce an unfavourable temperature effect. If two gauge-pressure transmitters are used, the head pressure is

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

Point

Vibration

Non-contact Continuous

Radar

Continuous

Non-contact Continuous

Ultrasonic

Guided radar Contact

Continuous

Contact

Hydrostatic

Contact

Solids

x

x

x

x

Continuous, point x

Point

Contact

Capacitance

Measurement

Conductance Contact

Type

Method

x

x

x

x

x

x

x

Liquids

Typical accuracy

Typical operating pressure

0.15–2.5 m to 0.42–10 m ±0.25 to ±1% of −1 to 10 bar 0.5–22 m for switches range to −1 to 100 bar 0.05–15 m – −1 to 10 bar to −1 to 160 bar 100 mbar to 10 bar ±1 to ±2% of – (= maximum ca. 10 m) ideal span 5–70 m ±3 mm or ±2% −0.7 to 2.5 of set range bar to −0.7 to 4 bar 20–70 m ±1 to ±15 mm −1 to 3 bar to −1 to 160 bar 4–45 m ±2 to ±10 mm −1 to 6 bar to −1 to 400 bar – – −1 to 6 bar to −1 to 100 bar

Max. measuring range/ switch length

−40 to 80°C to −60 to 280°C

−20 to 80°C to −196 to 450°C

−40 to 150°C to −60 to 400°C

−10 to 70°C to −10 to 100°C −40 to 80°C to −40 to 150°C

−10 to 100°C to −50 to 400°C

−20 to 80°C to −50 to 400°C

Typical operating temperature

Table 3.9 Properties of level transmitters used in the food industry. Values are guidelines only: for specific values see the manufacturer’s data sheet

Sensors for automated food process control 57

Fig. 3.6 Hydrostatic level measurement on a beer-fermentation tank. (Source: Courtesy of Endress+Hauser.)

available for control and the level is normally calculated by the controller. In the case of fieldbus devices, the value can be downloaded for display at the lower device. Hydrostatic pressure transmitters are the most used level measurement technology in the food industry. They operate continuously at temperatures up to +100°C (+212°F), are insensitive to build up of product and will withstand CIP temperatures of up to +130°C (+266°F) over short periods without permanent damage or the need for readjustment. Foam has no effect on the measurement. For gauge-pressure transmitters, a substantial overpressure range (20×) prevents damage due to pressure surges caused for example, by stirrers. The fill fluid must be suitable for food applications, and the housing requires a rating of IP 65 or above. Ultrasonic transmitters measure level by emitting a sonic pulse and measuring the time it takes to receive the reflection from the surface of the product, the so-called TOF (time of flight). In all cases, the sensor is mounted off centre at the top of the tank, for bulk solids, angled towards the product surface. The transmitter is set up by entering the desired empty and full distances. Where the atmosphere above the product is not air, a sonic velocity parameter must also be entered. It is also good practice to scan the empty tank, so that any spurious echoes from tank fittings can be filtered from the incoming signal.

© Woodhead Publishing Limited, 2013

58 Robotics and automation in the food industry

Fig. 3.7 Guided radar measurement in a flour silo (right). The sensor on the left is an older capacitance probe. (Source: Courtesy of Endress+Hauser.)

Ultrasonic transmitters can be used to measure liquids, provided they are not evaporating or degassing, and bulk solids. In general they can be used for pressures from 0.7 to 4 bar (sensor dependent); the diaphragm can withstand temperatures up to +150°C (+302°F). Measuring ranges up to 70 m are possible. For bulk solids, dusty conditions may temporarily reduce this value by up to 25%. For bulk solids, the surface of the material must reflect a significant proportion of the emitted pulse. An instrument will operate independently of the angle of the filling cone or if the granularity of the surface is greater than 1/4 of the wavelength used (>3–8 mm, depending on sensor). If the granularity is less than 1/6 of the wavelength, then the surface texture comes into play. If the surface has no pronounced texture, it will act as a mirror and the sensor receives an echo which has travelled back via the silo walls. Under these circumstances, level cannot be measured accurately. TOF radar level transmitters measure in exactly the same way as ultrasonic transmitters, but with a radar pulse. In the case of frequency modulated continuous wave (FMCW) radar transmitters, which find use in tank-gauging applications, the phase shift is measured, and this is also proportional to distance travelled. There are two types of TOF transmitter, a so-called ‘free space’ version, in which the pulse is launched from a stub, parabolic, rod or horn antenna, and a ‘guided’ version, in which the pulse travels down a steel rope or rod, see Fig. 3.7. The use of ‘free space’ radar transmitters is normally restricted to liquids that have a dielectric constant ≥1.8, which is the case for most substances encountered

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 59 in the food industry. Guided radar transmitters are suitable for both solids and liquids. Solids must be fine-grained (≤ 20 mm) and like liquids, have a dielectric constant ≥1.8. Powdery solids present no problems. The measurement is also independent of temperature and the pressure, with medium temperatures up to 150°C presenting no problems. Measuring ranges of up to 70 m are attainable in both solids and liquids. The original vibration switch has a sensing element in the form of a tuning fork (other versions use rods) that is forced to vibrate at a set frequency. The frequency changes and a switching signal is output when liquid or solid comes into contact with the fork tines. This signal might switch a relay or transistor output built in the sensor itself or be passed on to a separate switching unit. Typical operating temperatures are −40–280°C (−40–536°F) for the wetted parts and −40–70°C (−40–158°F) for the electronics. Higher temperatures can be withstood for short periods without damage to the sensor, for example, when cleaning or sterilising in place. The method is suitable for liquids having a viscosity of 10 000 cSt or less and for solids with grain size up to 10 mm. It is unaffected by changes in physical properties of the product, including viscosity as well as in process conditions such as turbulence, foam, build-up, gas bubbles or solid suspensions. The switch requires no calibration and maintenance is negligible.

3.3.5 Density In-line density measurements are possible with Coriolis flowmeters and tuned hydrostatic pressure transmitters. New on the market is a continuously measuring vibration transmitter, which is essentially an adaptation of the vibration switch, with a density computer. This is capable of measuring densities between 0.3 and 2.0 g/cm3 in liquids with maximum viscosity of 350 mPa.s and a flow velocity of up to 2 m/s. Process temperature may be in the range 0–+80°C (32–166°F), where +140°C (+194°F) is allowed for CIP, and process pressure from 0 to 25 bar absolute.

3.3.6 Analysis Often the analysis is done in the laboratory, but the demand for in-line measurement has driven manufacturers to develop an increasing number of sensing elements. There has also been some progress on standardised electrical interfaces, allowing the user to select probes independent of manufacturer. Classical in-line measurements are pH/ORP/Redox, conductivity and dissolved oxygen. The associated probes are generally connected to a single evaluating unit, which can be configured to make the required measurement and is easily integrated into a control system. Additional quantities that can be measured are turbidity, colour (concentration) and absorption. pH measurement has the widest range of applications in the food industry, for instance in the production of tomato sauce, mustard, cottage cheese, yoghurt and milk. It is used for quality control, for example, the freshness of milk before it is

© Woodhead Publishing Limited, 2013

60 Robotics and automation in the food industry

Fig. 3.8 pH measurement in a brewery, sensor mounted in a Varivent® housing. (Source: Courtesy of Endress+Hauser.)

processed or mustard before it is bottled. pH may also be used to control and optimise enzyme activity in certain food processes or simply to check that no cleaning agent has leaked into a beverage. Together with conductivity measurement, it is also used in the control of CIP cleaning systems, see Fig. 3.8. Conductivity measurements are also made in the production of various foods, sensors being installed in pipes to detect phase shifts. Turbidity with the associated colour and absorption measurement finds application at many stages in the brewing of beer and making of wine and is also used to detect phase shifts in milk. Dissolved oxygen measurement is used to monitor filter layer discharge, wort aeration and yeast propagation in brewing. Another application is the monitoring of the degassing of water fruit juices and concentrates.

3.4

Device integration

Depending on the extent of automation, devices must be integrated into controllers and in some cases plant asset-management systems. The former allows the process variables to be acquired by the controller, the latter allows the devices to be configured, monitored and diagnosed by a plant asset-management tool. The tool may be part and parcel of the control system or be based on an open technology such as EDDL (Electronic Device Description Language) or FDT (Field Device Tool) and operated in parallel to the controller. In this case a parallel path to the system is often provided by a gateway or multiplexer.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 61 The user must decide between integration over a 4–20 mA current interface or a fieldbus system. There are many different fieldbuses but suppliers of measuring instruments support only a few. These are ones used in process automation rather than factory automation, although some span both applications. This section takes a short look at the features and benefits of each; the main attributes of each bus are to be found in Table 3.10. More information on networks in general can be found in the Instrument Engineers Handbook (Liptak, 2012).

3.4.1 4–20 mA/HART The advantage of the 4–20 mA current interface is that it is simple to install and commission, if somewhat expensive on wiring and cabinet space. Instruments are connected to the controller via I/O cards, which in addition to current, also cater for voltage, pulse/frequency and binary signals, see Fig. 3.9. The controller sees only the process value and a fault signal. The wealth of information contained in a modern measuring instrument, for example additional measured values or diagnostic information, goes to waste. Nowadays, the majority of 4–20 mA devices also support the HART protocol, which is superimposed on the current loop (HCF, 2010). HART devices can be configured and faults diagnosed on a point-to-point basis by using a handheld terminal or a laptop running a FDT-frame program. If HART digital signals are to be used by the controller or a plant asset-management system, a HART card must be used or the 4–20 mA loops must be tapped and the signals read via a multiplexer. It is also possible to use HART as a purely digital multidrop bus, but this is seldom found in practice. HART devices are integrated into the controller and handheld by means of device description files (DDs). For asset management with an FDT frame, device type managers (DTMs) are required.

3.4.2 MODBUS serial/MODBUS TCP MODBUS is a programmable logic controller (PLC) orientated industrial standard for factory automation that also finds use in process automation (MODBUS, 2010). It has an installed base of over 7 million points worldwide and remains popular because little programming effort is required to get it running. Support from manufacturers both on the field and system side is excellent. MODBUS is basically a messaging service that runs on a variety of physical layers. Serial MODBUS, using RS-485 as physical layer, and MODBUS TCP, using Ethernet, are two variants of interest to instrumentation. Serial MODBUS allows the connection of MODBUS devices to a PLC in a bus structure, see Fig. 3.10. Such devices may be instruments such as flowmeters, drives or gateways. A gateway allows the integration of binary, analogue and pulse/frequency as well as HART signals. MODBUS TCP is often used to exchange data between controllers supporting different protocols. It is frequently offered as an interface for recorders or flow computers. As yet no measuring instruments are available with MODBUS TCP, but this may only be a matter of time.

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

Baud rate

Ethernet

Maximum Copper or Maximum 30 fiber 500 m optics

10 Mbit/s 100 Mbit/s

None

CSMA/CD

Decentral

None

Fieldbus Foundation www.fieldbus.org

Modbus users website www. modbus. org Profibus International www.profibus. com

Exi, Exd HART or Exe Communication Foundation www. hartcomm.org

Bus safety Further concept information

Central

Maximum 1200 bit/s 3000 m dependent upon number of devices Maximum 9600 bit/s 1200 m 115 kbit/s

Segment length

PROFIBUS PA IEC 61158, Standard Type 3 certification process for hazardous areas FOUNDATION IEC 61158, Function fieldbus HSE Type 5 blocks for decentralised control

RS-485

Maximum Copper 247

Maximum Copper 16

Medium

Maximum Copper or Maximum 9600 bit/s–10 None 126 fiber 1200 m Mbit/s optics (copper) Several kilometers with optical fibers Master-slave IEC 61158-1 Maximum Copper Maximum 37.5 kbit/s Exi to 32 1900 m FISCO model

Token passing

Master-slave RS-485 Ethernet

FSK

Nodes*

Central

Industrial Simple structure, Central standard widely used

MODBUS

Token passing

Physical layer

PROFIBUS DP IEC 61158, Optimised for Type 3 remote I/O

Industrial Integrates into Central standard existing 4 -.20 mA systems

Processing Medium access

HART

Special features

Standard

Protocol

Table 3.10 Attributes of various control and fieldbus standards

© Woodhead Publishing Limited, 2013

Central

TDMA

Maximum 1900 m

Maximum Copper, Maximum 30† fiber 100 m optics, Maximum wireless 3000 m Maximum 100 m IEEE Maximum Wireless Maximum 802.15.4; 250† 200 m 2006 between devices

Ethernet

Decentral

CSMA/CD

IEC 61158-1 Maximum Copper 32

Central/ Token decentral passing

250 kbit/s

10 Mbit/s 100 Mbit/s 1 Gbit/s

37.5 kbit/s

–

None

Exd or Exi

†

*

Number of nodes that can be physically connected per segment: in some cases the number of logical addressable nodes can far exceed this number. Practical guideline for control networks: the more nodes there are the longer the refresh time of the system.

WirelessHART IEC 62591 Self-organising wireless sensor network

FOUNDATION IEC 61158, Function fieldbus H1 Type 1 blocks for decentralised control Ethernet/IP IEEE 802.3 Spans factory and process automation

HART Communication Foundation www. hartcomm.org

ODVA www.odva. org/

64 Robotics and automation in the food industry Engineering

Maintenance (HART)

SCADA visualisation

FDT EDDDL

OPC

Ethernet

Controller

HART multidrop to HART card (4 mA)

HART mutiplexer

HART digital signal

4–20 mA/HART

Fig. 3.9 Integration of 4–20 mA/HART devices into a control and asset-management system.

Engineering

SCADA visualisation OPC

Modbus TCP

Controller

Remote I/O

Modbus Controller RS-485 Gateway

e.g. Fieldbus

Modbus devices 4–20 mA/HART 4–20 mA, pulse, binary

Fig. 3.10 Integration of field devices in a Modbus system.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 65 The MODBUS protocol exchanges data in a master-slave relationship. Each slave has a unique address and the data are identified by their location in the slave address register. Devices are integrated by simply providing the controller with the slave register where data are to be read or written. The use of an FDT tool with MODBUS depends on whether the associated DTMs are provided by the manufacturer of the controller and slaves. There is some support, but not to the extent of the other fieldbus systems.

3.4.3 PROFIBUS (DP, PA) PROFIBUS is a control and fieldbus system which has a solid installed base in the food industry (PROFIBUS, 2010). In fact, some of the first applications in both Germany and in England were in breweries. It is PLC orientated, with the protocol being basically master–slave. It caters for multi-master systems by supporting a virtual token ring, where each master in turn receives the token, giving it the right to communicate with its slaves. Communication may be cyclic, as used for control, or acyclic, as used for the configuration and diagnosis of field devices. PROFIBUS DP (Decentralised Periphery) addresses the general field of factory and process automation. PROFIBUS PA (Process Automation) fulfils the requirements of the process industries regarding device configuration/diagnosis and use in explosion-hazardous areas. The two versions have different physical layers (RS-485 and IEC 61158-2) but use the same protocol. The baud rates are also different, so that couplers or links must be used to connect the two. Together the two protocols find a tremendous amount of support from control equipment manufacturers, so that in addition to valves and measuring instruments, drives, frequency converters, low voltage switchgear, remote I/Os, etc. are all available. Figure 3.11, left, shows the integration of these components into a PROFIBUS system. GSD files (device database files) and EDDs (electronic device descriptions) are required for integration into the controller. For asset management either EDDs or DTMs are required. PROFIBUS also supports various device profiles, so instruments can always be set up, even if the native files are unavailable.

3.4.4 FOUNDATION fieldbus (HSE, H1) FOUNDATION fieldbus is orientated towards distributed control systems (DCS) and focuses on continuous control (FOUNDATION Fieldbus, 2010). It has made a little headway in the food industry in North and South America, but elsewhere its stronghold is in petrochemicals. For historical reasons FOUNDATION fieldbus H1 has the same physical layer as PROFIBUS, but of course, the protocols are different. The first H1 devices became available in 1998, but it was almost 2000 before the specification of the control layer, FOUNDATION fieldbus HSE (High Speed Ethernet) was completed. Beyond linking devices and controllers, manufacturers have been reluctant to support HSE and it is only in the past couple of years, that the first prototype Remote I/Os and HART or MODBUS interfaces

© Woodhead Publishing Limited, 2013

66 Robotics and automation in the food industry have appeared. The integration of control equipment such as drives and frequency converters is normally done by connecting to other bus systems, for example, MODBUS, see Figs. 3.10 and 3.11, centre. FOUNDATION supports a function block application process, whereby in addition to standard analogue and discrete inputs and outputs, field devices may have logic and control blocks inside them. The bus is driven by a link active scheduler, usually located within a linking device or controller, which allocates time slots to each fieldbus device in which they may publish their data. All devices listen, but act only on data that is relevant to them (publisher-subscriber mechanism). The so-called macro-cycle, that is, the time it takes to execute all the function blocks within a particular fieldbus segment is typically of the order of 300–500 ms. As can be seen, FOUNDATION fieldbus is not suitable for control applications requiring a quick response. Devices are integrated into the system by means of CFF files (Common Format Files); EDD files are required for operation of the devices. Many manufacturers have produced DTMs for their devices, so that they can be operated with an FDT-frame program. Indeed, the recent introduction of a so-called interpreter device type manager (iDTM) means that all devices registered with the Fieldbus Foundation can now be operated by FDT.

3.4.5 EtherNet/IP EtherNet/IP is device network that shares a common application protocol with ControlNet and DeviceNet, two open protocols originally developed by Rockwell Automation (ODVA, 2010). Its installed base is in factory automation, but like PROFIBUS DP it crosses the border to process automation: in North America there are many installations in the Food & Beverage industry. The first 4-wire hygienic flowmeter with Ethernet IP protocol appeared on the market in 2009. This is likely to be followed by analytical transmitters, registration devices and application computers, which basically integrate 4–20 mA/HART devices. EtherNet/IP is supported by the Open DeviceNet Vendors Association, ODVA (ODVA, 2010) which is now also responsible for testing and compliance. EtherNet/IP devices are integrated into the system by EDS (Electronic Data Sheets) files, which are the equivalent of device descriptions in other fieldbus systems. Depending upon the application, the maintenance tool may use FDT or EDS. Figure 3.11, right, shows the integration of these components into a EtherNet/IP system.

3.4.6 Wireless Since the introduction of WirelessHART devices a couple of years ago, local wireless sensor networks have come into focus as a new means of gathering information from remote or moving equipment. Unfortunately, there are rival standards: so-called ISA SP100.11a and a Chinese initiative that is in the process of becoming an international standard. At present, WirelessHART has the most backing

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 67 Engineering GSD (PROFIBUS) CFF (FOUNDATION Fieldbus) EDS (EtherNet/IP)

SCADA visualisation

Configuration and maintenance FDT EDDL EDS

OPC

Ethernet FF HSE (EtherNet) PLC

PLC/DCS

Controller

PROFIBUS DP

EtherNet/IP

FF HSE Remote I/O

Gateway

Devices, drives, MCs Coupler/link

PROFIBUS PA

PROFIBUS

Gateway Devices, drives, MCs

4–20 mA/HART pulse, binary

Up to 4×

4–20 mA/HART pulse, binary

FF H1

FOUNDATION fieldbus

PROFIBUS FF 4–20 mA/HART pulse, binary

EtherNet/IP

Fig. 3.11 Device integration in a fieldbus system. Left to right: PROFIBUS, FOUNDATION fieldbus and EtherNet/IP.

from instrumentation and system manufacturers, but at least two major players have opted for the ISA solution. Despite many similarities, the two are not compatible. There has been a lot of pressure placed on the ISA to migrate towards WirelessHART, but although a committee has been founded with this as its target, at the time of writing there appears to be no particular urgency in their schedule. Figure 3.12 shows the basic architecture for wireless sensor networks. A gateway acts as base station for the network, buffering measured values and making them available to an automation system via an Ethernet or RS-485 interface. For ISA SP100.11a a second possibility of integrating the wireless gateway directly onto the fieldbus network is foreseen. The process data are sent at regular intervals by the devices to the gateway. Configuration and maintenance of a WirelessHART network is possible by FDT, EDDL and depending on gateway, by web server. The major applications for wireless sensor networks are likely to be monitoring and diagnosis. Open-loop control is practical where cycle times are relatively long. Closed-loop control, safety applications and applications requiring continuous measurement are not within the scope of the technology at the current moment.

3.5

Applications of sensors in automated food process control

Process instrumentation is used everywhere in the food industry to monitor and control production processes. Many of these processes have been described in

© Woodhead Publishing Limited, 2013

68 Robotics and automation in the food industry Monitoring

Configuration and maintenance FDT EDDL Web browser

OPC

Ethernet Base station 1

Wireless sensor network 1

Base station 2

Wireless sensor network 2

Fig. 3.12 Integration of a wireless sensor network into a monitoring system.

detail in other books and instrument manufacturers themselves often publish case studies on the internet. For this reason, the three applications described in the following are concerned with aspects of food production which are not concerned with process control, but represent automation possibilities with a great potential for cost saving.

3.5.1 Inventory management Before any food production process can be started, the raw materials it requires must be present. This usually means they are stored on site in tanks and silos. Similarly, the final product is often stored before it is delivered to the customer. Often a significant amount of capital is tied up in stock, so that finding the optimum inventory strategy can significantly reduce costs. Any good strategy depends on information. The more readily available and up to date the information, the more flexibility there is in planning replenishment. Inventory may be managed in different ways: • The vendor manages the inventory of his customer and arranges delivery when the stock falls below a certain level – depending whether the material is inbound or outbound, the vendor may be the raw material supplier or the producer of the finished product. • Inventory is managed in-house by the producer and orders are sent out when stock falls below a certain level. • The management of inventory is trusted to a logistics company which transports the material from one site to another.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 69

Inventory management tool User

Database User intranet

Connection manager

Ethernet/internet connection

Modem connection

Fieldgate

Location A

GSM connection

Fieldgate

Location B

Fieldgate

Location C

Fig. 3.13 Example of an inventory management system.

In all cases data must be collected from various parts of a site as well as from locations at other (customer) sites or even from other countries. Figure 3.13 shows a possible solution. The level measurements made on tanks and silos are acquired by gateways, which may be integrated into the customers’ intranet by means of a modem, Ethernet/Internet or GSM connection. The gateways may be polled or send their information at regular intervals by e-mail to a database. If stock falls below a particular level, an event message is sent which demands immediate action from the managing software. The software itself allows tanks to be managed on a product, group, site or vendor basis and can be installed on site or hosted by an external server. Interfaces are provided to SAP and other logistics programs, so that ordering and subsequent delivery can be automatically invoked. The solution has been successfully used for grain distribution in North America.

3.5.2 Dosing and filling In recent years there has been a noticeable trend to replace mechanical piston and cup-filling machines with electromagnetic and Coriolis flowmeters in dosing applications where the fluid is not sticky or highly viscous. Firstly, they are easier to clean, and secondly, they are more versatile. Today the product or fill amount often changes from batch to batch, but the product must still be dosed exactly. For flow-based dosing, this is done by changing device parameters: piston fillers must be individually adjusted to new dosing volumes. Where products are sold by volume, for example, fruit juices, milk, tomato ketchup, etc., an electromagnetic

© Woodhead Publishing Limited, 2013

70 Robotics and automation in the food industry Pressure transmitter Control valve Pressure medium

24 VDC/0.5 A

Operator panel of dosing system

Flowmeter

24 VDC

4–20 mA e.g. PROFIBUS DP

24 VDC Pulse output, status and power

24 VDC/0.5 A Dosing valve

Machine controller

Fig. 3.14 Schematic diagram of a flowmeter-based dosing system.

flowmeter provides a precise measurement. Where products are sold by weight, for example, salsa, mustard, sweet and sour cream, yoghurt, etc., a Coriolis mass flowmeter does the same job. Here density and temperature are taken into account, so that a change of product has no effect on the dosing precision. In some applications the capacity of the packing material governs the process; in this case level control may be preferable to flow control. One challenge for the user of electrical flowmeters is their set up and integration in the machine control system. This knowledge lies partly with the instrument manufacturer and partly with the machine builder. A solution is shown in Fig. 3.14: a modular dosing control system integrates into the machine controller via a digital interface such as PROFIBUS DP. This provides integrated control, correction and monitoring functions such as air-bubble detection and provides accuracies of 0.2% or better at filling volumes from 50 μL to 5000 mL under normal operating conditions. The achievable accuracy is dependent upon process conditions, for example, medium, pressure and temperature, and the hardware deployed, that is, valves and filling nozzles. The instrument manufacturer is responsible for the filling-control system, the machine manufacturer for the supervisory system controlling the mechanical part of the filling machine (bin transport, moving the filling nozzles, etc.). The system is suitable for both linear and rotary fillers and has found immediate acceptance among machine builders.

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 71

Client server

Half hourly data collection

Energy management software Plant SCADA

Scheduler

Production data

Ethernet PROFIBUS controller

Fieldgate

Pulse inputs

RS485 HART multiplexer

DP/PA coupler PROFIBUS PA

Flow, pressure, temperature

Electricity

Flow, pressure, temperature

Fig. 3.15 Example of an energy management system.

3.5.3 Energy management Generating steam and hot water accounts for 40% of fossil fuels used in industry; compressing air accounts for 10% of electricity. It is not unusual to find leaks accounting for 30% of the demand. Improving energy efficiency in processes using these utilities, therefore, is an obvious way of reducing the utility bill. Optimisation means measurement or as Lord Kelvin is reported to have said: unless you measure the flow of a utility, you cannot manage it. Often there is a need for extra measuring points, so any investment must also be justified by the returns. Unfortunately, instruments do not save energy or money, they merely provide measurement data. It is what is done with the data that is important. In a modern energy-management system, therefore, two more components are required: data-collection devices and software analysing the data. Figure 3.15 shows an open solution for energy monitoring based on 4–20 mA/HART and pulse inputs, whereby similar systems can be built using PROFIBUS and FOUNDATION fieldbus communication. Openness means that the network can be used by other applications, for example, for Plant Asset Management. The instruments installed in the utility system monitor flow, pressure, temperature and electricity consumption. Where an energy value is derived from these quantities, the computing is done in separate module. The instruments and/ or modules are connected to data loggers or remote terminal units (RTUs). These publish the energy and utility consumption data at regular intervals to a client/ server application which stores them in a database. Normally this is done every half-hour, but any frequency is possible. Once the data have been collected, they can be used in a variety of ways. • Import into Microsoft Excel spread sheets for manual energy analysis and report writing.

© Woodhead Publishing Limited, 2013

72 Robotics and automation in the food industry • Transfer to existing SCADA/HMI systems. • Transfer to a packaged energy monitoring and targeting (M&T) tool. An M&T tool provides a wide range of facilities for energy analysis, financial analysis, exception handling, budgeting and reporting on energy data. In addition to monitoring monthly energy consumption it provides the means to correlate this information to a driving influence providing a measure of the efficiency of a process. It is estimated that on average, savings of 5–15% in energy cost are possible when simple measures such as those described are implemented.

3.6

Future trends

Although the sensing elements and associated electronics of a measuring device are always subject to continuous improvement, no radical changes are to be expected in the next few years. The same is true for digital communication: the fieldbus has established itself as a reliable alternative to the 4–20 mA current loop but it will be the market rather than new developments that determines the success or failure of a particular protocol. Plant Asset Management, on the other hand, is attracting a great deal of attention at the moment: since efficient operation also depends to some extent on the ability to plan and manage maintenance activities. Developments can be seen at several levels. At the sensor level, NAMUR recommendation NE107 (NAMUR, 2006) proposes a simplified device-diagnosis scheme that sorts error messages into four classes: ‘maintenance required’, ‘failure’, ‘functional check’ and ‘out of specification’. This allows the user to assess the severity of the alarm before deciding whether the detailed information held by the device is needed. The recommendation has been incorporated into both PROFIBUS and FOUNDATION fieldbus specifications and the first devices with this form of diagnosis are already on the market. As discussed briefly in Section 3.4 there are two open specifications that support device configuration and monitoring, EDDL and FDT. FDT was conceived with asset management in mind and has much broader support, which extends beyond measuring devices and valves. EDDs are part and parcel of every HART, PROFIBUS and FOUNDATION fieldbus instrument. Although FDT will bring out an enhanced specification in the near future, they are also co-operating with the EDDL group to specify a joint FDI (Field Device Integration) standard. This should open the way to easier integration of devices in plant management tools and better support of the solutions already offered, for example, condition monitoring. A step further is the use of a plant asset-management tool to connect to the life-cycle data of a device: certificates, calibrations, order codes, spare parts or, in the case of discontinued lines, pointers to on replacement devices. This is information which, of course, can be manually entered into any database; however, it requires a great deal of effort to do this and is basically duplicating data that is

© Woodhead Publishing Limited, 2013

Sensors for automated food process control 73 held by manufacturer. Opening an internet portal with life-cycle data by simply clicking on a device tag or serial number is a possibility which already exists for some devices. A further click allows the replacement device or spare part to be ordered. This service is likely to be offered by most instrument manufacturers in the next few years. At the time of writing, a concept for remote service management is being tested within in the so-called Future Factory Initiative (SAP, 2010), whereby measuring instruments are enabled as smart devices within an ERP tool. These can then create error tickets, notifications or alerts as well as transfer processing data to the service management application for installed-base configuration, management and maintenance. On detecting a problem, a device sends a notification to the service management application which forwards it to the service department of the user. Here the decision is taken whether the problem can be solved locally, or whether the device manufacturer should be involved. In the latter case, the message is again automatically forwarded to the manufacturer’s service centre, which then accesses the device remotely and proposes appropriate actions, if the fault cannot be remedied by reconfiguration. The result is an accelerated maintenance process which ensures that the best qualified person solves the problem.

3.7

Conclusion

The food industry represents a significant market for instrument manufacturers, who have not been slow to follow the demands made by increasing automation. Nowadays, reliable, hygienic instrumentation, which more often than not can be cleaned in place, is available for most process variables. Developments in the past few years have also led to more in-line analysis instruments, allowing automation of the corresponding processes. Finally, the user has a number of possibilities to integrate intelligent instruments into automation systems, allowing their full potential to be used.

3.8

References

Avesta Sheffield Corrosion Handbook (1994), Stainless steels for the Food Processing Industries. ISBN 91–630-2122-6. Also available on-line as Outokumpu Steel Professional Tool (http://www.outokumpu.com/applications/documents/start.asp) [Accessed 24 September 2010]. Bentley (1998), Handbook of Temperature Measurement Vol. 1: Temperature and Humidity Measurement, Springer; ISBN-10: 9814021091; ISBN-13: 978-9814021098. Berrie (2001), Instrumentation and Sensors for the Food Industry (2nd edition), ed. Kress-Rogers, Brimelow, Chapters 13 and 14, Woodhead Publishing, ISBN 1 85573 560 1. Council of Europe (2002), Guidelines on Metals and Alloys Used as Food Contact Materials. Available from: http://www.bfr.bund.de/cm/216/guidelines_on_metals_and_ alloys_used_as_food_contact_materials.pdf [Accessed 24 September 2010].

© Woodhead Publishing Limited, 2013

74 Robotics and automation in the food industry FDA (2010), Food Contact Substances (FCS), Food and Drug Administration, USA. Available from: http://www.fda.gov/food/foodingredientspackaging/foodcontactsubstancesfcs/default.htm [Accessed 24 September 2010]. FOUNDATION Fieldbus (2010), Fieldbus Foundation. Available from: www.fieldbus.org [Accessed 24 September 2010]. HCF (2010), HART Communication Foundation. Available from: www.hartcomm.org [Accessed 24 September 2010]. IEC (1995); IEC 60584-1 (1995) Ed. 2.0; Thermocouples – Part 1: Reference tables; International Electrotechnical Committee. (2001); IEC 60529 (2001-02) Ed. 2.1 Degrees of protection provided by enclosures (IP Code); International Electrotechnical Committee. (2008), IEC 60 751 (2008) Ed. 2.0, Industrial platinum resistance thermometers and platinum temperature sensors, International Electrotechnical Committee. Liptak (2012), Instrument Engineers Handbook: Vol 3, Process Software and Digital Networks, ed. Lipták, B. G, Eren, H, CRC Press, ISBN-13: 978-1-4398-1776-6 (Hardback). MODBUS (2010), Modbus Organization. Available from: www.modbus.org [Accessed 24 September 2010]. NAMUR (2006), Namur NE 107, Self-Monitoring and Diagnosis of Field Devices, NAMUR (International user association of automation technology in process industries), Available from: http://www.namur.de [Accessed 24 September 2010]. NEMA (2008), NEMA Standard 250, Enclosures for Electrical Equipment (1000 V Maximum), National Electrical Manufacturers Association, USA. ODVA (2010), ODVA. Available from: www.odva.org [Accessed 24 September 2010]. PROFIBUS (2010), Profibus International. Available from: www.profibus.com [Accessed 24 September 2010]. SAP (2010), SAP Future Factory Initiative. Available from: http://www.sap.com/about/ company/research/livinglabs/futurefactory/index.epx [Accessed 24 September 2010]. Stahl (2007), Basics of Explosion Protection, R.STAHL Schaltgeräte GmbH, 74638 Waldenburg, Germany. Available from: http://www.r-stahl.com/fileadmin/Dateien/ explosionsschutz/pdf/grundlagen_en.pdf [Accessed 24 September 2010].

© Woodhead Publishing Limited, 2013

4 Machine vision in the food industry E. R. Davies, Royal Holloway, University of London, UK

DOI: 10.1533/9780857095763.1.75 Abstract: Over the past decade, machine vision has been applied much more widely, uniformly and systematically in the food industry. This has been ensured by continual developments in the constituent methodologies, namely image processing and pattern recognition. At the same time, advances in computer technology have permitted viable implementations to be achieved at lower cost. To some extent, progress is now being held up by the need for tailored development in each application: hence future algorithms will have to be made trainable to a much greater extent than is currently possible. In addition, recent developments such as hyperspectral imaging should become accepted in a number of niche areas. Key words: machine vision, image processing, pattern recognition, object location, food inspection.

4.1

Introduction

Machine vision has long been seen as the way forward to help many parts of industry – including medicine, transport and a good number of other application areas – to monitor and even manage its varied operations (Davies, 2012). In particular, it has been applied to the inspection of products during manufacture, and in the case of food products it has also been important for inspecting the raw food materials arriving from farms and abattoirs (Davies, 2000). Above all, its use is seen as a way of maintaining control of quality, which is all the more exacting when the numbers of individual products arriving at a processing plant amount to millions per week, representing a typical flow of product of 20 items per second. Needless to say, for seeds or cereal grains, the corresponding figures are far greater, and it is more practical to measure flow rates in kg per second. To apply machine vision, images of the various items first have to be obtained, using suitable cameras: these can be normal ‘area’ cameras, not dissimilar to the digital cameras found on the domestic market, but better adapted for continuous

© Woodhead Publishing Limited, 2013

76 Robotics and automation in the food industry computer interfacing. However, with the steadily moving conveyor-belt system found in many plants and factories, it is often beneficial to use a ‘line-scan’ camera to obtain in turn each vertical line of pixels for (effectively) an infinitely long horizontal image of the belt, from which normal 2D images of the products and packages can be obtained. Once suitable 2D images have been acquired, they can be analysed to obtain the required visual information about the products. Much of the processing can be achieved by taking each image and converting it into a succession of other images that refine the data and steadily extract relevant details and measurements (Davies, 2012).1 Indeed, image processing is defined as the conversion of one image into another, in which the same objects occur at the same relative positions, but in successive images they may have noise removed, edges enhanced, edges detected, edges measured, and then whole objects identified and assessed. Alternatively, grey-scale images may be thresholded to obtain binary images representing shapes, and these may then be ‘thinned’ and analysed further. These processes will be described in more detail in later sections of this chapter. Here we note that relatively simple image-processing procedures are able to extract quite high-level information while working only in image space. However, if machine vision were limited to image processing, it would have severely restricted power, and it is necessary to add abstract pattern recognition processes to realise its full potential. Abstract pattern recognition is based on the idea that the computer should memorise both the characteristics of objects (such as their shapes) and their classifications, so that it can perform automatic interpretation whenever it sees such objects again. As we shall see, mere memorisation is virtually impossible because of the vast variations between different instances of any individual type of object, so special methods have to be invoked to perform abstract recognition (Webb, 2002). When this has been achieved, combining the capabilities of image processing and abstract pattern recognition provides the basis for viable machine vision systems. Once suitable vision systems have been devised, they can be used for inspection, quality control, and also to oversee and control ‘assembly’ of more complex foods, such as pizzas and layer cakes. Certain products need X-ray inspection (e.g. to detect bones in chicken fillets), and in general it is necessary to select the image modalities that can provide the most appropriate information about products. Finally, there is a need for real-time processing so that any vision system can keep up with the flow rate on the product line: while it is commonplace for dedicated hardware accelerators to help with the computer processing, we are at last moving into an era where computers will be able to do all the necessary processing unaided. In the following section we review the principles and methods of machine vision that arise from various image-processing and pattern recognition techniques; we also look briefly at available image modalities. In Section 4.3 we present a set of case studies that demonstrate the value of a good proportion of these techniques. In Section 4.4 the discussion is broadened to cover a much wider range 1

This book provides a basic source covering a good many of the underlying concepts, and theory, for this chapter, and especially for Section 4.2.

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 77 of applications and recent advances in the food industry. Section 4.5 appraises the need for special hardware for food inspection applications. The chapter finishes with a conclusion outlining likely future trends.

4.2

Machine vision: principles and methods

Machine vision is a composite subject whose methodology is based in two major disciplines, namely image processing and statistical pattern recognition. In fact, the first of these disciplines is very wide in scope and it is helpful to divide it into a number of smaller but closely linked areas including shape analysis, morphology, object location and texture analysis. Also of concern is the means by which images are acquired before analysis – as additional information is clearly available from X-ray, ultra-violet and infra-red images, as well as from colour. All the above aspects are considered in the following subsections.

4.2.1 Basic image-processing techniques A simple means of performing image processing is to take the pixel value at each pixel location and to place a modified value at the corresponding location in an output image space. For example, in the process known as thresholding, which is widely used for locating clearly defined objects in digital images, any pixel whose intensity is darker than a given threshold leads to a binary value of 1 in the output image space, while other pixels are given the value 0 (Fig. 4.1b). Such a process is called a ‘pixel–pixel’ operation. ‘Window–pixel’ operations are far more powerful and are achieved by examining the pixel values within a window around each pixel location and computing a value to be placed at the corresponding location in the output image space. Amongst the most general and widely used window–pixel operations are convolutions, and for (2k + 1) × (2k + 1) windows, these are defined by the formula: P′′ [

k

k

] ∑ ∑ P[ m

kn

]M [

]

k

[4.1]

where P is the original image, M is the applied convolution mask, and P′ is the output image. Common convolutions of this sort are local averaging to help remove noise, enhancing small holes, enhancing vertical and horizontal edges, and enhancing corners. Within 3 × 3 windows these respective operations can be achieved by applying the following convolution masks: ⎡1 1 1⎤ ⎡ −1 1⎢ 1 1 1 1⎥⎥ , ⎢⎢ −1 9⎢ 8 ⎢⎣1 1 1⎥⎦ ⎢⎣ −11

1⎤ ⎡ −1 0 1⎤ ⎡ 1 1 1 ⎤ ⎡ −5 4 4 ⎤ 1⎢ 1⎢ 1 ⎢ ⎥ ⎥ ⎥ 8 −1⎥ , ⎢ −1 0 1⎥ , ⎢ 0 0 0 ⎥ , −5 −5 4 ⎥⎥ . 3 3 20 ⎢ ⎢⎣ −5 4 4 ⎥⎦ 1 1⎥⎦ ⎢⎣ −1 0 1⎥⎦ ⎢⎣ −1 −1 −1⎥⎦ 1

© Woodhead Publishing Limited, 2013

78 Robotics and automation in the food industry (a)

(b)

(c)

Fig. 4.1 Thresholding and edge detection. (a) Original seed image. (b) Thresholded version, showing that some seeds are not well presented. (c) Edge-detected version, showing that edges are often broken and may be several pixels wide in some places.

Basic properties of these masks are shown in Fig. 4.2. While the second mask tends to give a large signal just inside the boundary of an object (Fig. 4.2c), the signal is boosted at corners – and even more inside holes – where the many boundary contributions accumulate. At first sight the corner enhancement mask (Fig. 4.2f) gives a similar response to that of the horizontal edge mask (Fig. 4.2d); in fact, this is natural as the two 3 × 3 masks are quite similar; however, the righthand corner does exhibit a greater signal in Fig. 4.2f than in Fig. 4.2d. The main fact to observe is that corners are more sizeable features than edges, and therefore require larger masks to identify them unambiguously (typically 7 × 7 pixels rather than 3 × 3).

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 79 (a)

(b)

(c)

(d)

(e)

(f)

Fig. 4.2 (a) Original figure of a square shape. (b)–(f) Effect of applying the five masks described in the text. Panel (b) shows a slight blurring effect that is able to smooth out noise; (c) shows how a Laplacian mask gives a positive signal just inside an edge and a negative signal just outside it; (d) shows a positive edge signal on the right of the object and a negative signal on the left; (e) shows a positive edge signal on the top of the object and a negative signal at the bottom; (f) shows a similar effect to (d) but there is a stronger signal on the right-hand corner. In (c)–(e) the signals are shown on a grey pedestal so that positive and negative signals are easily identifiable.

While convolutions have wide utility and considerable power, Equation [4.1] shows that they are always linear. In fact, non-linear filters are needed to perform useful recognition functions. In particular, for actually detecting enhanced objects and features, operations such as thresholding and non-maximum suppression (i.e. ignoring signals that are smaller than the surrounding ones) need to be applied. Next, we look more carefully at the noise suppression mask given above. This eliminates noise by smoothing it out over neighbouring pixels; in fact, it also smoothes out the image signal. To avoid this effect, median filters are often used. These involve taking the pixel values within the window, placing them in order of size, and taking the middle one in the resulting sequence. For a 3 × 3 window, the fifth of the nine values is taken as the output value; for example, if the nine values are 5, 3, 6, 2, 1, 18, 4, 6, 7, reordering them according to size gives the sequence 1, 2, 3, 4, 5, 6, 6, 7, 18; the middle one (the ‘median’) is 5, and is taken as the output value. Median filters are highly effective at eliminating noise without causing image blurring; in particular, they are especially good at removing ‘impulse’ noise, corresponding to sudden extreme values, such as 18 in the above sequence. In spite of these good qualities, this type of filter can introduce slight distortions by minutely shifting edges, and should therefore not be applied needlessly when accurate measurements are about to be made (Davies, 2012).

© Woodhead Publishing Limited, 2013

80 Robotics and automation in the food industry We now turn to the problem of edge detection. To achieve this we employ the outputs from the vertical and horizontal edge enhancement masks. In fact, these differentiate the image locally in the x and y directions, permitting the x and y components of intensity gradient, gx, gy, to be determined. As gx, and gy indicate how rapidly the intensity is changing along the x and y axes, it is not surprising that this information can be used to find the magnitude and direction of maximum rate of change of intensity. In fact, the magnitude and direction can be calculated from the following equations: g

12

⎡⎣ g x2 + g y2 ⎤⎦

(

θ = arctan g y g x

[4.2]

)

[4.3]

The final process of edge detection is now achieved by thresholding g (Fig. 4.1c). Next, we look more closely at the corner enhancement mask, which only enhances corners of a particular orientation. (The fundamental reason for this is that the mask is a model corner, and moving a model corner over the whole image will only give a close match for corners of similar orientation.) This means that detection of corners of arbitrary orientation will require eight masks of this type, each appropriately orientated. In general this situation will apply when designing detectors for any arbitrary feature, though with larger features, larger more intricate masks will be needed, and greater numbers of masks with different orientations will in principle have to be devised. In fact, designing general feature detectors illustrates the principle, well known in radar, that a ‘matched’ filter must be identical in form with the profile it is to be used to detect, in order to achieve optimum signal-to-noise ratio. However, in image processing there is the caveat that this would give rise to sensitivity to the level of illumination of the background. Hence matched filters used in image processing have to use masks whose coefficients sum to zero. This can readily be understood as follows: if the level of illumination rises by the same amount δ everywhere in the image, then the increase in response from the whole mask will be δ times the sum of the mask coefficients: and for a zero-sum mask the overall change in response will necessarily be zero. This applies for the last four of the convolution masks listed above. (The first mask is for suppressing noise and is not intended to emulate a matched filter.) To summarise, convolutions are highly important image-processing operations as they provide optimal detection performance for a variety of image features, but to realise the detection itself, which results in the actual recognition of the features, their outputs need to be subject to non-linear filters including thresholding and non-maximum suppression. In fact, non-maximum suppression is more important than might be thought, because thresholding depends on the reliability with which global thresholds can be selected: in many cases selection is difficult because it relies largely on there being a clear demarcation between true features

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 81 and noise; in addition, while effective local thresholds can often be selected, global thresholds will vary with the level of illumination and the contrast, both of which are liable to vary significantly from one part of the image to another. Much effort has been devoted to these problems, not merely in the past, but also right up to the present day (Davies, 2007). That this is still a hot problem, when the eye can discern almost all types of feature and object without trouble, is down to the very nature of image data, which is incredibly variegated, with many aspects that are confusing for the computer interpretation system. It will serve to highlight this situation if we here refer to one such problem – that of secondary (reflective) lighting which changes the illumination of a given object when another object is placed nearby.

4.2.2 Shape analysis Shape analysis is one of the prime means by which humans can distinguish and recognise objects: it is no less important for computer recognition. Here we assume that any input images have been thresholded to produce binary images in which the objects appear as 1s in a background of 0s, as indicated in Section 4.2. At this stage the computer will have to determine how many objects there are in an image and exactly what regions they cover. While all this is ‘obvious’ to a human examining the image, the computer has to determine it by ‘connected-components analysis’. This is not a trivial process. When scanning over the image in a forward raster scan (that used on an analogue TV screen), first one object is met, then another, then the first one again, so labelling the pixels in sequence will not give objects unique sets of labels (Fig. 4.3). While it is often found that combinations of scans (e.g. forward raster scans followed by reverse raster scans) can help to arrive at a unique labelling, it is more reliable and more efficient to examine the first set of labels and to make a ‘clash table’ showing which labels coexist. Careful iterative analysis of the clash table will then show how to achieve an ideal labelling. Once this has been carried out, the objects have necessarily been counted, and their areas, circumferences and linear measurements can be tabulated without difficulty. Clearly, connected-components analysis makes it possible for many further size and shape measurements to be made and recorded. If objects are to be sorted quickly, simple shape measurements that are invariant to size can be useful. Prime amongst these is the ‘circularity’ measure, C = A/P2, where A is the area and P is the perimeter of a blob: C is largest for a circle and smallest for a thin stick. C is widely used for cell counting in biological samples, but is also useful for sorting small objects such as seeds. While not a sophisticated measure, it is useful for making a preliminary identification of the objects in an image, preparatory to a full assay of any doubtful cases, using more advanced and accurate tools. One such tool is that of moments, in which the first, second, third and higher moments of any complex shape are computed: as this can be done to a high degree of accuracy, it can leave little doubt as to the type of object that is being tested – so long as a large enough database of values is available. However, this method is somewhat computationally intensive, so boundary-tracking methods

© Woodhead Publishing Limited, 2013

82 Robotics and automation in the food industry

Fig. 4.3 Result of a basic algorithm for connected-components analysis. In several cases objects have more than one label: analysis of label clashes is required in order to give a unique labelling.

that focus on object boundary points rather than all the points in the object shape are generally preferred. The archetypical boundary-tracking representation is the ‘centroidal-profile’ method, which describes the object boundary in terms of a (r, θ) polar graph relative to the centroid location (Fig. 4.4). For a circle this is a simple graph consisting of a horizontal straight line at r = a, where a is the circle radius. For a square it is a set of four curves meeting at four cusps, which represents the four corners of the square; the orientation of the square is given by the angles θ at which the cusps occur. Clearly, such a graph will indicate the object’s size and orientation, and will also permit any shape distortions to be detected and measured. Likewise it will permit the object to be recognised. The same situation applies for any of the objects that have been included in a suitable database. Unfortunately, this method is not very robust: for example, if two objects touch or overlap or if there is one object that is seriously misshapen, the centroid will be in an arbitrary location; as a result, the centroidal profile will be grossly distorted and will not be recognisable (Fig. 4.4). While there are alternative types of boundary profile, such as curvature–boundary distance (κ, s) curves, which are not reliant on accurate location of the centroid, these are clumsier to handle, and less robust, than the Hough transform approach, which we treat in Section 4.4. Skeletonisation is a further approach to shape recognition. The idea is at its most natural when dealing with objects that are composed primarily of narrow strips and curves – particularly alphanumeric characters, asbestos fibres, branches in a tree, and so on. In such cases, by eroding the boundaries of the objects repeatedly, the aim is to reach the innermost ‘medial’ lines, which in the case of

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 83 (a)

(b) r

0

2π

Fig. 4.4 Centroidal-profile method for object-shape analysis. (a) An original object. (b) Centroidal profile, which poses undue problems of interpretation.

Fig. 4.5 The skeleton concept. Here the skeleton is intended to show the path the pen travelled while drawing the shapes.

hand-drawn characters would correspond to the paths followed by the pen (Fig. 4.5). However, the skeleton concept is very closely defined. It is meant to represent the shape of the connected object, and must never become disconnected during the erosion process. This makes algorithm design quite intricate and exacting. Once formed, the skeleton retains all the limbs and holes of the original shape; indeed, it is intended to be a topologically exact description, with all the junctions and line ends giving a useful, idealised width-independent description of the shape. It is a widely used technique, but has the disadvantage that if noise in the original image leads to additional holes in the object shape, a misleading skeleton will result. However, if excessive noise can be eliminated, errors from this cause will be substantially reduced.

4.2.3 Morphology While intrinsically morphology means much the same as shape analysis, it grew out of a different approach – that of the application of simple window operations to modify shapes so that they could be filtered to locate particular structures. The starting point was that of eroding objects so that the outermost layer of pixels was eliminated, or dilating them, so that an additional boundary layer was added (Fig. 4.6): both methods could be applied isotropically or directionally, for example, parallel to the image x-axis. The purpose of these operations becomes clear if a set of vertical striations on a smooth surface is to be located. Eroding the image horizontally would eliminate the striations, and then dilating the image would return the image to its original state, but without the striations. Then, subtracting the

© Woodhead Publishing Limited, 2013

84 Robotics and automation in the food industry (a)

(b)

Fig. 4.6 Erosion and dilation. (a) Image of Fig. 4.1b eroded 4 times by a 3 × 3 erosion operator. (b) The same image dilated twice by a 3 × 3 dilation operator. Erosions of type (a) are useful for separating objects so that they can be counted; they also have the effect of removing tiny objects (in this case tiny flakes of extraneous vegetable matter (EVM)). Dilations of type (b) can be useful for showing regions where objects appear at high density.

final image from the original image would reveal the striations that had been eliminated. Such a method would for example reveal any scratches on a computer disc or polished metal surface. Figure 4.7 shows results in a similar but more exacting case – that of scratches on a wood surface. Although the algorithm works by eliminating the scratches and then subtracting so as to recover them selectively, the overall effect is to filter out most of the (horizontal) wood-grain markings, thereby isolating the vertical scratches. Morphology is formulated mathematically, because carrying out morphological operations efficiently requires them to be broken down into basic operations and then analysed to determine the best way of assembling them into useful procedures. Thus the process of dilating image A using a (possibly directional) mask B is written as A ⊕ B, while the process of eroding A using mask B is written as A ⊖ B. In accordance with the types of processing indicated above, dilation followed by erosion is described as ‘closing’, represented by ‘•’, and erosion followed by dilation is described as ‘opening’, represented by ‘○’. These operations are written respectively as: A • B = (A ⊕ B) ⊖ B

© Woodhead Publishing Limited, 2013

[4.4]

Machine vision in the food industry 85 (a)

(b)

(c)

Fig. 4.7 Use of morphology for scratch detection. (a) Original image of a scratched wood surface. (b) Removal of scratches by a horizontal closing algorithm. (c) Using subtraction to recover the scratches.

A ○ B = (A ⊖ B) ⊕ B

[4.5]

With this notation, we can immediately write the process of identifying the striations described above as the following: S=A–A○B

[4.6]

While this approach shows how one sort of fault can be located, finding cracks in an egg might well require an operation of the following type: T=A•B–A

[4.7]

© Woodhead Publishing Limited, 2013

86 Robotics and automation in the food industry This amounts to filling in the cracks and then comparing with the original image in order to accurately locate them. Space does not permit a full exploration of the mathematics underlying morphology. However, we can write down two rules that give further insight into the rigour it imposes: (A ⊕ B) ⊕ C = (A ⊕ C) ⊕ B

[4.8]

(A • B) • B = A • B

[4.9]

These can be construed as ‘it doesn’t matter in what order two dilation operations are carried out’ and ‘once a hole has been filled in, it remains filled in, so repeated closing is wasteful and unnecessary’. To make best use of morphology, the sizes and directionalities of the basic erosion and dilation operations need to be selected carefully. It should also be noted that grey-scale versions of these basic operations have been devised, some of these being as simple as maximum and minimum operations within pixel windows. Overall, morphology gives us an additional set of tools that are capable of efficiently locating a range of features in digital images, including particularly defects such as holes, concavities, cracks, spots, prominences and hairs – and also scratches, as mentioned earlier (Davies, 2012, chapter 7).

4.2.4 Efficient object location In Section 4.2.2 we considered the efficient location of small features, typified by edge, hole and corner points, using the matched filter paradigm. Unfortunately, it is difficult to scale up this method to locate large objects, (a) because of variations in illumination across the objects, (b) because variegated backgrounds will cause different parts of their boundaries to suffer low or even inverted contrast, and (c) because of the increased likelihood of partial obscuration or distortion. Hence object location methods have to be designed to be much more robust. (The need for this was already alluded to in Section 4.2.2 in the context of the centroidal-profile technique.) To achieve a high level of robustness, a radical new approach is needed: it becomes necessary to infer the presence of objects from partial evidence. The Hough transform approach achieves this by a voting scheme, in which only positive evidence for objects is counted. The simplest example of this is the detection of circular objects. All edge points in the image are made to vote for a circle centre by moving a distance equal to the supposed radius along the edge normal and casting a vote in a secondary image space called a parameter space (Fig. 4.8). If only p pixels are found on the circle boundary, p votes will still be cast at the circle centre, and a peak of this weight will be found at that position in the parameter space. Hence the circle can be found robustly, in spite of the fact that much of its boundary may be misplaced (Fig. 4.9) (Davies, 1984). In addition, the problem

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 87

Fig 4.8 Principle of the Hough transform. Here a partial circle is located robustly: irrelevant votes are scattered and do not interfere with object location.

Fig. 4.9 Use of Hough transform for biscuit location. The broken biscuit and the partial occlusion of another do not prevent any of the biscuits from being located reliably. (Source: Reproduced from Davies (1984) with permission; © IFS International Ltd, 1984.)

of unknown circle radius may be solved in various ways, for example, by trying a sequence of radius values. The method can be extended to allow straight lines and ellipses to be located, or other shapes composed of these shapes. It can also be used to locate arbitrary shapes as long as these can be specified accurately. As indicated above, the power of the approach depends on accumulating positive evidence: erroneous evidence needs to be ignored as it would only serve to bias the results. The method known as RANSAC works rather differently. Sets of data points are used to generate hypotheses and these are each tested to determine which gives greatest agreement with the image data. Despite its different strategy, it actually embodies a voting scheme, seeking the solution giving the greatest numbers of votes. A quite different approach called graph matching has frequently been employed when the objects being sought are characterised by sets of point

© Woodhead Publishing Limited, 2013

88 Robotics and automation in the food industry features such as corners or holes – as in the case of hinges, brackets and certain types of biscuit. In such cases it is necessary to search for subsets of the observed point features that exactly match with subsets in the object template. The solution that is taken to be the most probable one is the largest subset that gives an exact match. Here again, we see that this is a type of voting scheme, in which the best solution is the one for which the greatest number of votes are recorded. Unfortunately, applying this idea rigorously incurs a computational load that is exponential in the number of features to be matched. This is because the definition of a subset is quite general and so all subsets of all sizes and compositions have to be tried. Exponential problems such as this have the property of requiring trivial amounts of computing for say three model features being sought in an image containing five features (in which case there are just ten possibilities to be checked), but the amount of computing escalates disproportionately when six model features have to be sought in an image containing ten features (in which case there are 210 possibilities to be checked), while for nine model features in an image containing 15 features there are a massive 5005 possibilities to be checked. Fortunately, for 2D interpretation of biscuit images, this problem can be solved elegantly using a special form of the Hough transform, as shown in Section 4.3.2. The reader is referred to Davies (2012) for further details and discussion of the three techniques (Hough transform, RANSAC and graph matching) outlined above.

4.2.5 Texture analysis Object surfaces are said to have texture when they are rough, woven or composed of many small particles or strands. In such cases they acquire characteristic intensity patterns that have varying degrees of regularity and randomness. For the more regular patterns, periodicities will be measurable, but they may vary considerably over the surface; they will also exhibit strong directionality, which may be manifest in several directions – as for the weave of a fabric. A pile or sand or seeds will have a much more random texture than a fabric, and will not be directional (Davies, 2012). Because of the degree of randomness that is present in most textures, measuring texture patterns presents problems of statistics, so it becomes necessary to average over a sufficient area in order to arrive at an accurate assessment of the surface texture. This is important for the purpose of recognition, and for demarcation of object boundaries – and also for the location of any blemishes, defects or foreign bodies. Getting a unique characterisation of a texture can involve considerable computation, as each pixel must be accessed many times in order to measure textural coherence over different distances and directions. However, short-cuts may be acceptable in certain circumstances. For example, on a hypothetical ‘bixit’ line, only bixits should be present, so simple tests for (a) presence of a bixit, (b) presence of a non-bixit, (c) presence of a faulty bixit may be sufficient. Perhaps the simplest test is that for ‘busyness’, which may be measured as the number of edge

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 89 (a)

(b)

Fig. 4.10 Basic texture detection using the ‘busyness’ concept. (a) An original image of rape and charlock seeds. (b) Light highlights in the seeds have been located and are indicated by white dots. If more than about 20 white dots appear in the vicinity of the object centres (shown by white crosses), the corresponding seed is very reliably classified as a rape seed.

points, or the number of tiny spots per unit area (spots can be detected using the second mask in Section 4.2.2) (Fig. 4.10).

4.2.6 Statistical pattern recognition As indicated in Section 4.1, image processing alone will not permit reliable judgements to be made about classes of objects. However, it will permit the aggregation of data that can then be used in a complete machine vision system that is capable of making useful decisions. Here it will be important not only to classify and sort objects into one of many categories, or to place products into rejection bins, but also to know the ‘false positive’ (FP) and the ‘false negative’ (FN) rates, so that the classifier system can be optimised. It should be noted that there will be costs associated with erroneous decisions and it will usually be necessary to minimise these costs rather than the errors themselves. Such considerations are crucial when trying to eliminate poisonous moulds, such as ergot for example. To achieve all this, an abstract pattern recognition scheme will be needed. One of the simplest such schemes is the so-called ‘nearest neighbour’ method, which may be implemented by setting up a ‘training set’, consisting of a database of instances of each class, including the feature measurements that have been made on them. Then, during testing, the features of a test pattern are compared with

© Woodhead Publishing Limited, 2013

90 Robotics and automation in the food industry those of each training pattern, and the class of the pattern giving the closest match is assigned to that test pattern. This method is quite computationally intensive, because of the need to make comparisons with a very large number of stored training set patterns when testing. However, if a sufficiently large training set is used, the method has the advantage of a close to ideal (Bayes) error rate. There are many methods giving better performance than the nearest neighbour method outlined above. However, whichever is used, it will be necessary to arrange the best tradeoff between the numbers of FPs and FNs. This can be regarded (Fawcett, 2006) as obtaining the best balance between S, the sensitivity of detection (also called ‘recall’), and D, the discriminability against other types of object (also called ‘precision’), where: S=

TP TP + FN

[4.10]

D=

TP TP + FP

[4.11]

In each case TP is the number of ‘true positives’. To obtain the correct balance, it is useful to optimise the ‘F-measure’, for which the parameter γ is set appropriately, 0.5 often being a suitable value: F=

γD

SD (1 γ ) S

[4.12]

To ensure setting up the system correctly, it is useful to plot the ‘receiver operating characteristic’ (ROC), which is a plot of the true positive versus the FP rate (Fig. 4.11). However, it can be difficult to do this rigorously, if at all, when attempting tasks such as the optimum detection of rare defects. 4.2.7 The variety of image modalities So far, the discussion has concentrated on the analysis of grey-scale and binary images. In fact, modern cameras normally have colour acquisition capabilities and this can be immensely useful, though it tends to involve more complex processing techniques. Here RGB input is usual, but it is frequently useful to convert RGB images into other formats such as HSI (H is the ‘hue’, which is the main colour component, S is the ‘saturation’ or degree of colour vis-à-vis whiteness, and I is the intensity) before proceeding with detailed image analysis. The particular advantage offered by HSI is that the effects of any random variations in illumination will be concentrated in the I channel and hardly echoed in the S and H channels. This means that when observing apples, for example, the H channel should indicate unambiguously whether green or red patches are being observed.

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 91 1.0

PFP 0.5

0.0 0.0

0.5 PFN

1.0

Fig. 4.11 ROC. This graph of probability of false positives (PFP) against false negatives (PFN) is an example of an ROC curve. A typical working point is the position where PFP = PFN.

For some purposes, infrared imaging or thermal imaging can be useful. The latter applies to the detection of live objects such as insects or slugs, while the former can be useful for identifying the emission bands of certain molecules in vegetation or flesh. Ultraviolet light can also indicate traces from live matter, for example, fingerprints, rat urine, ergot and moulds that produce aflatoxin. While multispectral image data, covering visible light, infrared and the microwave regions, is frequently available from satellite data, this modality may be too complex for most food applications, but ways of combining the outputs of the various channels to obtain coherent information, for example, using principal components analysis (PCA), are highly effective. The other modality that is of immense value in the food industry is that of X-rays. Its penetrating powers are well known, but the prospect of radiation damage to the human body brings with it safety problems and the requirement for heavy screening. The result is dedicated machinery that is heavy, costly and non-portable. However, in recent years vast improvements have been made to semiconductor linear array detectors and at the same time dual-energy detection (DEXA) has been achieved and is in wide use. The value of DEXA technology is that by taking the ratios of the dual-energy detector outputs, the thickness of a specimen can largely be cancelled out. Hence careful analysis can reveal more detailed structure in the X-ray images; thus detection of foreign bodies becomes much more reliable. Producing and processing X-ray images are rather specialised topics and cannot be covered in full here. For further details and guidance the reader is referred to Batchelor et al. (2004) and Kröger et al. (2006).

© Woodhead Publishing Limited, 2013

92 Robotics and automation in the food industry (a)

(b)

Fig. 4.12 Chocolate cakelets and mode of illumination. With this mode of illumination, a dark ring appears above the edge of jam positions. (a) Top view of cakelet. (b) Side view of cakelet showing profile and directions of illumination and reflected light. (Source: Reproduced from Davies (2001) with permission; © Éditions Cépaduès, 2001.)

4.3

Applications and case studies

The case studies below relate to baked product and cereal grain inspection, and are important in illustrating a wide variety of image processing and machine vision techniques, including a fair proportion of those covered in Section 4.2. In particular, good use is made of edge detection, Hough transforms, texture analysis, morphology, thresholding, matched filtering, median filtering and abstract pattern recognition methods, while the necessity for robustness is underlined. The coverage of applications will be broadened well beyond the confines of these case studies by the review in Section 4.4, though many of the same principles will be kept in sight.

4.3.1 Case study: chocolate cakelet inspection Many modern foods have high added value relative to the original raw product – or even relative to the basic cooked product. This applies particularly to biscuits and cakes, and in some cases even small cakes are marvels of innovative manufacture. Here we consider the inspection of a certain type of cakelet, made with a round sponge base, to the centre of which a spot of jam has been added, and the top of the product coated with chocolate (Davies, 2000). Chocolate is an expensive commodity, and hence it is important to control the amount of it that reaches the product: too much, and the cost will be too high; too little, and the layer of chocolate will be broken and the consumer will not be satisfied. In addition, the spot of jam must not spread too far away from the centre of the product, or else the characteristic dark ring that appears in the chocolate just above the edge of the jam will be absent, misshapen or distorted, and the final product will lose its pleasing symmetry (Fig. 4.12). Hence, while inspection is generally concerned predominantly with defects or contaminants, here the appearance of the product has also become important. While it is difficult to make a computer think like a

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 93 human, the following computer-implementable rules can be laid down for acceptability with this type of cakelet: 1. The product should be round. 2. The diameter should lie within well-defined limits or the packing machine will be liable to jam. 3. The spot of jam should be central. 4. The spot of jam should be round and of the right diameter. 5. The chocolate should cover the whole of the top of the cakelet. 6. The chocolate should have a clear texture (imparted to it by the dipping grid). 7. The cakelet should be the right way up. 8. The cakelet should not be overlapped by, or even touching other cakelets, as this means it will almost certainly be stuck to them. Checking all these rules for each product at the required rate of some 20 products per second is computationally demanding. It is also a costly enterprise, because cameras, lighting systems and rejection machinery must be linked carefully to the product line. In addition, it is questionable where to place the inspection station. Placed at the end of the line, it can successfully monitor the overall quality of the product, but this will be too late to locate early faults such as lack of jam, so value added at a later stage (including expensive chocolate) will end up being wasted (Davies, 2000). So there is something to be gained by having an early inspection station as well as a final one, but this is seldom done for food products. (However, the reverse situation applies for bread, where the raw cereals are more likely to be scrutinised than the final product.) The first stage of inspection is to locate the product, and in this case the Hough transform is employed and works well. In particular, it is robust to products that are stuck together and to smudges of chocolate on the conveyor. Once the product has been located, a circular region of interest is defined around it. Then the degree of chocolate cover can be assessed highly accurately and the overall diameter can be measured. Next, the surface can be measured using an approach known as a radial intensity histogram, which contains information on the radial distribution of intensities in the texture of the chocolate (Davies, 1984). This permits the general appearance of the product to be assessed; it also permits the dark ring around the edge of the jam region to be measured for size and uniformity. With these approaches, all the rules listed above can be tested and the product flagged as ‘acceptable’ or ‘reject’. It is important to note that it is the line manager’s job to define the limits that have to be set on any one day: it is the machine’s job to implement his wishes using the measurements made using the vision system. Finally, some remarks about the lighting are in order. This had to be set up using four lights that were symmetrically placed around the camera, so as to provide reasonably uniform lighting over the product: great care over this was not necessary to the operation of the Hough transform, which is highly robust. However, reasonable care was still required in setting up the lighting (Davies,

© Woodhead Publishing Limited, 2013

94 Robotics and automation in the food industry (a)

(b)

(c)

Fig. 4.13 Cream biscuit inspection. (a) Ideal appearance of cream biscuit. (b) Appearance when the two wafers are misaligned. (c) Appearance when cream is oozing out from between the wafers. (Source: Reproduced from Davies (2001) with permission; © Éditions Cépaduès, 2001.)

2000), so as to minimise the number of separate measurements to be made and so that simple averaging procedures could be used to save computation.

4.3.2 Case study: cream biscuit inspection Cream biscuits present a challenge for inspection because they consist of two wafer biscuits between which a layer of cream is sandwiched: exact alignment of the wafers is relatively unlikely, so from overhead the biscuit will not have an exact rectangular shape but one in which each side is composed from different parts of the upper and lower wafers. It would be better if the vision system could normalise itself on the upper wafer, so that any deviation between it and the ideal shape could be attributed to misalignment of the lower wafer and/or to cream oozing out from between the two wafers (Fig. 4.13). As for the cakelets already described, either of these two eventualities will lead to the product losing its symmetry and being less attractive to the consumer. Again, there are potential problems of packing machines becoming jammed. In order to locate the upper wafer, location of the boundary will be hindered by either of the two deficiencies mentioned above, so it will be useful to locate it from its surface pattern. In this context it is useful to locate it from the ‘docker’ holes it possesses (Davies, 2000, 2001). As any of these may be rendered invisible by excess cream, and as additional holes or points on the boundary can sometimes appear to be docker holes, all potential docker holes will have to be considered for matching to the ideal wafer template. While this can be carried out using the graph matching approach outlined in Section 4.2.4, a special version of the Hough

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 95

L

Fig. 4.14 Use of Hough transform for locating biscuit centre. For the two ‘docker’ holes located on the left of this biscuit, two potential centre locations are identified and votes are cast at these positions in parameter space. After all such votes have been cast, peaks show the location L of the true centre. (Source: Reproduced from Davies (2001) with permission; © Éditions Cépaduès, 2001.)

transform was eventually used to locate the centre of the wafer and its orientation. The basic concept was to use each pair of features to predict the position of the centre of the wafer and to place a vote at that location in parameter space (Fig. 4.14). By this means wafer location was determined accurately and robustly and also more efficiently than by graph matching. (In fact, there is a break-even point for this computation: this works out at about five docker holes, and for eight or more docker holes the Hough transform approach is hugely more rapid in operation.)

4.3.3 Case study: location of non-insect contaminants in cereals In this and the following section we examine an important application of inspection – that of cereal grain inspection. There are several aspects that need to be assessed: (a) variety, (b) degree of contamination by other varieties, (c) quality, (d) presence of damaged or diseased grains, and (e) presence of contaminants such as insects, rodent droppings and ergot. In the space available we will concentrate on the detection of some of the most important of these contaminants. Many of the contaminants that can appear amongst wheat grains, such as insects, rodent droppings and ergot are significantly darker than the rather grey wheat grains considered here. To the human eye this is what appears to make these contaminants recognisable; thus the computational task seems at first to be a rather trivial application of thresholding. However, tests show that this surmise is far from the truth, first because of shadows between the grains and second because of dark patches on the grains themselves, while the presence of chaff, rape seeds and other natural substances, commonly known as extraneous vegetable matter (EVM), complicates the situation further. First we attend to the detection of rodent droppings and ergot, all of which are dark and have characteristic elongated shapes, while rape seeds are relatively small, dark and round. It turns out that detection is made more complicated

© Woodhead Publishing Limited, 2013

96 Robotics and automation in the food industry because rodent droppings are often speckled, while ergot is shiny and often exhibits highlights, which appear as small light patches. A morphological approach Although thresholding did not prove to be a complete solution to the problem, it presented a useful starting point, the main difficulty being the shadows between the grains. Thus it seemed best to apply erosion operations to eliminate the shadows, and then to apply dilation operations to restore the shapes of the contaminants (Davies et al., 1998). While successful in removing the shadows, this procedure was made worse by the problem of light patches due to speckle and highlights within the contaminant shapes (Fig. 4.15c). Thus a closing rather than an opening type of operation seemed to be needed to ‘consolidate’ the contaminants before eliminating the shadows. However, tests showed that this procedure merely served to make the shadow problem worse (Fig. 4.15d and 4.15e). What was needed to aid recognition of the contaminants was a way of (a) preserving the overall shapes, (b) eliminating the light patches within the shapes, and (c) eliminating the shadows around the shapes – a highly exacting task. Eventually it was found that applying median filters in unusually large windows was able to achieve all this (Fig. 4.15f). The reason this was successful is that median filtering eliminates minority intensity values, whether dark or light, and thus tackles both (b) and (c): curiously it even eliminates both at the same time as it tries to find the central intensity value, and in fact achieves a best balance when minority dark and light pixels are present in equal measure. Nevertheless, some overall expansion of the contaminant shapes occurs, because of the high density of shadows around the grains, but it was found that this could readily be offset by a final erosion operation operating in a small window (Fig. 4.15g). To ensure that ad hoc and thereby possibly deleterious processes were not introduced, the theory underlying this procedure was developed and the latter was shown to be a necessary and quantifiable inclusion in the overall algorithm. Using this approach, good segmentation of the contaminants was achieved, and good preservation of their overall shapes, thereby making the subsequent recognition task more straightforward. In this case, final recognition has to be passed to an abstract pattern recognition system in order to discriminate between rat droppings, mouse droppings, ergot, large beetles, rape seeds and EVM. However, it was found that there was little need for use of a nearest neighbour or similar classifier, as a simple rule-based scheme was sufficiently accurate to deal with most of the possibilities. (Essentially, a rule-based classifier is organised as a tree, in which the computer starts at the root and makes successive decisions about which branch to take until it gets to the final solution in each case.) The main disadvantage of the above method is the need to apply a large median filter. In fact, there are many ways of speeding up median operations: in this case, by starting with thresholding, finding the median merely involves counting dark and light pixels within the chosen window, and this can be carried out very quickly. An interesting lesson is that, like dilation and erosion, median filtering is a morphological operation, though it is not often recorded as such. This may be

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 97 (a)

(b)

(c)

Fig. 4.15 Use of morphology to recover rat dropping shapes. (a) Original image of rat droppings amongst wheat grains. (b) Thresholded version of (a). (c) Result of erosion + dilation on (b). (d) Result of dilation + erosion on (b). (e) Result of erosion on (d). (f) Result of applying an 11 × 11 median filter on (b). (g) Result of erosion on (f). All erosions and dilations take place within a 7 × 7 window. (Continued)

© Woodhead Publishing Limited, 2013

98 Robotics and automation in the food industry (d)

(e)

(f)

(g)

Fig. 4.15 Continued

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 99 due partly to its slow speed making it less popular, though with modern computers, this is nowadays a diminishing disadvantage. A thresholding approach Further work in this area has shown that under some lighting conditions ergot can be located reliably by thresholding. As indicated in Section 4.2.1, thresholding is subject to two major problems: one is the difficulty of finding a suitable global threshold, and the other is the questionable validity of any global threshold under poor lighting conditions. However, the latter difficulty can often be solved by care in the setup of an inspection system (particularly by eliminating extraneous natural lighting), so finding a suitable threshold becomes the main difficulty. In part this is due to the fact that any contaminant forms a very small proportion of the area of an image, so its effect on the intensity histogram used to determine the thresholding level becomes minimal. However, a new ‘global valley’ approach that is applicable to ergot detection has shown how this may be achieved (Davies, 2007). The global valley approach involves searching for the valley in the intensity histogram that gives maximum response to a function f dependent on the depth dl and dr of the valley when measured respectively on its left and on its right: the most suitable formula for f was found to be the product of dl and dr. This makes it highly sensitive to tiny peaks near the left (dark) end of the intensity histogram, and thus provides a close to perfect threshold for discriminating ergot (Fig. 4.16).

4.3.4 Case study: location of insects in cereals In spite of the undesirability of rodent droppings, and the poisonous nature of ergot (whose incidence has decreased markedly in recent years), insects pose an even greater threat to the cereals market: this is because insects have a breeding cycle that is measured in weeks, and thus the integrity of a store or shipment of grain can quickly be compromised. Hence it is important to detect insect infestations at the earliest opportunity. The problem is exacerbated by the huge numbers of grains that are involved and the speeds with which they need to be checked: note, for example, that 30 tonne lorries may arrive at mills or shipping depots at intervals ranging from 3 to 20 min; while even a 3 kg sample will contains some 60 000 grains. Thus devising effective inspection algorithms involves substantial computational problems (Davies et al., 2002; Ridgway et al., 2002). In this case thresholding was far from being a good starting point, because of shadows and dark colorations on the grains, together with rapeseeds and other artefacts. Hence it was necessary to take account of the known shapes of insects, which included Oryzaephilus surinamensis (the saw-toothed grain beetle). For this purpose, bar detector algorithms turned out to be a promising line of investigation. In fact, the regions near the centres of the insects were readily detected by relatively large masks approximating to rings of pixels. The optimum situation was when the rings were half within and half outside the insects, in this case following an ‘equal area’ rule based on the matched filter concept. To achieve

© Woodhead Publishing Limited, 2013

100 Robotics and automation in the food industry (a)

(b)

(c)

Fig. 4.16 Location of ergot amongst wheat grains. (a) Original image. (b) Result of applying two global thresholds, one locating the ergot and one identifying the background. (c) Analysis of the intensity histogram (top) using the ‘global valley’ approach (middle and bottom) to determine the best two threshold locations (two short vertical lines at the bottom). Note that the dark threshold (left) is determined surprisingly accurately considering the paucity of information at that position in the intensity histogram (top). (Source: Reproduced from Davies (2007) with permission; © IET, 2007.)

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 101 (a)

(b)

(c)

(d)

Fig. 4.17 (Continued overleaf)

© Woodhead Publishing Limited, 2013

102 Robotics and automation in the food industry (e)

Fig. 4.17 Insect location. (a) Original image. (b) Result of applying bar detector. (c) Result of applying end detector. (d) Combined result: ‘bar-plus-end’ detector. (e) comparison with thresholded version of (a), showing the large number of false alarms arising in that case. (Source: Reproduced from Davies et al. (2002) with permission; © IOP Publishing Ltd, 2002.)

minimum computation, two masks were employed instead of the anticipated eight (see Section 4.2.1). This was found to be possible first by analogy with edge detection, which requires two masks, and second because line segments have 180º rotation symmetry, which means that a rigorous mathematical mapping to 360º rotations makes detection using two masks appropriate. Tests of this procedure resulted in a FN rate of around 1% against a total of 300 insects: the FNs were due to cases where insects were viewed end-on or were partly hidden by the wheat grains. A small number of FPs arose from dark boundaries on some of the grains and the condition known as ‘black-end-ofgrain’, though it was found that either could be eliminated if they were shown to lie within the region of a grain. To further increase discrimination relative to chaff, which can sometimes resemble insects, improvements were made to the bar detector algorithm. Instead of merely detecting the central part of the bar, the ends were also detected using a spot detector (which was quite close to being an end-of-bar matched filter); merging the main part of the bar with the ends then gave a very good approximation to the true size and shape of the insect and thus the capability for highly accurate recognition (Fig. 4.17). In fact, other insects could be found by combining the dark-area approach, devised to detect non-insect contaminants (see the previous section), with the bar-plus-end detector approach described above (Davies et al., 2002). This permitted detection of several types of insect ranging from the saw-toothed grain beetle to very large insects similar in size to mouse droppings. To achieve this, the areas obtained by the dark-area detector and the bar-plus-end detector were combined, and used in a rule-based recognition system to classify the various objects detected; however, small insects were flagged directly from the bar-plusend detector output. The architecture of the overall system is shown in Fig. 4.18.

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 103 Bar detector Input image

End detector

Shape fusion

Morphological processor

Insect recognition

Dark object recognition

Fig. 4.18 Architecture of complete contaminant detection system. The dark-object recognition system starts by carrying out connected-components analysis and then identifies each component using rule-based recognition.

4.4

Recent advances in the application of vision in the food industry

In recent years the main driver of progress has been the need for greater control of quality, to more exacting standards, for a much wider range of foodstuffs. While the challenges were already there ten years or so ago, computer power was insufficient and machine vision had not been developed far enough to permit wide implementation. However, the former difficulty has now largely been overcome, and any remaining lack of computer power can be made up by relatively cheap dedicated processors based for example on field programmable gate array (FPGA) technology. Meanwhile, advances in algorithms and their gradual adaptation from one product to another have meant that the coverage of foodstuffs is much less patchy (Davies, 2009). For example, problems of defects on fruit and vegetables have been a matter of some importance for a good many years, with the result that surface defects, including spots, bruises, scab, scars, cracks, wrinkling, injury, mould, rot, discolouration and misshapes are all looked for much more consistently throughout the industry (Jahns et al., 2001; Bennedsen et al., 2005; Blasco et al., 2007; Xing et al., 2007; Riquelme et al., 2008). In fact, over the past decade or so, there has been a move away from concentration on highly controlled aspects of manufactured items such as cakes and biscuits, to more careful scrutiny of the raw products as a whole. Whereas a decade ago publications centred largely on staple diet foods such as potatoes and apples, nowadays we see publications on quite specialised fruit and vegetables, including olives, pomegranates, dates and kiwifruit (Moreda et al., 2007; Lee et al., 2008; Riquelme et al., 2008; Blasco et al., 2009). For processed foods, some attention has been focussed on cheeses, including those containing additional ingredients such as garlic, parsley and vegetables – which have to be checked for ingredient distribution (Jeliński et al., 2007). Use of X-rays has become more widespread, a phenomenon that has been accelerated (e.g. Batchelor et al., 2004, Kröger et al., 2006) by the arrival of DEXA systems (see Section 4.2.7). At the same time, the use of faster computers

© Woodhead Publishing Limited, 2013

104 Robotics and automation in the food industry has permitted greater attention to be paid to the analysis of colour (Stajnko et al., 2004; Riquelme et al., 2008; Blasco et al., 2009). In addition, artificial neural networks (Zhang et al., 2005; Mahesh et al., 2008), fuzzy logic (Jahns et al., 2001; Ureña et al., 2001) and decision tree classifiers (Gómez et al., 2007) have been successfully applied to a variety of foods, while morphology (Choudhary et al., 2008; Jiang et al., 2008) and thresholding (Davies, 2007; Chen and Qin, 2008) are probably the two most important means of discerning defects that have consistently moved forward in the last few years. (Because of the relative simplicity of these approaches, this was not a foregone conclusion, but with the support of well-controlled lighting, they have proved valuable for pre-processing ready for the application of rigorous pattern recognition methods.)

4.5

Appraisal of the need for special hardware for food inspection applications

In this section we look back over the development of computer hardware for image processing with particular reference to product inspection and handling. From the 1970s, when the possibilities for 100% inspection of products on product lines were first contemplated, it was found that available serial computers were nowhere near fast enough to meet real-time needs – the prime reasons being the large amounts of data in each image and the rate of processing that was entailed. Parallel processing using arrays of dedicated very-large-scale-integration (VLSI) logic chips was felt by many to be the key, though dedicated transputer farms and ‘bit-slices’ (very fast early types of microprocessor) were also fashionable. But during the 1990s the digital signal processing (DSP) chip, originally posed as a solution to computing fast-Fourier-transforms (FFTs) for processing speech waveforms, emerged as the dominant approach, only to be largely usurped in the early 2000s by FPGAs. It was soon found that these random logic arrays could be vastly increased in power by including a few modest microprocessors on the same chip. This meant that any computer with which they were used could spend most of its highly flexible processing power doing the more complicated high-level vision tasks, while the FPGAs did the repetitive low-level tasks such as edge detection and morphology. A decade later on, PCs and other computers have become so much more powerful that in a number of applications there is no need for special dedicated hardware accelerators such as DSPs or even FPGAs with embedded microprocessors. This is advantageous because both types of device need to be programmed using special software, while a PC can be programmed in a simple high-level language such as C++, Java or Matlab. However, it should be added that applications requiring only standard unaided PCs (or stripped down single-tasking PCs) are not yet very numerous, but as their speed and power increase we are at the exciting stage of envisaging more and more applications following this ideal. It should also be remarked that any PC chip needs a certain amount of random logic around it, so some FPGAs are also needed in any applied vision system. The result of all this is that for many

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 105 vision systems the costs of vision processing are almost negligible vis-à-vis (to take an important example) the cost of an X-ray setup on a food conveyor, including all the necessary screening: the total cost could easily be ~£100 000, compared with ~£10 000 or less for the PC vision processor. A further advantage of using an easily programmed PC as against a DSP or FPGA with embedded microprocessors is the reduced expertise and cost necessary to get the system working. Overall, the problem to be tackled is that of understanding which inspection processes and algorithms can be implemented on a single PC and which need FPGA or other processing hardware as well. Added to this, there is great difficulty in knowing exactly what available commercial vision systems can achieve.2 In fact, most of them have to be tailored to any new requirement, and prices can only be ascertained by extended discussions with the vendors and integrators: unlike the situation for other marketable devices and machines, prices are not available on the web; nor is it realistic to buy such a machine off the peg as the specifications of each application are so different. Indeed, the same system operating in the same application will be very dependent on the effectiveness of the lighting system, including the incidence of shadows, glints, ambient lighting, secondary illumination, and so on. These factors mean that an article of this length can hardly scratch the surface in recommending how to approach the problem. However, a case study approach is invaluable in helping to understand this sort of situation, as we shall now see. Looking again at the insect and non-insect contamination detection system outlined in Sections 4.3.3 and 4.3.4, the specification was that of producing a system that would inspect ~60 000 wheat grains in 3 min while costing only ~£5000. By the target date (around the year 2000) this was essentially achieved using a single PC (within a year advances in standard PC technology would in any case bring the speed fully up to specification). Thus no extra processors were involved, and no special (e.g. FPGA, DSP) programming or handling skills were required – apart from C++ programming capability. But if the task had been more sophisticated or much greater speed had been required, the unit costs would very quickly have increased substantially. As indicated earlier, it is in general difficult to estimate in advance the degree of sophistication that might be required – not least as lighting can dramatically alter the whole scenario: hence it only remains to remark that much can be achieved already with single PC vision systems, and that over the coming decade very much more will clearly be possible.

4.6

Conclusion and future trends

This chapter has aimed to give a background on available machine vision methodology as applied in the area of food processing: the basic principles are primarily those of image-processing and abstract pattern recognition, which have to work in tandem in order for complete machine vision systems to be built. Other factors in 2

A great many such systems can readily be found on the Internet using ‘X-ray, inspection, food’ as typical search terms.

© Woodhead Publishing Limited, 2013

106 Robotics and automation in the food industry complete designs of these systems are the image modalities to be used and appropriate methods of image acquisition. However, having computer hardware that is capable of the high speeds necessary for real-time operation is also a crucial factor. Over the past decade or so we have seen considerable progress in the march to control quality, and its achievement for a great many different products and raw materials; as a result, coverage is now distinctly less patchy and also more soundly based and reliable. All this would not have been possible without considerable advances being made in the design of vision algorithms over this period. At the same time computer power has advanced and its cost has been reduced so much that real-time implementation should be achievable for almost any envisaged process. Nevertheless, while it may seem that the subject is now ‘mature’, it is still the case that the algorithms for each new application have to be designed carefully and individually, making the process quite manpower intensive. What are currently needed are means to permit algorithms to be developed or adapted automatically. To some extent this can already be achieved by the use of trainable algorithms: for example, abstract pattern recognition algorithms are necessarily trainable as they have to be applied in a training-plus-testing regime. On the other hand, their training data normally have to specially prepared and selected, and this is itself a manpower-intensive process. However, over time this side of the subject will progress much further, with the machine gathering its own training data, and this will have to encompass not only the classifier itself but also the image-processing system that feeds it. A separate area that has been progressing in recent years, and has gradually been catching on, is that of hyperspectral processing (Gómez-Sanchís et al., 2008; Mahesh et al., 2008; Qin and Lu, 2008). This involves feeding each input pixel not just with a single intensity or colour but with a whole spectrum that is output from a spectrum analyser. Such a system is complex, expensive and slow in operation. Basically, it works by using an area camera in place of a line-scan camera, the additional dimension being used for the spectral input. This type of system is slow not only to acquire images but also to process them, as a typical image ‘hypercube’ contains hundreds of megabytes of data. Indeed, this is an area where new types of algorithm will need to be developed to cope with the superabundance of data, but it has already been applied to food inspection, for example, for the inspection of citrus fruit (Gómez-Sanchís et al., 2008). It seems likely that over the coming decade this technique will be developed much further and used for everyday application in the food industry.

4.7

Acknowledgements

The author is pleased to acknowledge financial support from United Biscuits (UK) Ltd, UK Ceńtral Resŏurces Ltd ańd the UK Scieńce ańd Eńgińeerińg Research Council (EPSRC) for financial support for the baked product work; and

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 107 from the UK Home-Grown Cereals Authority (HGCA) for financial support for the cereal grain inspection work. The author is grateful to Dr J. Chambers and Dr C. Ridgway of the Central Science Laboratory, MAFF (now Defra), York, UK for useful discussions on the needs of the grain industry, and for providing the original images used to test the cereal grain inspection algorithms. He is also grateful to the HGCA Cereals R&D Committee for guidance on the relative importance of different types of contaminant. Finally, he would like to thank former RHUL colleagues Mark Bateman and David Mason for vital assistance with the implementation of key cereal inspection algorithms.

4.8

Sources of further information and advice

The following books and reviews are particularly relevant to the present chapter: Davies, E. R. (2000), Image Processing for the Food Industry, Singapore: World Scientific. Davies, E. R. (2009), ‘The application of machine vision to food and agriculture: a review’. Imaging Science, 57, 197–217. Davies, E. R. (2012), Computer and Machine Vision: Theory, Algorithms, Practicalities, 4th edition, Oxford, UK: Academic Press. Edwards, M. (ed.) (2004), Detecting Foreign Bodies in Food, Cambridge, UK: Woodhead Publishing Ltd. Graves, M. and Batchelor, B.G. (eds.) (2003), Machine Vision Techniques for Inspecting Natural Products, London: Springer Verlag. Mirmehdi, M., Xie, X. and Suri, J. (eds) (2008), Handbook of Texture Analysis, London: Imperial College Press. Pinder, A.C. and Godfrey, G. (eds) (1993), Food Process Monitoring Systems, London: Blackie. Webb, A. (2002), Statistical Pattern Recognition, 2nd edition, Chichester: Wiley.

In addition, the reader will find the journals and associations listed below to be of value. • Major food-related associations: American Society of Agricultural and Biological Engineers (ASABE) – formerly the ASAE Campden and Chorleywood Food Research Association (UK) Home-Grown Cereals Authority (HGCA) (UK) Leatherhead Food Research (UK) • Major machine vision and engineering associations: British Machine Vision Association (BMVA) (UK) European Machine Vision Association (EMVA) Institute of Electrical and Electronic Engineers (IEEE) (USA) Institution of Engineering and Technology (IET) (UK) United Kingdom Industrial Vision Association (UKIVA)

© Woodhead Publishing Limited, 2013

108 Robotics and automation in the food industry • Major food-related journals: Biosystems Engineering Cereal Chemistry Computers and Electronics in Agriculture Journal of Agricultural Engineering Research Journal of Food Engineering Journal of the Science of Food and Agriculture Meat Science Transactions of the ASABE Trends in Food Science and Technology • Major machine vision journals: IEEE Transactions on Pattern Analysis and Machine Intelligence Imaging Science Pattern Recognition Pattern Recognition Letters Real-time Image Processing – formerly Real-time Imaging

4.9

References

Batchelor B G, Davies E R and Graves M (2004), ‘Using X-rays to detect foreign bodies’, in Edwards M (ed.), Detecting Foreign Bodies in Food, Woodhead Publishing Ltd, Cambridge, UK, 226–264. Bennedsen B S, Peterson D L and Tabb A (2005), ‘Identifying defects in images of rotating apples’, Computers and Electronics in Agriculture, 48, 92–102. Blasco J, Aleixos N, Gómez J and Moltó E (2007), ‘Citrus sorting by identification of the most common defects using multispectral computer vision’, Journal of Food Engineering, 83, 384–393. Blasco J, Cubero S, Gómez-Sanchís J, Mira P and Moltó, E (2009), ‘Development of a machine for the automatic sorting of pomegranate (Punica granatum) arils based on computer vision’, Journal of Food Engineering, 90, 27–34. Chen K and Qin C (2008), ‘Segmentation of beef marbling based on vision threshold’, Computers and Electronics in Agriculture, 62, 223–230. Choudhary R, Paliwal J and Jayas D S (2008), ‘Classification of cereal grains using wavelet, morphological, colour, and textural features of non-touching kernel images’, Biosystems Engineering, 99, 330–337. Davies E R (1984), ‘Design of cost-effective systems for the inspection of certain food products during manufacture’, in Pugh A (ed), Proceedings of 4th Conference on Robot Vision and Sensory Controls, London, 9–11 October, 437–446. Davies E R (2000), Image Processing for the Food Industry, World Scientific, Singapore. Davies E R (2001), ‘Some problems in food and cereals inspection and methods for their solution’, Proceedings of 5th International Conference on Quality Control by Artificial Vision (QCAV 01), Le Creusot, France, 21–23 May, 35–46. Davies E R (2007) “A new transformation leading to efficient multi-level thresholding of digital images”, Proceedings of IET International Conference on Visual Information Engineering, Royal Statistical Society, London, 25–27 July, paper 26, 1–6. Davies E R (2009), ‘The application of machine vision to food and agriculture: a review’, Imaging Science, 57, 197–217.

© Woodhead Publishing Limited, 2013

Machine vision in the food industry 109 Davies E R (2012), Computer and Machine Vision: Theory, Algorithms, Practicalities, 4th edition, Academic Press, Oxford, UK. Davies E R, Bateman M, Chambers J and Ridgway C (1998), ‘Hybrid non-linear filters for locating speckled contaminants in grain’, IEE Digest no 1998/284, Colloquium on Non-Linear Signal and Image Processing, IEE, 22 May, 12/1–5. Davies E R, Chambers J and Ridgway C (2002) ‘Combination linear feature detector for effective location of insects in grain images’, Measurement Science Technology, 13, no 12, 2053–2061. Fawcett T (2006), ‘An introduction to ROC analysis’, Pattern Recognition Letters, 27, 861–874. Gómez J, Blasco J, Moltó E and Camps-Valls G (2007), ‘Hyperspectral detection of citrus damage with a Mahalanobis kernel classifier ’, Electronics Letters, 43, no 20, 1082–1084. Gómez-Sanchís J, Moltó E, Camps-Valls G, Gómez-Chova L, Aleixos N and Blasco J (2008), ‘Automatic correction of the effects of the light source on spherical objects An application to the analysis of hyperspectral images of citrus fruits’, Journal of Food Engineering, 85, 191–200. Jahns G, Nielsen H M and Paul W (2001), ‘Measuring image analysis attributes and modelling fuzzy consumer aspects for tomato quality grading’, Computers and Electronics in Agriculture, 31, 17–29. Jeliński T, Du C-J, Sun D-W and Fornal J (2007), ‘Inspection of the distribution and amount of ingredients in pasteurized cheese by computer vision’, Journal of Food Engineering, 83, 3–9. Jiang J-A, Chang H-Y, Wu K-H, Ouyang C-S, Yang M-M, Yang E-C, Chen T-W and Lin T-T (2008), ‘An adaptive image segmentation algorithm for X-ray quarantine inspection of selected fruits’, Computers and Electronics in Agriculture, 60, 190–200. Kröger C, Bartle C M, West J G, Purchas R W and Devine C E (2006), ‘Meat tenderness evaluation using dual energy X-ray absorptiometry (DEXA)’, Computers and Electronics in Agriculture, 54, 93–100. Lee D-J, Schoenberger R, Archibald J and McCollum S (2008), ‘Development of a machine vision system for automatic date grading using digital reflective near-infrared imaging’, Journal of Food Engineering, 86, 388–398. Mahesh S, Manickavasagan A, Jayas D S, Paliwal J and White N D G (2008), ‘Feasibility of near-infrared hyperspectral imaging to differentiate Canadian wheat classes’, Biosystems Engineering, 101, 50–57. Moreda G P, Ortiz-Cañavate J, García-Ramos F J and Ruiz-Altisent M (2007), ‘Effect of orientation on the fruit on-line size determination performed by an optical ring sensor ’, Journal of Food Engineering, 81, 388–398. Qin J and Lu R (2008), ‘Measurement of the optical properties of fruits and vegetables using spatially resolved hyperspectral diffuse reflectance imaging technique’, Postharvest Biology and Technology, 49, 355–365. Ridgway C, Davies E R, Chambers J, Mason D R and Bateman M (2002), ‘Rapid machine vision method for the detection of insects and other particulate bio-contaminants of bulk grain in transit’, Biosystems Engineering, 83, no 1, 21–30. Riquelme M T, Barreiro P, Ruiz-Altisent M and Valero C (2008), ‘Olive classification according to external damage using image analysis’, Journal of Food Engineering, 87, 371–379. Stajnko D, Lakota M and Hŏcevar M (2004), ‘Estimation of number and diameter of apple fruits in an orchard during the growing season by thermal imaging’, Computers and Electronics in Agriculture, 42, 31–42. Ureña R, Rodríguez F and Berenguel M (2001), ‘A machine vision system for seeds germination quality evaluation using fuzzy logic’, Computers and Electronics in Agriculture, 32, 1–20.

© Woodhead Publishing Limited, 2013

110 Robotics and automation in the food industry Webb A (2002), Statistical Pattern Recognition, 2nd edition, Wiley, Chichester. Xing J, Saeys W and DeBaerdemaeker J (2007), ‘Combination of chemometric tools and image processing for bruise detection on apples’, Computers and Electronics in Agriculture, 56, 1–13. Zhang G, Jayas D S and White N D G (2005), ‘Separation of touching grain kernels in an image by ellipse fitting algorithm’, Biosystems Engineering, 92, no 2, 135–142.

© Woodhead Publishing Limited, 2013

5 Optical sensors and online spectroscopy for automated quality and safety inspection of food products C. B. Singh and D. S. Jayas, University of Manitoba, Canada

DOI: 10.1533/9780857095763.1.111 Abstract: Due to increasing consumer demand for healthy, high quality, and safe food, food processing industries are exploring/adopting new non-destructive, fast, and accurate techniques for quality and safety inspection of their products. Optical sensing and spectroscopic techniques have high potential for automated real-time quality and safety inspection of agricultural and food products. These techniques have already been adopted successfully in several food processing units for automated quality monitoring of their products. This chapter discusses various optical sensing techniques under three broad categories namely: spectroscopic, fiber optic and image sensing. Working principles, instrumentation, advantages, disadvantages, and limitations of these techniques have been described and various applications for quality monitoring in the agriculture and food processing industries have been reviewed. Key words: food quality and safety, automated inspection, NIR spectroscopy, IR spectroscopy, Raman spectroscopy, ultraviolet spectroscopy, machine vision, hyperspectral imaging, fiber optic sensing.

5.1

Introduction

Food material should be inspected raw as well as processed on-line, in-line, or at-line for its quality (color, texture, flavor, shape, nutritional value, composition, and functionality) and safety (chemical and microbial contamination, foreign material, impurity, and adulteration). Human inspection of food products is slow and subjective, whereas chemical wet analysis for food compositional analysis is destructive, time consuming, and expensive. The food processing industry is looking for alternative in/on-line, fast, accurate, non-destructive, and less expensive techniques for automated food quality monitoring. Optical sensing

© Woodhead Publishing Limited, 2013

112 Robotics and automation in the food industry and spectroscopic techniques meet these criteria and can be used in automated systems in the food processing industry. Optical sensors are devices used to measure radiation signals in the form of absorbed, scattered, attenuated, or fluorescent light from the sample with characteristic optical properties for classification/quantification. The interaction between sample molecules and radiated energy is wavelength specific, thus carries distinguishable characteristics. The detectors in optical sensors convert the collected radiometric light as electric signals, which are then recorded and used for further analysis. The optical sensors can have single, multiple, line or array of detectors based on the technique and end use. In optical-spectroscopic instruments, mostly single-point-based detector elements are used, whereas in optical imaging sensors mostly a linear or two-dimensional (2D) array of detectors is used. If the sensor uses a fiber optic technique for delivering and receiving the radiated light, it is called fiber optic sensing technology. The choice of detector is heavily dependent on the specific use of a particular range of electromagnetic radiation (e.g., ultraviolet (UV), visible (VIS), near-infrared (NIR), infrared (IR)). An ideal detector should have high sensitivity, large dynamic range, wide spectral response, low drift, and low thermal noise. Irrespective of the type of sensor, most of the sensing techniques consist of a detector, a light source, and a wavelength filter/selection device. There are several classifications of optical sensors to be found in the published literature based on: magnitude and nature of the quantity to be measured (mechanical, thermal, electromagnetic, chemical composition, radiation, and typical biomedical measurements); spatial distribution (point, integrated, distributed, and quasi-distributed); sensing technology (optical fiber technology (OFT), integrated optical technology (IOT), integrated opto-electronic (IOE), and hybrid optic technology); or modulations (amplitude or intensity sensor, phase or interferometric sensor, polarimetric sensors, and spectroscopic sensors) (Lopez-Higuera, 2002). For simplicity and basic understanding of the topic, we have divided optical sensors into three broad categories namely, spectroscopic, fiber optic and image sensing techniques. Each of these categories is discussed along with the principles, instrumentation, advantages and disadvantages, and limitations. Various applications of these techniques for quality monitoring in the food processing industry have also been reviewed and presented.

5.2

Optical sensing and spectroscopic techniques

This section describes the working principles and hardware components of various spectroscopic, fiber optic, and image sensing techniques.

5.2.1 Spectroscopic techniques Point-based optical sensors are mainly used in spectroscopic instruments. They consist of a single detector element or an array of detector, radiation source, and wavelength filter devices. Wavelength filter or grating devices disperse the light

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 113 Long wavelength Low energy Low frequency

Short wavelength High energy High frequency

Gamma

UV X-Ray

NIR VIS

Microwave IR

Radiowave

Fig. 5.1 Electromagnetic spectrum. (Source: Singh and Jayas, 2010.)

onto a material, and the detector element records the amount of radiation absorbed (UV, IR, and IR) or scattered (Raman spectroscopy) after its interaction with the material. The detector, also known as a photodetector or photodiode, is the core element of the sensing unit, converting the measured input signal (radiation) into an electrical output signal. An ideal photodetector should have a wide spectral detectability range, high response, high speed (large bandwidth), low noise, high stability, high reliability, compact size, and low cost (Lopez-Higuera and Madruga, 2002). The full electromagnetic spectrum spans from very high energy gamma rays and X-rays, UV, VIS, NIR, IR, and microwaves, to the low energy radio waves (Fig. 5.1). Different types of spectroscopic techniques are available for different regions of electromagnetic radiation. The spectral characteristics depend on the material to be inspected/analyzed, region of electromagnetic radiation, and type of instrument used for the analysis. Molecules in space can possess energy in several forms, such as vibrational, rotational, or electronic. Vibrational energy causes periodic displacement of atoms from their equilibrium position, and rotational energy causes rotation about a center of gravity. The energy required for rotational transition is low compared to the energy required for vibrational transition (NIR–mid IR). Rotational energy is absorbed in far IR and microwave regions, and the spectra in this region are very broad (except for the gases). Vibrational motion is generally used in the spectral analysis of solids and liquids (Osborne et al., 1993). Electronic energy associated with electron excitation in UV-VIS region is studied in the UV spectroscopy. Infrared, Raman, and NIR spectroscopy are part of vibrational spectroscopic techniques. IR spectroscopy In IR spectroscopy, fundamental vibrations arising from molecular transitions between adjacent energy states (allowable change in vibrational quantum number is ν = ±1) are studied. Photons in the IR region have lower energies, corresponding to covalent bond stretch and bend vibrations capable of producing only fundamental transitions. The IR spectra are mainly studied in the mid IR region (650–4000 cm−1), which has well-resolved peaks. The IR spectra are very useful in the structural and qualitative analysis of food materials. IR spectroscopy has

© Woodhead Publishing Limited, 2013

114 Robotics and automation in the food industry the capability to analyze solid, liquid, and gas samples; however; the sample preparation requirement limits rapid analysis. In the past, dispersion IR spectroscopic instruments using grating devices limited the application because of slow data collection rates. With the advent of Fourier transform infrared (FTIR) techniques the sample can be scanned and spectra can be recorded in only a few seconds. FTIR spectroscopy has several advantages, such as high sensitivity, increased brightness, high resolution, and very low background noise since only an interference signal is used in the Fourier transformation. An FTIR instrument mainly consists of an IR radiation source (referred to as a light source and not as a thermal source), a fixed mirror, a moving mirror, a beamsplitter, and a detector. The beamsplitter splits the incoming radiation into two beams and directs them respectively to a moving and a fixed mirror. These two beams are reflected back to the beamsplitter, causing interference (constructive interference if the mirrors are equidistant, and destructive interference if one mirror moves away), which is sensed by the detector after passing through the sample. The detector output is a fluctuating cosine wave also known as an interferogram, which is complex and difficult to analyze, and hence Fourier transformation is used to transform this data into a wavelength and intensity. The PbS detectors are used for the low IR spectral range. The mercury cadmium telluride (MCT) is a widely-used photodetector in mid IR that has high sensitivity, low noise, and high spectral resolution. The PbSe and InSb detectors are also used in the mid IR spectral range. Quantum well infrared photodetectors (QWIP) with wide IR spectral response are also available nowadays and are cost effective, but have higher dark current and lower quantum efficiency compared to the MCT detectors (Lopez-Higuera and Madruga, 2002). Deuterated triglycine sulphate (DTGS) is used in IR thermal detectors. Raman spectroscopy Raman spectra are also produced by molecular vibration, but the vibrations in Raman spectroscopy are symmetric and the main criteria is the polarizability of the compounds. Raman spectra are the result of scattering of incident radiation on a sample. If the scattered light has the same frequency as the incident light, due to elastic collision between striking photons and vibrating molecules, the scattering is called Rayleigh scattering. However, in the case of inelastic collisions, the scattered light has a different frequency from the striking radiation, due to energy exchange between photons and molecules, and the scattering phenomenon is then described as Raman scattering. The majority of the vibrating molecules are present in the ground energy state, but some are present in excited energy states. When an incoming photon strikes a molecule of relatively low energy present in the ground energy state, the molecule gains energy and the scattered photon loses its energy, resulting in Stokes Raman scattering. When a photon strikes a molecule with relatively high energy, the photon gains energy from the molecule, and the scattered photon has higher energy, giving rise to the anti-Stokes scattering phenomena. Stokes scattering lines are usually stronger than anti-Stokes lines due to the high energy of the scattered photons. Stokes lines are used in Raman spectral analysis. In the

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 115 Raman spectrometer, various excitation sources (such as visible excitation source in VIS excitation Raman spectroscopy, UV lasers in resonance Raman spectrometers, and diode-bar pumped Nd-YAG lasers in Raman NIR spectrometers) are used for excitations in various wavelength ranges (Kizil and Irudayaraj, 2008). For VIS and UV excitation, charge-coupled device (CCD) detectors are used, whereas in NIR excitation germanium detectors are the preferred sensing element. IR and Raman spectroscopic techniques give complementary information for analyzing a sample. Peaks in IR spectra are produced due to molecular vibrations caused by a change in dipole moment of the molecule, and peaks in Raman spectra are produced due to vibrations caused by a change in the electronic polarizability, that is, polar bonds (C–H, O–H, C=O, C==O, C–O, C–F, Si–O, and N-H) give strong IR absorption peaks but homonuclear apolar symmetric bonds (C==C, C=C, C–C, S–S) are poor absorbers of IR radiation but give strong Raman Stokes lines. Raman spectroscopy is not very popular for several reasons, including small sample volume, interference by fluorescence, low efficiency, low intensity of desired Stoke’s scattering, and high intensity of the undesired Rayleigh scattering. NIR spectroscopy NIR spectroscopy is the most common among the vibrational techniques currently being used in the agri-food industry. Overtones (transition energy for transition from ν = 0 to higher transition levels for example, ν = 2, 3) and combination bands (combinations of overtones and fundamental vibration) that occur in the range 700–2500 nm (14285.71–4000 cm−1) are studied in the NIR spectroscopic analysis (Osborne et al., 1993). Absorptions in the NIR region are caused by vibrations of –CH, –OH, and –NH bond groups. Overtones and combination bands are weaker than fundamental absorption bands in the IR spectral region, which is an advantage as samples of thickness of several mm – such as solid food, packed food, and intact whole seeds – can be analyzed in an NIR spectrometer without any sample preparation. An NIR instrument consists of a light source (Tungsten halogen lamp, quartz halogen lamp, light emitting diodes (LED)), a wavelength filtering device (prism, grating, interference filters, and advanced electronically tunable filters (ETF)), and a diode array detector (Silicon (Si), lead sulphide (PbS), lead selenide (PbSe), indium gallium arsenide (InGaAs)), and a data acquisition and processing system and software. Laboratory and research based instruments require full spectral analysis, hence prism, grating, or ETF are required. Two advanced ETFs, namely acousto-optical tunable filters (AOTF) and liquid crystal tunable filters (LCTF), are commonly used in NIR scanning instruments for full spectral analysis in contiguous wavebands. In an industrial inspection setting, sample scanning at only a few discrete wavelengths (identified from full scanning research instruments) is required, which can be accomplished using less expensive interference filters. Interferogram based FT-NIR instruments are also available with improved noise level. The spectral data acquired from a spectroscopic instrument can be analyzed both quantitatively and qualitatively, and calibration (prediction) and classification models can be developed using various advanced chemometric tools for multivariate analysis (Naes et al., 2002). Prior to model development, the spectral data must be pretreated to remove/overcome the

© Woodhead Publishing Limited, 2013

116 Robotics and automation in the food industry noise and non-linearity between spectral signal (absorbance) and the concentration of food material. This non-linearity may arise due to several physical and chemical phenomena, such as scattering due to particles, interferents, molecular interactions, changes in refractive index at high concentrations, stray light, and changes in sample size (ground or whole) and thus path length, and the effects of growing location and growing season on the crop/food. Several preprocessing methods, such as multiplicative scatter correction (MSC), de-trending, standard normal variate (SNV), normalization, baseline correction, and derivation (finite difference and SavitzkyGolay smoothing) are used in preprocessing spectral data (Naes et al., 2002; Rinnan et al., 2009). UV spectroscopy UV spectra arise as a result of electron excitation into high energy levels. Several organic molecules containing unshared electrons in sulphur, bromine, and iodine have absorption peaks in the UV region (Meurens, 2003). The UV region is divided into two subgroups, namely vacuum UV (100–190 nm) and near UV (190–400 nm) (Cooper, 1980). The spectroscopic instrument for the vacuum UV region requires absolute vacuum conditions to eliminate/reduce UV absorption in this region, and hence is not suitable for in-line-processing. The near UV region is commonly used in UV spectroscopic analysis. Optical material in UV cells for holding the sample is made of quartz (for use above 210 nm) or silica (for use up to 165 nm). A UV spectrophotometer consists of a light source (deuterium or xenon lamp, LEDs, diffraction grating, and a detector (diode array or CCD detector). Pure solvents (CCl4, CHCl3, CH2Cl2, ethanol, methanol, water, and dioxane), with no absorption in the UV spectral region of interest, are used. The UV absorption spectra are generally broad and hence not common in trace identification, but they have potential for the quantitative analysis of food products.

5.2.2 Fiber optic sensors Optical fiber sensors offer several advantages, such as being passive, dielectric, light weighted, compact, minimal or no interference, inexpensive, safe, and easy to install (Rogers, 1992). In fiber optic technique, fiber materials (in the form of thin wire) are used in transmitting light through optical components (filters, beamsplitters) that guide the light to the sample and/or sensing element or detector. The fiber line consist of three layers, namely an inner core layer made of glass, silica or other fiber material, and a middle layer (cladding) surrounded by an outer protective layer. The light travels through the inner core by means of internal reflection. The fiber optic lines are able to deliver high quality light to a detector several meters away even if the cables are bent, like any other wire, and are not affected by the surrounding environment. A simple fiber optic sensing instrument consists of an optical source, detector and optical fiber line (single or multimode). LEDs or laser diodes are mainly used as light sources. Semiconductor devices, such photodiodes, are used as detectors. Distributed optical fiber sensing (DOFS)

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 117 is used to map two-dimensional data with spatial labeling similar to other imaging techniques. Sometimes fiber material itself is part of the sensing element or detector (intrinsic sensor), however, in the present context, fiber material is referred to and meant only for transmission of light to the sample and detector element. Detailed information about the fiber optic technology is beyond the scope of this chapter and readers can find the information elsewhere (LopezHiguera, 2002).

5.2.3 Imaging sensors In imaging sensors, a line or 2D array of sensors is used, instead of just the single detector used in point-based instruments. The imaging sensing system consists of a light source and filter devices, as used in point-based instruments. The advantage with the image sensor is that it provides the spatial labeling of each pixel in conjunction with absorbance/reflectance value. The spatial resolution of the imaging device depends on the size of the detector element, that is, a small detector will give higher resolution but, at the same time, produce more noise as less intense light reaches the detector area. 2D data acquisition at several wavebands (thus forming a 3D hypercube) is possible, using either grating devices or tunable filters. In a color imaging system, data are collected at red, green, and blue bands. If 2D image data are collected at up to ten wavebands, this technique is termed multispectral imaging technique. 2D image data can also be collected at hundreds of narrow contiguous wavebands, using hyperspectral imaging. Therefore, samples can be analyzed as spectra for each pixel or as an image at each wavelength. The 3D hyperspectral data needs to be corrected for the same type of noise as in spectroscopic data, by applying similar preprocessing techniques. However, hyperspectral data also require correcting geometric distortions in images caused either by moving grating instruments or optical errors in ETFs. The details of calibrating hyperspectral imaging systems have been reported elsewhere (Geladi and Grahn, 1996; Lawrence et al., 2003a). In real-time imaging applications, very high speed data acquisition (scanning speed) by the sensor (camera) is required. The scanned data from the detector (output signal) should also be transferred to the host computer for storage and analysis. CAMERA Link is a standard communication interface that is being used for high speed, high resolution, and very good quality data transmission applications. Other communication interfaces such as fire wire (as referred as IEEE 1394) and GiGE VISION are also used. The host computer should be able to handle the amount of data transferred by the communication interface.

5.3

Applications in the food industry

Various applications of spectroscopic, fiber optic, and image sensing techniques have been reviewed and are presented in Tables 5.1, 5.2, and 5.3, respectively.

© Woodhead Publishing Limited, 2013

118 Robotics and automation in the food industry 5.3.1 Applications of spectroscopic techniques Infrared spectroscopy has been applied to the analysis (composition, quality, and authenticity) of a variety of food products, such as meat, cheese, and apples, and this technology can provide rapid characterization of food products (Fagan and O’Donnell, 2008). Sivakesava et al. (2001) used mid FTIR spectroscopy to detect sugar adulterants in apple juice. A linear discriminant analysis (LDA) classifier discriminated juice adulterated with beet and cane sugar invert. Partial least squares (PLS) and principal component regression (PCR) models were also developed to determine the quantity of beet and cane invert sugar in apple juice. Results demonstrated the applicability of IR techniques for rapid detection and quantification of adulteration in apple juice and other beverages. Inon et al. (2004) used FTIR spectroscopy to predict five nutritional parameters of milk, namely total fat, total protein, total carbohydrates (CH), calories, and calcium. The FTIR spectra were collected in total attenuated reflection mode (ATR) from commercial milk samples covering wide parameter ranges and different milk types, namely whole, semi-skimmed and skimmed, enriched with calcium and vitamins, and modified by altering lipid or sugar content. Spectral data were preprocessed and the calibration set was selected using hierarchical cluster analysis (HCA), and prediction models were developed using PLS. Spectral data collection using ATR mode instead of transmission mode gave good signal-to-noise ratio as ATR was less influenced by the presence of water in milk samples. Applications of Raman spectroscopy have been reported for authenticity detection of fermented milk, discrimination between glutinous and non-glutinous rice, authentication of extra virgin olive oil and honey (Meurens, 2003). Afseth (2005) compared the performance of Raman spectroscopy and NIR spectroscopy to determine the fatty-acid composition and contents of main constituents in a complex model food sample composed of 70 different mixtures of protein, water, and oil blends and representing fish and meat. The model samples, as well as pure oil mixtures, were scanned in Raman and NIR spectroscopy, and PLS models were developed for prediction of fatty acid in terms of iodine value and as contents of saturated, monounsaturated, and polyunsaturated fatty acids. Raman spectroscopy gave the best iodine value prediction accuracy; however, NIR gave the best accuracy in determining contents of main constituents in the model samples. Batsoulis et al. (2005) used FT-Raman spectroscopy to simultaneously determine fructose and glucose content in honey. Reference data were collected by a high performance liquid chromatography (HPLC) method, and the PLS prediction model was developed using Raman spectral and HPLC reference data. The Raman spectrometer and standard HPLC method were found to be statistically equivalent in terms of accuracy and reproducibility, but HPLC requires more analysis time. Muik et al. (2005) used FT-Raman spectroscopy to directly monitor lipid oxidation in edible oils. From the pure component spectra, formation of aldehydes was detected and both saturated and unsaturated aldehydes were identified. The conjugated double-bond systems formation and cis to trans isomerisation of double bonds was noticed in the C=C stretching region with a

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 119 unique pattern for each oil. These differences were related to fatty-acid composition. This study demonstrates the capability of Raman spectroscopy to monitor the oxidative degradation of lipids in edible oils without any sample preparation requirement. Several applications of NIR spectroscopy for on/in-line quality monitoring of meats, fruit and vegetables, grains and grain products, dairy and dairy products, fish and fish products, oils, and beverages have been reported (Huang et al., 2008). NIR spectroscopy can potentially be used for multicomponent analysis (moisture, fat, protein, oil, etc.) for a variety of food products, such as biscuits, chocolates, cocoa powders, cheese, flour, cereals and cereal based products, milk and milk powder, oil seeds and oils, minced meat (beef, pork, turkey, chicken, lamb), snack foods, and animal feed (Benson, 2003). Nicolai et al. (2007) presented an overview of NIR instrumentation, chemometrics, calibration transfer, and various applications of NIR spectroscopy for soluble solid content (SSC) determination in a variety of horticulture products (apple, apricot, cherry, citrus fruit, grape, guava, kiwifruit, mandarin, mango, melon, nectarine, papaya, peach, pear, pineapple, plum) and quality evaluation (firmness, pH, dry matter, insoluble solids, stiffness, sugar, color, starch, acidity, sensory attributes, bruise, optimum harvest date, essential oils and their composition, carotenoids, defects, maturity, volatiles, pectins, hardness, density, citric acid, oil, moisture, oleic acid, linoleic acid, ripeness, chlorophyll, fresh weight, and contaminants) of fruits and vegetables (apple, apricot, cherry, citrus fruit, grape, guava, kiwifruit, macadamia nuts, mandarin, mango, melon, mushroom, nectarine, olive, onion, papaya, peach, pear, pepper, pineapple, plum, tangerine, and tomato). Marquez et al. (2005) used NIR transmittance spectroscopy for on-line quality control and characterization of virgin olive oil. The NIR equipment was installed in a production line to collect spectra from real-time samples. Calibration models were developed for acidity value (AV), bitter taste (k225) and fatty-acid composition (FAME), using PLS regression and reference data. The results of PLS models demonstrated the potential of NIR spectroscopic technique for on-line and real-time quality monitoring and characterization of the oil during the oil-extraction process. Kawamura et al. (2007) developed an online NIR sensing system to monitor milk quality during milking. The NIR system was able to obtain spectra of unhomogenized milk during milking of individual cow and predict major milk constituents (fat, protein, and lactose), somatic cell count (SCC), and milk urea nitrogen (MUN) using a developed calibration model. The precision and accuracy of the validated models showed the capability of the developed NIR systems for real-time quality monitoring of milk during milking. Chen et al. (2006) used NIR spectroscopy for rapid determination of organic acid composition in Japanese apricot juice. Spectral data were collected from juice samples of different organic acid composition, and calibration models were developed using PLS and validated using an independent validation set. The predicted values of organic acid composition (citric and malic acid contents) were in good agreement with reference values. The results demonstrated the capability of NIR techniques for quantitative compositional analysis of fruit juices.

© Woodhead Publishing Limited, 2013

120 Robotics and automation in the food industry UV spectroscopy has been used for detection of fruit juice and vegetable oil adulteration. Fruit juice and wine contain polyphenols whose concentration can be measured using the UV spectrum. This technique has been used to detect adulteration in Florida orange juice and authenticity of virgin olive oil (Meurens, 2003).

5.3.2 Applications of optic fiber sensing techniques Castillo et al. (2005) used optic fiber sensing technology to determine whey-fat concentration in cheese making. The light sidescatter and transmittance spectra from the samples were collected in the 300–1100 nm wavelength range. Normalized spectral data were correlated with reference whey-fat concentration using a power function. The model using sidescatter waveband ratio gave better prediction accuracy than the transmission waveband ratio, demonstrating the potential of light scattered technique in automated process control in the food industry. Everard et al. (2009) used an on-line fiber optic sensor and derived color parameters to monitor curd syneresis during cheese making. An on-line fiber optic color sensor was installed in a cheese vat to monitor curd syneresis. Color parameters derived from the sensor predicted curd moisture and whey-fat content during syneresis. The results showed that color optic fiber sensor can monitor the syneresis process and improve cheese quality. Lewis et al. (2008) used optical fiber sensing technology to monitor food quality during cooking in industrial large-scale cooking ovens. During cooking food material changes its color due to various chemical reactions. A color sensor was used to monitor the cooking process and visible spectra were recorded at each stage. Various food products, namely steamed skinless chicken fillets, roast whole chickens, marinated chicken pieces, sausages, pastry, breadcrumb coating, and char-grilled chicken fillets, were used in the analysis and visible spectra were recorded. A probe, encased in stainless steel, consisting of an illumination optic fiber (attached to a tungsten halogen lamp), and a fiber to guide diffuse reflected light to the spectrometer, was designed to puncture the food material at high speed for sampling without causing any damage. Dimensionality of the spectral data was reduced by principal component analysis (PCA) and classification models were developed by artificial neural networks (ANN) using PCA reduced spectral data. Food products were successfully classified into various cooked stages as raw, light, correct, and dark. Alvarez et al. (2009) used an optical fiber sensor to monitor stability, during the meat emulsification process, by visible light scattering. Normalized backscatter spectra were derived and correlated with chopping times and fat/lean ratios, which had a high influence on meat emulsion stability. Increasing the distance between emitting and detecting optical fibers decreased the scattered light intensity logarithmically. Light propagation also decreased due to increases in chopping duration and fat/lean ratio. The fattier and overchopped emulsions were more unstable and caused more cooking loss. The optimum sensor parameters from this study indicated that multiple measurements

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 121 groups (detecting fibers surrounding an emitting fiber), spaced over a relatively large area with less than 2.50 mm radial distance between the emitting and the detecting fibers within a group, would be more effective in emulsion monitoring. Mateo et al. (2010) used an on-line NIR optical sensor to determine curd moisture content and solids in whey in cheese making. An optic fiber line was used to emit and collect the backscattered light. The reflected backscattered light was transmitted through another fiber and collimating lens, which focused the reflected light onto another fiber connected to a detector. Spectra were recorded by the detector and fitted to a cross-validation (leave-one-out) regression model for selection of useful wavelengths using variance as criteria. The PLS models were developed to predict the curd moisture content and whey solids using single wavelength, single wavelength with other compositional and technique parameters (time after gel cutting, gel cutting time, and milk fat level), and broad spectra. Single wavelength gave reasonable results, whereas single wavelength with other parameters gave the best prediction accuracy. Whey solid could be predicted only using the full spectral range.

5.3.3 Applications of image sensing technology Imaging sensors (cameras) have been extensively used in several applications in agriculture and food industries for quality and safety inspection, damage detection, sorting, grading, and monitoring of real-time food processing systems. Most common imaging systems use sensors sensitive to the VIS region of the electromagnetic spectrum. Color cameras used in color imaging systems for industrial process inspection are the least expensive cameras used in machine vision technology. Machine vision technology has been used in the assessment of fruits and nuts (apple, oranges, strawberries, nuts, tomatoes, peaches, and pears), vegetable inspection (mushrooms, potatoes, and others), grain classification and quality evaluation (wheat, corn, rice, barley, oats, rye), and applications in food products (pizza, bakery products, cheese, meat, and meat products) (Brosnan and Sun, 2004). Three chip CCD detectors are used to acquire images at red, green, and blue bands of the VIS spectrum and combined to simulate a near true color image of the object as seen by human eyes. Color images contain a lot of information in terms of shape, size, color, and texture of an object, which are quantitatively described by extracting color, textural, and morphological (shape and size) features. These features are used in pattern recognition and in learning/training the classifiers/prediction models for future classification and prediction of constituents in food products. The learning techniques used in machine vision technology are ANN, statistical learning (SL), fuzzy logic, genetic algorithms, and decision trees; however, ANN and SL are the most widely adopted learning methods (Du and Sun, 2006). Color, size, shape, and textural features have been extracted and applied in the quality analysis of several products in the agri-food industry namely apple, beef, chicken, pork, cocoa butter, fish, lettuce, maize noodle, nut, pizza, peach, tomato, cucumber, corn, cherry, orange, grape,

© Woodhead Publishing Limited, 2013

122 Robotics and automation in the food industry rice, grains, and poultry (Zheng et al., 2006). Anami and Burkpalli (2009) developed color-based models to identify different boiled grain types and then classify boiled grain into full, medium, and half boiled grains. Five types of grain, namely Bengal gram (Cicer arietinum), tur dal (Fragaria), green gram (Vigna radiate), black gram (Vigna mungo) and red gram (Caroluslinnaeus) were used in the study. Color features were extracted to develop identification and classification models. The results from this study indicated the potential application of the color imaging system for automated monitoring of the cooking process in the food industry. De-Silva et al. (2005) proposed a method based on a watershed segmentation algorithm to subgroup color images of food samples (pizza, butterfly cakes, and biscuits) into groups of similar color characteristics. Each segment in an image was individually analyzed and its color characteristics were compared with a threshold value for each food product. Samples falling below the threshold value were rejected. The developed algorithm was able to detect the color changes in the food samples with very high accuracy. Savakar and Anami (2009) explored the possibility of using the color imaging system for classification of grain into ten grain types; classification of grapes into ten varieties; classification of mangoes into five varieties; classification of jasmine flowers into ten varieties. Color and textural features (morphological features were also added for grapes) were given as input to the back propagation neural networks (BPNN). The grains were classified with the highest accuracy. Mangoes gave the lowest accuracy due to the similar color and texture of different varieties. Du and Sun (2008) used a color machine vision system to classify the pizza base, sauce spread, and topping. Image features (shape features of the pizza base and color features of pizza sauce spread and topping) were extracted and given as input to a support vector machine (SVM) for classification (dimensionality of the color features was reduced before being given an input to the SVM). The results showed the ability of color imaging systems in the automatic multi-classification of pizza base, sauce spread, and toppings. Choudhary et al. (2008) used a color imaging system to classify Canada Western Red Spring wheat, Canada Western Amber Durum wheat, barley, oats, and rye. Wavelet, color, morphological, and textural features were extracted from the color images of non-touching single kernels and classification algorithms were developed by LDA and QDA using different combinations of extracted features. Linear discriminant classifiers using combined wavelet, color, morphological, and textural image features as input gave the highest classification accuracy (89.4–99.4%). Color images do not provide information about the chemical composition and its distribution in the object and rely mainly on external surface features of the object under investigation. Hyperspectral imaging provides the spectral information in a spatially resolved manner, by scanning a sample at hundreds of narrow wavelengths. Hyperspectral imaging has been successfully applied in pharmaceutical formulations to analyze compound distribution and quantify active pharmaceutical ingredients (API). This technique has demonstrated high potential for real-time automated quality monitoring in the agriculture and food processing

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 123 industries. Hyperspectral imaging has been used to detect fecal and ingesta contamination on poultry carcasses (Lawrence et al., 2003b). The poultry carcasses were scanned using an NIR hyperspectral imaging system and also an NIR spectrophotometer. Spectral data were analyzed using PCA scores and loadings to identify significant wavelengths. Image ratios at significant wavelengths (434, 517, 565, and 628 nm) in various wavelength combinations were used to obtain wavelength ratio images. Histogram stretching and fecal thresholding were applied in the mapped ratio images for qualitative and quantitative analyses, and 100% accuracy in a qualitative model and more than 96% in quantitative model were achieved. Use of only four wavelengths reduced the dimensionality of hyperspectral data, making it multispectral analysis for real-time applications. Dimensionality of hyperspectral data can be reduced by applying the multivariate image analysis (MVI) technique, which uses PCA to identify significant wavelengths for multispectral analysis. Multivariate image analysis will remove the need for additional NIR spectroscopic instruments and will overcome system variation effects. Chao et al. (2008) developed a high throughput on-line spectral imaging system for wholesomeness inspection of chicken in a high speed commercial chicken processing line. The line-scan system consisted of an electron-multiplying charge-coupled device (EMCCD) detector and a line-scan spectrograph. Hyperspectral images of wholesome and unwholesome chicken carcasses were acquired in reflectance mode and converted into relative reflectance using a standard. The relative reflectance images were then analyzed to optimize the region of interest (ROI), the region providing the highest spectral differences between wholesome and unwholesome chicken carcasses across all wavebands, and significant wavebands were identified for multispectral analysis. In the multispectral analysis significant wavelengths and two-wavebands ratios were used and classification algorithms were developed by applying a fuzzy logic membership functions that identified 99% wholesome and 94% unwholesome chickens at a minimum sorting speed of 140 birds per minute (BPM). The system was capable of reaching food safety standard and high speed production requirements, reaching thus increasing production efficiency. Kim et al. (2002) used multispectral imaging, based on significant wavelengths identified from hyperspectral imaging data, to detect fecally contaminated apples. A simple threshold was applied to an NIR band (significant wavelength) image to create a mask for removing the background. Then PCA was applied to background removed images to discriminate the uncontaminated apple surface area from contaminated surface area with the help of principal component (PC) score images (pseudo color maps of PC scores). Lu (2003) used NIR hyperspectral imaging to detect the bruises on apple surfaces. Fresh apples and bruised apples were scanned and a minimum noise fraction (MNF) transform and PCA were applied to the selected ROI. A transformed image was developed by multiplying first and third PC images. Mean pixel intensity values in the transformed image corresponding to the dark and light areas in MNF image were calculated for the comparison. The areas on the bruised surfaces had higher mean pixel intensity. Discrimination accuracy of the system was affected by apple variety, bruise development stage (from first to 47th

© Woodhead Publishing Limited, 2013

124 Robotics and automation in the food industry Table 5.1 Applications of spectroscopic techniques in the food industry Technique

Product

Analysis

Reference

IR

Food products

Quality

IR IR

Apple Milk

Adulteration Nutritional parameters Authenticity

Fagan and O’Donnell (2008) Sivakesava et al. (2001) Inon et al. (2004)

Raman

Milk, rice, olive oil, honey Raman and NIR Complex food Raman Honey Raman Oil NIR Food products NIR Virgin olive oil NIR NIR

Fruit and vegetables Food/feed

NIR

Milk

NIR

Apricot juice

UV

Fruits and oil

Fat composition Fructose and glucose Lipid oxidation Quality AV, bitter taste, fattyacid composition SSC, quality Moisture, protein, fat, oil Quality and composition Organic acid composition Adulteration

Meurens (2003 Afseth (2005) Batsoulis et al. (2005) Muik et al. (2005) Huang et al. (2008) Marquez et al. (2005) Nicolai et al. (2007) Benson (2003) Kawamura et al. (2007) Chen et al. (2006) Meurens (2003)

Table 5.2 Applications of fibre optic techniques in the food industry Technique

Product

Analysis

Reference

NIR spectrometer Color sensor NIR spectroscopy VIS reflection spectroscopy Fiber optic spectrometer

Cheese

Whey-fat concentration

Castillo et al. (2005)

Cheese Cheese

Curd syneresis Curd moisture content and solids in whey Food products Quality

Lewis et al. (2008)

Meat

Alvarez et al. (2009)

Emulsification process monitoring

Everard et al. (2009) Mateo et al. (2010)

day of bruise development), and spectral resolution. Similar parameters should be taken into consideration in future model developments. Chilling injury in cucumbers was also detected by hyperspectral imaging by applying both spectral and image analyses (Liu et al., 2006). A dual-band image ratio (811 nm/756 nm) gave better discrimination than PC score images and correctly classified more than 90% of the cucumbers. ElMasry et al. (2007) used hyperspectral imaging to predict moisture content, total soluble solids, and pH in strawberries. The ROI was selected after removing the background by creating a binary mask, and then averaged spectra from ROI in each hypercube were obtained. Averaged spectra were used in developing

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 125 Table 5.3 Applications of image sensors in the agricultural and food industries Technique

Product

Analysis

Reference

Computer vision

Food

Quality

Computer vision

Food

Color imaging Hyperspectral imaging

Learning techniques Pizza Classification Boiled food grains Identification and classification Food products Classification Grains, fruits, and Classification flowers Food products Quality Poultry Quality

Brosnan and Sun (2004) Du and Sun (2006)

Hyperspectral imaging Hyperspectral imaging

Chicken Apple

Du and Sun (2008) Anami and Burkpalli (2009) De-Silva et al. (2005) Savakar and Anami (2009) Zheng et al. (2006) Lawrence et al. (2003b) Chao et al. (2008) Kim et al. (2002)

Hyperspectral imaging Hyperspectral imaging Hyperspectral imaging

Apple Cucumber Strawberry

Hyperspectral imaging Hyperspectral and color imaging

Tomato Wheat

Color imaging Color imaging Color imaging Color imaging

Wholesomeness Fecal contamination Bruise detection Chilling injury Moisture, total soluble solids, pH Ripeness Damage detection

Lu (2003) Liu et al. (2006) ElMasry et al. (2007) Polder et al. (2003) Singh et al. (2010a, 2010b)

PLS calibration models. Multilinear regression (MLR) prediction models were also developed using β coefficients from PLS analysis. Texture analysis was done to classify strawberries into unripe, ripe and overripe stages. The variation in the texture was associated with light intensity or gray levels. A color image for each sample was formed by extracting red (650 nm), green (500 nm), and blue (450 nm) band images from a hypercube. Color images were then analyzed by graylevel co-occurrence matrix (GLCM) to determine textural features and used for classification by discriminant analysis. Hyperspectral imaging has been also been investigated for sorting tomatoes (Polder et al., 2003). The dimensionality of hyperspectral data was reduced by independent component analysis (ICA), also known as a blind source separation technique. Unsupervised classification models were developed by pseudo-color mapping of independent component scores similar to PCA scores. The ICA needs to specify the independent components before starting computation, this requires a longer time to converge, and may not be suitable for rapid online inspections. Hyperspectral imaging has been investigated for its potential applications in the grain handling and processing industries for grain quality evaluation. Singh et al. (2010a, 2010b) used hyperspectral imaging to detect midge-damaged and insect-damaged wheat kernels. Multivariate image analysis was used to identify significant wavelengths, and statistical and histogram features from the significant wavelength NIR image were extracted

© Woodhead Publishing Limited, 2013

126 Robotics and automation in the food industry and given as input to statistical classifiers and BPNN classifiers. Performance of the classifiers was compared with color imaging systems. Color, textural, and morphological features from the color images of the healthy damaged wheat kernels acquired using the color imaging system were extracted and used in classification. The highest classification accuracy was achieved by combining the NIR and colorimaged features.

5.4

Future trends

For on-line measurements several factors, such as near uncontrolled sample presentation, variation in the distance between sample and sensor, variable sample amount passing the sensor, non-uniform sample (appearance, particle size, thickness, moisture), and variable ambient conditions (temperature, humidity, light intensity), may affect the performance of the calibration/classification model (Benson, 2003). Therefore, on-line systems should be designed to withstand these factors, and robust calibration classification models should be developed using a representative sample incorporating large sample/product variation. Composition and visual characteristics of agricultural products are influenced by growing region, growing season, and growing year. Therefore, these parameters should be incorporated into calibration model development, hence requiring a large representative sample. The reference data used in training the calibration model should also be very accurate otherwise the prediction model will give inaccurate results.

5.5

Conclusion

Selection of a technique for automated process monitoring in the food industry depends on the specific type of application – food material, type of analysis, ease in sampling, rate of process, ambient condition, state of material (solid, liquid, hot, cold, etc.). Infrared spectroscopy is not very suitable for in/on-line quality and compositional analysis of food products for several reasons, including very low reflectivity of solids above 2500 nm (hence poor signal-to-noise ratio), intense and fine absorption bands that are difficult for quantitative analysis (Benson, 2003), and sample preparation requirement. NIR is considered the most suitable spectral region for in/on-line analysis in the food industry allowing rapid and multi constituent analysis (e.g., moisture, fat, oil, and protein, fiber, starch, sugar salt, and solid content) simultaneously with near zero sample preparation. Fiber optic sensors have demonstrated potential to scan (in back scatter mode) the solid food materials in industrial food processing set ups using ambient (e.g., heat) resistant fiber probes that can penetrate the sample without causing any damage to the products. Machine vision systems using color cameras have been successful in sorting/analyzing food products using their surface characteristics (color, shape and size). Multispectral imaging systems based on wavelengths selected from hyperspectral imaging can inspect and analyze food products at high speed and

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 127 provide reflectance/absorbance information with spatial labeling. With advances in optical instrumentation, computer hardware, and computational methods, optical sensing and spectroscopic technique are poised to become an objective tool for automated food safety and quality inspection in the agri-food industry.

5.6

References

Afseth N K (2005), ‘Raman and near-infrared spectroscopy for quantification of fat composition in a complex food model system’, Appl Spectrosc, 59, 1324–1332. Alvarez D, Castillo M, Payne F A and Xiong Y L (2009), ‘A novel fiber optic sensor to monitor beef meat emulsion stability using visible light scattering’, Meat Sci, 81, 456–466. Anami B S and Burkpalli V C (2009), ‘Color based identification and classification of boiled food grain images’, Int J Food Eng, 5(7), Article 7. Batsoulis A N, Siatis N G, Kimbaris A C, Alissandrakis E K, Pappas C S, Tarantilis P A, Harizanis P C and Polissiou M G (2005), ‘FT-Raman spectroscopic simultaneous determination of fructose and glucose in honey’, J Agri Food Chem, 53, 207–210. Benson I B (2003), ‘Near infra-red absorption technology for analyzing food composition’, in Lees, M, ed., Food Authenticity and Traceability, Cambridge, UK: Woodhead Publishing Limited, 101–130. Brosnan T and Sun D W (2004), ‘Improving quality inspection of food products by computer vision – a review’, J Food Eng, 61, 3–16. Castillo M, Payne F A, Lopez M B, Ferrandini E and Laencina J (2005), ‘Optical sensor technology for measuring whey fat concentration in cheese making’, J Food Eng, 71, 354–360. Chao K, Yang C C, Kim M S and Chan D E (2008), ‘High throughput spectral imaging system for wholesomeness inspection of chicken’, Appl Eng Agri, 24, 475–485. Chen J Y, Zhang H and Matsunaga R (2006), ‘Rapid determination of the main organic acid composition of raw Japanese apricot fruit juices using near-infrared spectroscopy’, J Agri Food Chem, 54, 9652–9657. Choudhary R, Paliwal J and Jayas D S (2008), ‘Classification of cereal grains using wavelet, morphological, colour, and textural features of non-touching kernel images’, Biosys Eng, 99, 330–337. Cooper J W (1980), Spectroscopic Techniques for Organic Chemists, New York, NY: John Wiley & Sons. De-Silva L C, Pereira A and Punchihewa A (2005), ‘Food classifications using color imaging’, In Conference on Image and Vision Computing, New Zealand (ivcnz) University of Otago, Dunedin. Du C J and Sun D W (2008), ‘Multi-classification of pizza using computer vision and support vector machine’, J Food Eng, 66, 234–242. Du C J, and Sun D W (2006), ‘Learning techniques used in computer vision for food quality evaluation: a review, J Food Eng, 72, 39–55. ElMasry G, Wang N, ElSayed A and Ngadi M (2007), ‘Hyperspectral imaging for nondestructive determination of some quality attributes for strawberry’, J Food Eng, 81, 98–107. Everard C D, O’Callaghan D J, Mateo M J, Castillo M, Payne F A and O’Donnell C P (2009), ‘The use of colour parameters derived from an online fibre-optic sensor to monitor curd syneresis during cheese making’, J Food Eng, 94, 1–6. Fagan C C and O’Donnell C P (2008), ‘Application of mid-infrared spectroscopy to food processing systems, in Irudayaraj J and Reh C, eds, Nondestructive Testing of Food Quality, Oxford, UK: Blackwell Publishing Ltd, 119–142. Geladi P and Grahn H (1996), Multivariate Image Analysis, Chichester, UK: John Wiley & Sons.

© Woodhead Publishing Limited, 2013

128 Robotics and automation in the food industry Huang H, Yu H, Xu H and Ying Y (2008), ‘Near infrared spectroscopy for on/in-line monitoring of quality in foods and beverages: a review’, J Food Eng, 87, 303–313. Inon F A, Garrigues S and Guardia M D L (2004), ‘Nutritional parameters of commercially available milk samples by FTIR and chemometric techniques’, Analytica Chimica Acta, 513, 401–412. Kawamura S, Kawasaki M, Nakatsuji H and Natsuga M (2007), ‘Near-infrared spectroscopic sensing system for online monitoring of milk quality during milking’, Sens Instrum Food Quality Safety, 1, 37–43. Kim M S, Lefcourt A M, Chao K, Chen Y R, Kim I and Chan D E (2002), ‘Multispectral detection of fecal contamination on apples based on hyperspectral imagery: Part I. Application of visible and near-infrared reflectance imaging’, Trans ASAE, 45(6), 2027–2037. Kizil R and Irudayaraj J (2008), ‘Applications of Raman spectroscopy for food quality measurement’, in Irudayaraj J and Reh C, eds, Nondestructive Testing of Food Quality, Oxford, UK: Blackwell Publishing Ltd, 143–163. Lawrence K C, Park B, Windham W R and Mao C (2003a), ‘Calibration of a pushbroom hyperspectral imaging system for agricultural inspection’, Trans ASAE, 46(2), 513–521. Lawrence K C, Windham W R, Park B and Buhr R J (2003b), ‘A hyperspectral imaging system for identification of faecal and ingesta contamination on poultry carcasses’, J Near Infra Spectrosc, 11, 269–281. Lewis E, Sheridan C, O’Farrella M, Flanagan C, Kerry J F and Jackman N (2008), ‘Optical fibre sensors for assessing food quality in full scale production ovens – a principal component analysis and artificial neural network based approach’, Nonlinear Analysis Hybrid Systems, 2, 51–57. Liu Y, Chen Y R, Wang C Y, Chan D C and Kim M S (2006), ‘Development of hyperspectral imaging technique for the detection of chilling injury in cucumbers; spectral and image analysis’, Appl Eng Agri, 22(1), 101–111. Lopez-Higuera J M (2002), ‘Introduction to fibre optic sensing technology’, in LopezHiguera J M, ed, Handbook of Optical Fibre Sensing Technology, Chichester, UK: John Wiley & Sons Ltd, 1–21. Lopez-Higuera J M and Madruga F C (2002), ‘Photodetectors for sensing’, in LopezHiguera J M, ed, Handbook of Optical Fibre Sensing Technology, Chichester, UK: John Wiley & Sons Ltd, 149–164. Lu R (2003), ‘Detection of bruise on apples using near-infrared hyperspectral imaging’, Trans ASAE, 46(2), 523–530. Marquez A J, Diaz A M and Reguer M I P (2005), ‘Using optical NIR sensor for on-line virgin olive oils characterization’, Sensor Actuators, 107, 64–68. Mateo M J, O’Callaghan D J, Everard C D., Castillo M, Payne F A and O’Donnell C P (2010), ‘Evaluation of on-line optical sensing techniques for monitoring curd moisture content and solids in whey during syneresis’, Food Res Int, 43, 177–182. Meurens M (2003), ‘Spectroscopic techniques’, in Lees M, ed, Food Authenticity and Traceability, Cambridge, UK: Woodhead Publishing Ltd, 184–196. Muik B, Lend B, Molina-Diaz A and Ayora-Canada M J (2005), ‘Direct monitoring of lipid oxidation in edible oils by Fourier transform Raman spectroscopy’, Chem Phys Lipids, 134, 173–182. Naes T, Isaksson T, Fearn T and Davies T (2002), A User-friendly Guide to Multivariate Calibration and Classification, Chichester, UK: NIR Publications. Nicolai B M, Beullens K, Bobelyn E, Ann P, Saeys W, Theron K I and Lammertyn J (2007), ‘Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: a review’, Postharvest Bio Tech, 46, 99–118. Osborne B G, Fearn T and Hindle P H (1993), Near Infrared Spectroscopy in Food Analysis, Singapore: Longman Singapore Publishers.

© Woodhead Publishing Limited, 2013

Optical sensors and online spectroscopy for automated quality inspection 129 Polder G, Van-der Heijden G W A M and Young I T (2003), ‘Tomato sorting using independent component analysis on spectral images’, Real-Time Imaging, 9, 253–259. Rinnan A, Norgaard L, Van den Berg F, Thygesen J, Bro R and Engelsen S B (2009), ‘Data pre-processing’, in Sun D W, ed, Infrared Spectroscopy for Food Quality Analysis and Control, London, UK: Academic Press, 29–50. Rogers A J (1992), ‘Optical-fiber sensors’, in Wagner, E, Dandliker R, and Spenner K, eds, Optical Sensors, Weinheim: VCH Publications, 355–398. Savakar D G and Anami B S (2009), ‘Recognition and classification of food grains, fruits and flowers using machine vision’, Int J Food Eng, 5(5), Article 14. Singh C B and Jayas D S (2010), ‘Spectroscopic techniques for fungi and mycotoxins detection in agriculture and food’, in De Sager S, ed., Determining Mycotoxins and Mycotoxigenic Fungi in Food and Feed, Cambridge: Woodhead Publishing, 401–414. Singh C B, Jayas D S, Paliwal J and White N D G (2010a), ‘Detection of midge-damaged wheat kernels using short-wave near-infrared hyperspectral and digital colour imaging’, Biosys Eng, 105, 380–387. Singh C B, Jayas D S, Paliwal J and White N D G (2010b), ‘Identification of insect-damaged wheat kernels using short-wave near-infrared hyperspectral and digital colour imaging’, Comp Elect Agric, 73(2), 118–125. Sivakesava S, Irudayaraj J M K and Korach R L (2001), ‘Detection of adulteration in apple juice using mid infrared spectroscopy’, Appl Eng Agric, 17, 815–820. Zheng C, Sun D W and Zheng L (2006), ‘Recent developments and applications of image features for food quality evaluation and inspection – a review’, Trends Food Sci Tech, 17, 642–655.

© Woodhead Publishing Limited, 2013

6 Supervisory Control and Data Acquisition (SCADA) and related systems for automated process control in the food industry: an introduction J. F. Holmes and G. Russell, Georgia Tech Research Institute, USA and J. K. Allen, The University of Oklahoma, USA

DOI: 10.1533/9780857095763.1.130 Abstract: Supervisory Control and Data Acquisition (SCADA) is becoming the standard for industrial control systems. Despite the basics of SCADA dating back to the 1960s, this control technique has only become more evident in food processing environments over the last decade. The recent adoption of SCADA is due to increases in the cost of labor, reductions in the cost of control systems, and the general advances in food processing systems requiring more advanced control systems. This chapter provides an overview of SCADA, some high-level examples, and what should be expected of those with plans to implement a SCADA system into a food processing environment. Key words: Supervisory Control and Data Acquisition (SCADA), remote terminal unit (RTU), human machine interface (HMI), food processing, industrial controls, automation, Georgia Tech Research Institute (GTRI).

6.1

Introduction to Supervisory Control and Data Acquisition

Supervisory Control and Data Acquisition, better known as SCADA, is a term used to describe a set of industrial control standards that have been evolving over several decades to improve the control of complex processes. Initially an extension of programmable logic controllers (PLC), SCADA has evolved to become highly integrated within systems from manufacturing facilities to massive utilities, such as water processing facilities or nuclear power plants. Although implemented in several food

© Woodhead Publishing Limited, 2013

SCADA and related systems for automated process control 131 processing facilities, many have been slow to adopt such new technologies, because of the complexity of such integrations and justification of the cost-to-benefit ratio. 6.1.1 Need for SCADA and high-level overview The need for more complex control systems is driven by the increasing emergence of control technologies within industrial systems and the demand for systems to become increasingly more efficient. Due to this high level of complexity, it is impossible for humans alone to control the operations in a way that can satisfy the demands of the end user. For instance, a high-volume operation where a recipe is to be controlled carefully for quality purposes may require decisions to be made each second as various flow rates and temperatures change within the process. SCADA addresses this by creating a master terminal unit (MTU) that monitors several lower-level operations, referred to as RTUs (remote terminal units). These RTUs typically operate at high data rates within the local subsystem, and report process information to the MTU so that parameters can be adjusted to streamline the overall process. It is at the MTU where the high-level control algorithms can be implemented to optimize the performance of the system. For example, consider a typical high-volume baking operation where bread is being produced. In this case, there are several steps that must occur throughout the facility including: (1) the transfer of flour and water into a mixer, (2) dough mixing and kneading, (3) fermentation, (4) forming, (5) baking, (6) slicing, and (7) packaging (Matz, 1991). One can imagine the number of machines required to facilitate this operation, which will include conveyors, mixers, slicers, ovens, and packaging machines, among many others. Also, one can imagine the many parameters that affect this operation, such as water temperature, dough baking temperature, baking time, color of cooked goods, temperature of the facility, cost of power, storage space, packaging materials, and slicing operation speed, to name a few. The complexity involved in managing this operation is tremendous, as any parameter may have a considerable impact on the quality of the finished product. The timeliness of the decisions is also critical to quality. For example, if the bread color becomes too dark leaving the oven, all products in the oven may become unusable, resulting in the loss of thousands of products. It is apparent that a high-level controller could add considerable value to this process. Figure 6.1 illustrates a high-level block diagram of how SCADA could be implemented in a typical baking operation. Notice in Fig. 6.1 that the term human machine interface (HMI) is introduced as an extension of the main terminal unit. At some level, the need for a human operator is required to monitor the overall process. However, several HMIs can be present throughout the process with varying levels of permissions and control that enable humans to either enter process data, monitor the performance of subsystems, or to make changes as required. 6.1.2 General benefits and drawbacks of SCADA The previous example creates a framework in which several process improvements are immediately comprehensible. However, these benefits cannot be considered without also considering the negative aspects a SCADA system can introduce.

© Woodhead Publishing Limited, 2013

132 Robotics and automation in the food industry Human machine interface (HMI)

MTU server

RTU

RTU

RTU

RTU

Conveyors

Mixer

Fermentation

Forming

RTU

RTU

Packaging

Slicer

SENSOR Water temp.

RTU Oven controller SENSOR

SENSOR

Oven temp.

Baking color

SENSOR Label inspection

Fig. 6.1 Notional SCADA network for baking process.

The following list covers the major points to consider when deciding whether to implement a SCADA system or not (Bailey and Wright, 2003): Benefits • • • • • • • • •

storage of large amounts of data, customization of data displayed, ability to integrate thousands of sensors, ability to incorporate advanced control algorithms, ability to monitor several types of data, ability to view data remotely, better consistency of products, automatic data logging to free up employees for other tasks, automated warning systems.

Drawbacks • system complexity increases in software as well as hardware, • possible need for new skills within an organization, • new security threats created that can compromise the safety of the process or proprietary process data, • overall cost of implementation. As noted in previous sections, it is clear that an integrated SCADA system can allow a process to operate at a whole new level of efficiency yet the drawbacks, such as security threats, are not to be overlooked.

© Woodhead Publishing Limited, 2013

SCADA and related systems for automated process control 133

6.2

History of SCADA

The idea for SCADA came out of the need for a technology that could monitor numerous different weather and atmospheric stations and relay that information to a centralized station in order to better predict weather conditions. Soon the potential benefits to industry were seen and utility companies started to use SCADA systems to control critical portions of infrastructure in 1960s (Ezell, 1998). By the 1980s, both water and power companies were deploying SCADA on a large scale, controlling everything from Extra High Voltage (EHV) power substations (Woods and Serafin, 1982) to the flow in petroleum pipelines (Pharris and Kolpa, 2007). During this time, each application had to have custom-written protocols to control the communication between the many RTUs and the MTU. This soon led to numerous different protocols being implemented that were not compatible with one another. This state of muddled communication protocols remained until the early 1990s when a clear need for an open SCADA communication protocol was realized. During this time period, two competing protocols were developed. During this time period, two competing protocols were developed: IEC 870 and DNP3. The first was introduced by the International Electrotechnical Commission (IEC), and then the second by Harris Controls Division, Distributed Automation Products (Clarke et al., 2004). The communication standard developed by the IEC was eventually termed IEC 60870.5 and was designed to be very general so as to be applicable in many different SCADA applications. The standard developed by Harris Controls Division is called the Distributed Network protocol Version 3.0 (DNP3). DNP3 was focused more on specific industries such as electricity and oil, than IEC 60870.5. Since the protocol’s inception, DNP3 has become more widely accepted and supported by industry around the world, excluding Europe where IEC 60870.5 is still a major competitor. With the proliferation of the Internet during the late 1990s and early 2000s, security of the communication system became the priority. The need for a focus on security in SCADA systems was often pushed forward by the government (Barr, 2004) with numerous documents being written on the subject. The current state of SCADA security is discussed in more detail in a later section.

6.3

SCADA standards and applications

Tracing SCADA back to its roots reveals the magnitude of advances made in the past several decades. The primary shift has been from a number of proprietary packages and communication protocols to widespread support for several devices and process types. The result is a large number of integrators offering a broad range of support for several devices. The other convergence point is the protocols that SCADA networks are built upon. Three protocols have become the industry standards due to their open nature, reliability, and vendor support – IEC60870-5101/104, IEC61850, and DNP3. Lower level control protocols, such as Modbus and Allen Bradley DF1, still exist among many RTUs, but the high-level control executed by SCADA software packages typically support devices using these lower level protocols (Kalapatapu, 2004).

© Woodhead Publishing Limited, 2013

134 Robotics and automation in the food industry When considering SCADA software, it is expected that the following features be present: • support for multiple device protocols (Modbus Organization 2006), Profibus (International 1998), etc.), • PC-based implementation, • remote terminal support, • easy addition of more devices to the network, • fully customizable graphical user interface (GUI) for HMI, • alarm definition, • open access to historical data, • maintenance, • security measures present.

6.3.1 Examples of installations in industry The reliability, flexibility, and scalability of SCADA have made it an attractive choice when automating complex systems. SCADA has already been used successfully in a broad range of industries, ranging from energy production to factory automation to agriculture. In each case, the specific application required that various types of hardware be connected and communicate efficiently in order to perform their operation. The following case studies will dive into the specific challenges faced in each of the applications and demonstrate the performance gains by applying SCADA.

6.3.2 Thermal power plant Thermal power plants are complex systems involving numerous boilers, turbines, and pumps, which need to be monitored and controlled in order to achieve peak performance. Security is also of paramount concern in critical infrastructure facilities like power plants. This level of security and complexity combined with the requirement that each component functions efficiently independently of the others as well as in conjunction with the other components, makes SCADA an appealing solution (Lakhoua, 2009). The implementation of SCADA in a power plant prevents information overload for control workers and allows the consolidation and interpretation of data, to get a better global perspective on the operations and forecast possible problems. One foreseeable future advance in power regulation is to use SCADA to integrate the power plant with a smart grid to allow for real-time modulation of power production based directly upon current demand.

6.3.3 Water treatment implementation SCADA continues to be an elegant solution for control of utilities, as is the case with the example of a water treatment facility in Hawaii. Designed in 1975, the

© Woodhead Publishing Limited, 2013

SCADA and related systems for automated process control 135 Schofield Wastewater Treatment Plant was controlled manually with heavy use of man power aided by alarms on various controls within the operation. A detailed case study was generated to illustrate how a SCADA-based system could be integrated within schedule and budgetary estimates. In this particular example, Allen Bradley PLCs were used with a Rockwell Automation SCADA monitoring system. The duration of the effort was 24 months and the cost of the installation was $575 000 US, and more importantly, the installation did not result in any downtime of the facility (Rutemiller, 2006).

6.3.4 Computer production Global computer manufacturers have used SCADA to solve some of the problems associated with producing customizable computers quickly and efficiently in large volumes. Often, as is the case in many other industries, custom applications created for specific hardware are produced with little to no thought as to how they will interface with the other machinery in a factory. This creates unnecessary complexity and makes it difficult to optimize the manufacturing process. This is an excellent case for SCADA to be used to improve operating efficiency and allow for more detailed control over the entire production process (Swartz et al., 2007). In a particular computer production facility, the company needed to build customized computers on a rapid assembly line. The process needed to be controlled remotely, so that a user could place an order for a PC online and the computer would begin being assembled quickly in a factory around the globe. The assembly line was fully automated from initial installation through distribution using SCADA. The company also wanted to be able to monitor the assembly process remotely from both computer and Internet enabled telephones. The combination of monitoring the process and adding new items to the assembly line remotely caused the security of the system to be put at the forefront. Security was built in through a layered approach, with every component that connected to the Internet first passing through the company’s firewall. Increased security precautions were taken in regard to the mobile phones because the risk for loss or theft was considerably higher than for the other PC-based monitoring systems. Once this new SCADA system was installed, the company found increased productivity and reduced downtime due to the increased visibility of each of the systems within the plant. Maintenance workers would often perform fixes before workers on the assembly line knew of a problem, improving not only reliability but system safety. One other advantage that was found due to the installation of the SCADA system was the ability to monitor the work in progress in real time, providing both management and customers with a better understanding of the production process. After the SCADA system was installed, the company considered it to have a large return on investment, with over $3.7 million being saved annually per factory via the increased operations visibility alone with the total annual return on an investment being in excess of $40 million for the company.

© Woodhead Publishing Limited, 2013

136 Robotics and automation in the food industry

6.4

SCADA in food processing

The general nature of SCADA has been covered in detail, but there are specific concerns to the food processing industry that have affected the adoption and implementation of SCADA. A primary concern is the prevalence of low profit margins within the food industry. In the case of poultry processors, the recent acquisition of Gold Kist, Inc. – a major producer of poultry products – by Pilgrim’s Pride, Inc. can largely be attributed to the rising cost of feed in North American markets (Reuters, 2006). Because profit margins are much lower in food processing than in many other industries, great care has to be taken prior to the costly implementation of full scale SCADA integration. Additionally, due to the nature of food processing operations, a large information technology (IT) staff may not be present to support SCADA systems, meaning the investment required could increase substantially. Lastly, concerns with security are important due to food safety oversight and the potential negativity that can result from a large recall. Despite these challenges, many food processors are beginning to implement SCADA control systems in their operations in order to take advantage of the many benefits that they can provide. The largest benefits that SCADA can bring to the food processing industry are reductions in labor and increases in quality. It is clear that labor reduction increases the competitive advantage of an operation, but this cannot be achieved at a loss of quality. The perception of the product by the consumer remains an objective and quality also extends into the area food safety. Organizations such as the Food Standards Agency, or the Food and Drug Administration report on major recalls that can affect millions of products and consumers. By incorporating better controls and monitoring of a food processing operation, it is feasible that quality and thus safety could be drastically improved while at the same time cutting operating costs.

6.4.1 Cost versus benefit There are many installations of SCADA systems, yet the true cost to install and maintain a complex system needs to be considered prior to making a commitment. The previous example, related to a SCADA installation at a waste-water processing facility in Hawaii, documented a cost of $575K USD, but large-scale systems such as the upgrade to SCADA interfaces for Florida’s District’s Water Management System can be as high as $292M USD (Vann et al., 2002). It is worth mentioning that the latter example included maintenance over 10 years. There are cases where simple SCADA installations, such as those used in irrigation systems, can be as little as $3–22K (2001 US) (Hansen, 2001). The following table provides an idea of the costs related to a small system such as a single plant versus the costs related to a large-scale operation involving SCADA integration over a large number of sites. The key to take away from the presentation of data in Table 6.1 is that the cost of hardware is rather small in comparison to infrastructure updates, IT support, and maintenance. Also, the table is built upon examples

© Woodhead Publishing Limited, 2013

SCADA and related systems for automated process control 137 Table 6.1 Estimates of installation and maintenance costs of SCADA Single facility (USD) Multiple facilities (USD) Non-recurring Installation/labor Hardware Software Professional integration of system Recurring (annually) Software Hardware IT Infrastructure Totals Non-recurring Recurring (annually) Cost of implementation over 5 years

$150K $70K $40K $160K

$24M $3M $6M $13M

$10K $10K $80K $10K

$300K $1.2M $5.3M $12M

$420K $110K $970K

$112M 17M $198M

outsourcing many of the integration tasks whereas using existing internal sources for installation and infrastructure items could reduce overall costs. Keep in mind that Table 6.1 refers to the general estimated cost of a full installation on a complete and highly automated SCADA system. A more practical approach is to slowly integrate new equipment into a SCADA process using a simple yet scalable interface initially. As the system becomes more complex, the MTU Server and HMI can be upgraded to support a larger number of RTUs and new technologies. Going back to the baking example discussed in Section 6.1.1 and described in Fig. 6.1, an initial installation could consist of a basic MTU server, HMI interface and two major RTU devices such as an oven controller and inspection device to inspect the products leaving the oven. All of the benefits of SCADA could be realized, such as real-time data logging, remote monitoring, quality alarms, etc., for the baking process. This could be done at a cost of less than $50K USD considering hardware, software, installation and integration costs. Many of the benefits of SCADA have been discussed, but there are several benefits that are unique to the food processing industry. For instance, oversight by food agencies, such as the United States Department of Agriculture or the World Health Organization often requires a number of metrics to be reported to inspectors. The activity of capturing and reporting these measures can be time consuming and difficult in complex systems. Incorporating history-capture abilities into a SCADA-based network could alleviate some of the effort required in this area and provide great improvements to manual techniques. Benefits related to food safety could provide additional value. With food safety concerns always rising, any additional monitoring of parameters related to food safety could provide significant value to the operator. For example, having tight controls over temperatures in the process can provide more

© Woodhead Publishing Limited, 2013

138 Robotics and automation in the food industry quality to the consumer, better control of pathogens, and optimize energy usage. Benefits related to energy savings can be of utmost importance in food processing due to the nature of operating large ovens, freezers, and temperature-controlled plants. The potential cost savings related to heating venting and air conditioning (HVAC) alone could justify the installation of a SCADA system. 6.4.2 Security and operation concerns A major drawback to SCADA is the potential of vulnerability created by a security breach or other disruption in operation of the MTU. These issues can arise from the following: • malicious attacks, – internal (disgruntled employee, etc.), – external (terrorism), • natural causes, – power surges/outages, – operator error, – hardware failure, – software design flaws. SCADA developers are sensitive to these issues and have incorporated several features to combat many of these threats, such as embedding firewalls and various intrusion detection systems. However, there have been a number of SCADA-related attacks or failures documented (Tsang, 2009). The following list is only a sample: • Intentional – January 2000: Maroochy Shire Sewage Spill: flooding of 264 000 gallons of raw sewage. – March 1997: Worcester Air Traffic Communications Attack: a member of the public switched off the telephone network in the area which compromised air traffic control. – 2000 – Gas Pipelines in Russia: SCADA network compromised, but threat neutralized prior to damage. • Unintentional – August 2005: Automobile Plants in Zotob Worm: 13 automobile plants shut down for 1 h. – August 2003: CSX Train Signaling System and the Sobig Virus: shutdown of critical signal and dispatching systems. These attacks are well documented in critical infrastructure areas, but it is clear that food processing attacks or failures could end with potentially disastrous results. Concerns are the basic encryption techniques used for password-protection services, which are becoming easier to undermine as was demonstrated with using low cost graphics processors to crack seven character passwords in a matter of hours making them ‘hopelessly inadequate’ (BBC, 2010). Additionally, many organizations are offering wireless solutions for industrial control hardware due

© Woodhead Publishing Limited, 2013

SCADA and related systems for automated process control 139 to the obvious convenience of implementation. Although there are proven encryption techniques concerning wireless protocols, wireless transmission of data is inherently less secure than wired transmission of data. Overlooking proper security protocol at a single access point could create a more substantial vulnerability if using a wireless protocol. A breach in security can result in the loss of personal data, corporate data, or affect the throughput of an operation and safety. Although these cases demonstrate major failures in SCADA systems, these systems can be protected if proper security measures are in place. Resources, such as those offered by The Center for SCADA Security from Sandia National Labs or the US President’s Critical Infrastructure Protection Board, can enable more informed decision making in regards to SCADA (Energy, 2002). These security resources often stress that the most important piece of security is to maintain policies and a plan for SCADA security (Stamp et al., 2003). This includes the proper training of personnel, commitment of resources, personnel security controls, designating roles and responsibilities, and creating penalties for those that do not follow protocols (Krutz, 2005). To expand on the commitment of resources, it is paramount that resources are dedicated not only for the installation of SCADA hardware/software but also for future technology procurements. As threats evolve, so must the SCADA system evolve to diminish the effectiveness of those threats.

6.5

Laboratory study: implementation of SCADA

Researchers at the GTRI Food Processing Technology Division generated a small scale SCADA environment for the purpose of evaluating multiple SCADA software packages. This was a simple system involving a supervisory level controller with several control level entities. Through the implementation, an evaluation of three SCADA software packages was generated to highlight specific challenges and benefits of using these programs. The three programs used were Cimplicity, Citect, and VTSCADA. Cimplicity is a product from GE Intelligent Platforms, Citect was recently acquired by Schneider Electric, and VTSCADA is a Trihedral product. Researchers did find that the small scale system could easily be extended to include more devices and demonstrate a higher level of control between multiple pieces of equipment with any of the three software packages. Listing of the three software packages with a brief overview taken from the manufacturers are given below: • CIMPLICITY (Electric, 2011a) – ‘CIMPLICITY is a client/server based visualization and control solution that helps you visualize your operations, perform supervisory automation and deliver reliable information to higher-level analytic applications. With the latest features of CIMPLICITY 8.1, including a powerful graphics engine, dynamic time handling and the add on option digital graphical replay (DGR) 2.1, operators and engineers have the power and security to precisely monitor and control every aspect of their environment, equipment and resources. The results: faster responsiveness, reduced costs and increased profitability’.

© Woodhead Publishing Limited, 2013

140 Robotics and automation in the food industry • Citect (Electric, 2011b) – ‘The CitectSCADA range of automation software has been delivering world-class solutions and services to customers around the world for over 35 years. We specialise in the delivery of solutions that are innovative, reliable, and cost effective. Our offerings connect people in business and industry to real-time information in ways that improve business strategies and decisions’. • VTScada (Trihedral 2011) – ‘Designed for the water and wastewater industry, VTScada is a software tools layer that adds powerful remote asset management features to a visual tag system (VTS) for telemetry applications. Industryspecific reports, auto-generated graphic displays and automatic historical data logging provide a platform for quick application development without compromising on features. Applications range from stand-alone workstation applications to unified, county-wide systems tying plants and SCADA centrals together’. These various SCADA software solutions were compared and contrasted during the laboratory implementation, and the key points are shared in Table 6.2. The three software packages range from a simpler interface with less flexibility when using VTSCADA, to more complex and configurable systems such as Cimplicity and Citect. Depending on the size of the SCADA network, level of proficiency of IT personnel, and other general requirements, an organization has many options available for SCADA software tools. Furthermore, there are several integrators available to assist in all levels of SCADA implementation, with some providers focused on the food and beverage industry.

6.6

Future trends in SCADA

SCADA systems will continue to offer the same advantages in coming years with a focus on easing the burden of maintenance and operation for the user while offering more useful features. It is also imperative that security be at the forefront of SCADA developments because threats constantly evolve. Specifically for food processing applications, the future of SCADA as a technology is not as important as the adoption of SCADA control. Corporations and plant managers must be cognizant of the impact of increased use of SCADA for food processing operations – especially to security as it applies to food safety. Food processing operations will begin to operate more efficiently by using more advanced industrial controls, and the industry will see a shift in competition moving towards more technologically advanced processing plants. Furthermore, the need for more technologically advanced plants will be more prevalent in markets such as Europe and the United States, where wages continue to rise thus driving pressures to incorporate more automation into food processing facilities. Increases in automation will lead to increasing pressures to implement advanced industrial controls, and SCADA will become commonplace in successful food processing operations of the future.

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

General usability comments

Remote level

Control level

• Opens three screens for the programmer to work with: – Project Explorer – Project Editor – Graphics Builder window • Project Explorer allows the programmer to navigate within and between projects • Project Editor allows the user to define and change program parameters • Graphics Builder allows the programmer to design screens for the end user All permit communication using standard protocol (i.e. Modbus over TCP/IP). If product does not support standard protocol, custom drivers can be created to allow for integration. VTS Internet Server states that IE 6 or 7 is required. Uses Webview as webserver when only Multiple parts of the software monitoring. Uses terminal services when support remote monitoring. actual setpoints need to be changed. Simplest program of the three to use. Minimal setup Complex software, yet flexible for CITECT is highly configurable, to generate a Modbus network and read data. integrating custom scripts using the necessitating a certain degree of Once the programmer understands the tag system, Integrator’s Toolkit. The Workbench complexity. The GUI is attractive commands are relatively clear and logical. Good is logically put together, but requires and consistent between all three help file and good technical support. significant learning from the users. Some programming windows. inconsistencies in menu file structure among programming modules.

Routine installation.

Overview

Detailed installation involving SQL database configuration. • Aspects of the program are accessed from • Workbench is the front end of the application pages program, and provides a logically • The programming and run-time environments organized way to access all are blank user screens with a menu button at the programming information bottom left that allows navigation between screens • All parameters are listed in a tree structure (Overview, Alarms, etc.) on the left of the screen, similar to • The programmer and the user can access parameters Windows Explorer on application pages • Screens are created in a separate screen • A configuration toolbox provides tools for defining editor, called CimEdit, that can access program parameters data points as defined in the Workbench • Uses VTS to define all parameters • Uses a tree structure to allow the programmer to define variables

CITECT (v7.10r1)

Routine installation.

CIMPLICITY (v8.00)

Install

VTSCADA (v9.1.03)

Table 6.2 Various attributes of VTSCADA, CIMPLICITY, and CITECT

142 Robotics and automation in the food industry

6.7

References

Bailey, D. and Wright, E. (2003). Practical SCADA for Industry. Burlington, MA, Newnes. Barr, D. (2004). Supervisory Control and Data Acquisition (SCADA) Systems. National Communications System. Chantilly, VA 20151, Communication Technologies, Inc.: 76. BBC. (2010, August 13, 2011). ‘Call to improve password security’. Retrieved July 31, 2011, from http://www.bbc.co.uk/news/technology-10963967. Clarke, G. R., Reynders, D. and Wright, E. (2004). Practical Modern SCADA Protocols. Burlington, MA, Newnes. Electric, G. (2011a). ‘Proficy HMI/SCADA – Cimplicity’. Retrieved July 31, 2011, from http://www.ge-ip.com/products/2819. Electric, S. (2011b). ‘Citect has become part of schneider electric’. Retrieved July 31, 2011, from www.citect.com. Energy, U. S. D. o. (2002). 21 steps to improve cyber security of SCADA networks, Office of Energy Assurance, Office of Independent Oversight and Performance Assurance, from http://www.oe.netl.doe.gov/docs/prepare/21stepsbooklet.pdf. Ezell, B. C. (1998). Risks of cyber attack to supervisory control and data acquisition for water supply. In School of Engineering and Applied Science, Charlottesville, VA, University of Virginia. Masters of Science, Systems Engineering. Hansen, R. Hilton, A., Berger, B., Pullan, W., Gao, Z. and Lee, C. M. (2001). Low-cost automation and SCADA: A Pacific Rim Perspective: 10. 1st Asian Regional Conference, Seoul, Korea, 16-21 September, 2001, pp. B04 International, P. A. P. (1998). PROFIBUS Specification, Order No. .0.32. Normative Parts. Karlsruhe, Germany. European Standard EN 50 170 Volume 2. Kalapatapu, R. (2004). SCADA protocols and communication trends. The Instrumentation, Systems and Automation Society. Reliant Center Houston, Texas, Instrumentation, Systems and Automation Society. Krutz, R. L. (2005). Securing SCADA Systems. Hoboken, NJ, Wiley. Lakhoua, M. N. (2009). ‘Application of functional analysis on a SCADA system of a thermal power plant’. Advances in Electrical and Computer Engineering 9(2): 90–98. Matz, S. A. (1991). Bakery Technology and Engineering. New York, NY, Springer. Modbus Organization (2006). Modbus Application Protocol Specification V1.1b. Hopkinton, MA. Pharris, T. C. and Kolpa, R. L (2007). Overview of the Design, Construction, and Operation of Interstate Liquid Petroleum Pipelines. Oakridge, TN, Argonne National Laboratory. Reuters (2006). Gold kist OKs takeover deal by pilgrim’s pride. Los Angeles Times. Los Angeles, CA. Rutemiller, B. (2006). Implementing a State of the Art Automation Control, SCADA and Information System into an Existing Manually Operated 20 MGD WWTP Below Budget, on Schedule and Without Process Interruption. WEFTEC.06. Dallas, TX. Stamp, J., Dillinger, J., Young, W. and Depoy, J. (2003). Common Vulnerabilities in Critical Infrastructure Control Systems. Albuquerque, NM, Sandia Corporation. Swartz, P., Johnson, C. and Brown, R. (2007). ‘Dell IT implements factory automation and visibility in real time’. Dell Power Solutions (November, 2007): 66–69. Trihedral. (2011). ‘Trihedral software for monitoring & control’. Retrieved July 31, 2011, from www.trihedral.com. Tsang, R. (2009). Cyberthreats, Vulnerabilities and Attacks on SCADA Networks: 23. Vann, A., Beirnes, T. and Lynch, J. (2002). Audit of the Proposed Upgrade/Replacement of the SCADA System. West Palm Beach, FL, South Florida Water Management District. Woods, D. E. and Serafin, R. D. (1982). ‘A multi-microcomputer based distributed front end communications subsystem for a power control center ’. Power Apparatus and Systems PAS-101: 180–184.

© Woodhead Publishing Limited, 2013

7 Gripper technologies for food industry robots T. K. Lien, Norwegian University of Science and Technology, Norway

DOI: 10.1533/9780857095763.1.143 Abstract: This chapter gives an overview of the challenges of gripping food objects, the main answers to these challenges and some indications of possibilities offered by new gripping technologies. Conventional pinch gripping technologies are discussed and newly developed methods based on penetrating, suction and freeze methods are described. The basic governing equations for pinching and suction gripping are presented. The hygienic quality of the different methods is discussed. Finally a qualitative evaluation of the suitability of the different methods in food handling is presented. Key words: pinching grippers, penetrating (needle) grippers, Coanda effect grippers, Bernoulli grippers, Peltier freeze grippers.

7.1

Introduction

Since the invention of the first industrial robot by George DeVol in 1959, the gripper has been an integral part. The whole idea of robots is based on the concept of some non-human device that can perform tasks normally done by humans. It can be in manufacturing or any other human activity where a machine is used to replace the human action. In many of these cases the task involves gripping some object and doing something with it. It might be a simple transportation from one place to another, or a more complex task involving the use of some sort of tool and processing. The modern ideas and implementations of robots have expanded from the original industrial applications for the first robots based on DeVol’s patent. The original industrial robots were truly material handling devices and nothing more. For that reason the gripper was an essential part of the robot. But unlike the robot arm and control system, the grippers have not been the subject of the main research activities in robotics. There has been research on

© Woodhead Publishing Limited, 2013

144 Robotics and automation in the food industry mechanical hands with human-like fingers, but still the bulk of research seems to have been concentrated on robot control theory, arm structures, kinematics and dynamics, and on robot sensory systems. Grippers can be profoundly simple. A device with two mechanically actuated fingers that can pinch an object is all that is needed in many simple handling cases. Since there are so many applications where this gripper is sufficient, the research effort has been spent on the more interesting and mathematically challenging aspects of robotics. The well-functioning, simple two-finger gripper is too simple to challenge the minds of the researchers. But times have changed. All the simple cases are so well studied. Instead the robotic world is facing challenges in handling soft, limp, ‘unstructured’ objects. These are the objects that are found in great numbers in food industry. Handling fruits and vegetables, and meat from sea or land animals, poses problems that are not immediately solvable with the simple gripping technology of the first generation robotics. Slowly, in the period from around the new millennium, insight has been developed into more sophisticated gripping methods and technologies for all of these objects of soft, limp or unstructured character. As these new gripping methods and technologies become well understood and developed to industrial standards, new opportunities for automation in a vast number of industrial applications will appear. In this chapter the term ‘hard gripping’ will refer to any gripping task where the object has a well-defined shape and is strong enough to withstand the force of a pinching or similar gripping method. Typically, gripping a metal screw or plate can be considered to be a hard gripping case ‘Soft gripping’ will be used to characterize any gripping task where the object is of soft materials that can be easily influenced by the force of the gripper, and by any acceleration, and in addition the object may be of irregular shape that will vary from object to object of similar kind. Gripping a piece of beef is a good example of soft gripping. Food objects are often soft-grip objects, even if many food objects could be handled with hard grip technology. This chapter will cover both types of gripping technologies, but it will put most emphasis on soft gripping.

7.2

Gripper challenges in food process automation

In general food materials present challenges that are not normally seen with non-edible products. Some of these challenges are also seen in products other than food, notably softness and limp behaviour of textiles and certain foam plastics and in rubber products. But many of the challenges in food handling are unique when the combination of challenges is considered (Chua et al., 2003). Certain well-developed gripper solutions can be transferred from the non-food sector into food handling. But all special requirements in the food sector have to be considered. This consideration has lead to some interesting and unique gripper solutions for automatic handling of food materials.

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 145 7.2.1 Soft materials Much food material is soft. This comes from the nature of food. Food material is either vegetable or meat. Meat is always soft while vegetable food can be soft or hard. The hard food material causes fewer problems from the gripping side. But all soft vegetable food materials are potential problems in handling. There are two major problems with soft materials: 1. The material changes shape so that a fixed-width pinching device simply loses its grip on the object. 2. The gripping force applied can damage the surface or interior of the object so that quality is reduced. This generally leads to lower selling price or shorter life of the object. For this reason soft materials require a gripping method that eliminates these problems. For meat from land animals the problems of type two are insignificant. Meat is generally robust enough to tolerate normal forces from pinch gripping. The softness does mainly lead to the problem of the first type for meat. This might not matter if it were not for the aggregation of problems due to uneven surfaces and non-uniform shapes. Fruits are in a different category. Soft fruits require very gentle handling. Even semi-hard fruits like apples and tomatoes are very sensitive to denting and local pressure. In general the pinching grip is not suitable for any of these. Other vegetables are easier to handle – nuts and roots of different kinds are generally more tolerant to gripping forces.

7.2.2 Uneven surfaces All naturally grown objects have uneven surfaces. For some specimens the unevenness is only slight, while others have very rough surfaces. The surface quality can influence the range of gripping options. Hard objects with rough surfaces are normally the easiest to grip. Such objects are tolerant to high gripping forces. Soft objects with rough surfaces, on the other hand, present challenges. Kiwi fruits can serve as examples here. These fruits are very sensitive to denting and local pressure, and the skin is rough, making it difficult to use vacuum gripping techniques. Meat from both fish and land animals also presents problems due to uneven surfaces. Vacuum techniques are difficult to apply because of the unpredictable surface structure in many cases.

7.2.3 Non-uniform shapes All organic objects have imperfect geometrical shapes. It is only fabricated objects that can present perfect outer geometry. Some food materials come in this category; most notable are chocolate products that are cast in dies that

© Woodhead Publishing Limited, 2013

146 Robotics and automation in the food industry produce perfectly regular shapes. But all naturally grown objects show some irregularities. Many processed object are similarly non-uniform. Different cuts of meat from animals will always have some irregularity even if the cutting process is set up to produce very regular shapes. Breads, cakes and pastries can be produced to quite regular shapes; however the preference of many customers is that these products should be irregular to give a ‘home-made’ appearance. Non-uniform shape is the biggest challenge for gripper design and methods. But in recent years the development of technology that supports the gripping task has provided many new options for reliable gripping solutions

7.2.4 Hygienic requirements Since food is to be eaten there is an absolute, no compromise, requirement for hygienic treatment of food in harvesting and processing. Contamination of any kind must be avoided. Basically there are three major classes of contamination to be concerned with: 1. Toxic contamination: in the harvesting and processing no toxic substance should come in contact with the food. Traces of toxic substances can adhere to the product and thus be transmitted to the consumer. This implies that all gripping devices and methods should be of non-toxic materials. This is important to consider, since some engineering materials have toxic effects. 2. Bacteriological and fungal contamination: bacteria of all kinds have the possibility of growing where there is organic material, and favourable humidity and temperature. Fungi can also find favourable living conditions in humid areas. These conditions can easily be present in any food processing facility. All gripping methods that rely on physical contact with the object do also generate the risk of leaving small amounts of the handled material on the gripping device. Temperature and humidity will often be favourable for bacteria and fungi on the premises if there are humans in the processing area, due to the requirement for acceptable working conditions for humans. Therefore all gripping devices should be designed so that the accumulation of organic material is reduced to the absolute minimum, and that efficient cleaning methods exist for the equipment. 3. Discolouring: some handling and gripping methods may lead to discolouring, by leaving small traces of material that might not be harmful but will be visible. Local pressure from gripping actions may lead to discolouring because of small changes in material structure. Even if such discolouring has no effect on the nutrient value of the food it will often be considered as a quality defect by the consumer. The first and second classes of hygienic requirements are absolute. The third class is a ‘nice-to-have’ requirement but it does not represent a true health hazard.

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 147

7.3

Gripping physics

All grippers should perform the basic function of transferring the necessary force from the robot arm to the object in order to move the object. Any movement involves acceleration, through gravity and the local movement path. The acceleration in the local path is a function of any change in magnitude and direction of the speed vector. Acceleration due to gravity is always directed perpendicular to the local horizontal plane. For vertical movements this acceleration adds to the local path acceleration for a movement upwards. This will normally give the largest force exerted during a handling operation. Some gripper methods require at least partial enclosing of the object. This is the case with pinching and enclosing grippers. Others act on a single surface of the object. This is the case with pinning, underpressure and surface effect grippers. For food products the following options for force transfer through the gripper exists: • Pinching: the gripper has two or more fingers that apply force to the handled object. Straight fingers rely on friction force between the object and fingers to transfer the necessary force to keep the object in a safe grip during handling. Grooved or circular shaped fingers can also transfer force through local perpendicular force vectors. Since a large force is always exerted on the object this is a typical hard grip. Release is obtained by moving the fingers away from the object. • Enclosing: the gripper relies on fingers with large surfaces that enclose the object partially or completely. The enclosure must be complete in the sense that all possible escape directions from the gripper are closed by suitable surfaces. The required force transmission to the object is obtained from the vector sum of all normal force vectors of the surfaces in contact with the object. In this case there is no force larger than the force of gravity and the acceleration forces from the gripper walls that act on the object so this is a soft grip. • Release is obtained by moving the enclosing surfaces sufficiently away to create an opening large enough for the object to escape. • Pinning, penetration: the gripper has several sharply pointed pins that are pressed into the object in a manner that creates a locking pattern. Usually this means a set of pins pointing inwards toward the central area of the object. There are two types of pins: 1. Short pins for surface penetration of flat objects. 2. Long pins for deep penetration of irregular objects. The objects must be of types that are not harmed by pin penetration. In food processing this can be the case for certain meat products. Release is obtained by pulling the pins back into the gripper body.

© Woodhead Publishing Limited, 2013

148 Robotics and automation in the food industry • Underpressure, suction: these grippers are often called vacuum grippers. They do not however rely on total vacuum, but rather on a certain amount of underpressure in the gripping area. Underpressure means that the local pressure in the gripper area is lower than the ambient pressure. This creates a force due to the pressure difference, which forces the object against the gripping area. Release is obtained by removing the underpressure; sometimes a slight overpressure is required for consistent release. • Surface effects, surface phase transitions: gripping can be obtained by cooling down a contact area to a temperature well below the freezing point of water. When this area is brought in contact with a wet object, a thin ice film is created immediately on the object surface in contact with the cold gripper surface. This ice creates a sufficient holding force for many objects. Meat does normally have enough water content on the surface to enable this gripping action. Release is obtained by heating the contact area, or by slicing away the object by means of a thin knife (Lien and Gjerstad, 2008; Gjerstad et al., 2006).

7.4

Pinching and enclosing grippers

The pinching gripper is the original gripper concept for robots. It is a very simple and robust gripping method, and it is very efficient for a large number of different gripping tasks. There are several variants of this gripping principle, from two stiff-finger pneumatically operated grippers to multiple-jointed finger variants (Wolf et al., 2005). For practical robot applications the two and three stiff-finger variants dominate. Figure 7.1 shows examples of these gripper types.

7.4.1 Two-finger solutions Typical two-finger grippers have the appearance of a prismatic body with two fingers extending from one side. The fingers are either moving in parallel or rotating around a shaft at their base. Normally they are actuated by pneumatic power, but electric and hydraulic actuation does also exist. The fingers are either flat-faced or have circular or v-grooves. Two-finger flat-faced finger grippers rely on friction to generate the necessary force for holding the object. The force required depends on the direction of the acceleration in the movement. Figure 7.2 shows the four different principal force states that can occur depending on the direction of movement. This is called a force-fit grip. In all cases the gripper fingers transmit a force Ff opposing sliding along the gripper surface: Ff

[7.1]

nF µ

where:

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 149 (a)

(b)

Fig. 7.1 (a) A typical two-finger gripper, (b) a typical three-finger gripper.

Ff: the friction force opposing sliding, F: the normal force on the finger face, μ: the friction coefficient between the object and the fingers, n: the number of fingers. To avoid sliding, Ff for a two-finger gripper must satisfy the following conditions depending on the direction of movement: Vertical up: Ff

m ( g + a)

© Woodhead Publishing Limited, 2013

[7.2]

150 Robotics and automation in the food industry Move up

Ff

Move down

Ff

Ff

F

F

Ff

F

F

m(g + a)

Move sideways

m(g − a)

Move sideways Ff F

Ff ma

F

Ff ma

F

mg

Fig. 7.2 The forces acting on an object in different acceleration cases: m is mass of the handled object, a is the acceleration of the object due to the robot movement and g is the acceleration due to gravitation.

Vertical down: Ff

m ( g − a)

[7.3]

Horizontal parallel to finger face: Ff

m g 2 + a2

[7.4]

Horizontal normal to finger face: Ff

[7.5]

mg

For the horizontal movement normal to the finger faces, the acceleration force from the movement will actually increase the total force on one face and reduce it with the same amount on the other face. But the sum of forces will remain constant rendering the equation valid for all accelerations. The grooves will give gripping force through form enclosing action. This is called form-fit gripping. The gripping force is the resultant of the individual normal force vectors from the contact surfaces. The calculation for this type of grip is presented in the section for multi-finger gripping. The grip finger force F for pneumatically operated grippers is normally dependent on the supply pressure to the gripper. In most cases the force will exceed the level for minimum Ff by a very large margin, often several orders

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 151

Fig. 7.3 Two-finger servo gripper with v-groove finger for object centering.

of magnitude. These grippers will always close to a position where the reaction force from the pinched object equals the finger force F. In the case of handling hard metallic, wooden or plastic components this is of no consequence. For fragile food materials the case is different. The gripping force must often be adjusted to the absolute safe minimum to avoid damage due to local pressure on the objects.

7.4.2 Two-finger servo gripper One solution to the problem of excessive gripping force is the use of an electric servo gripper. This type of gripper is controlled to give a specified gripper gap, or to grip with a specified grip finger force F. Externally these grippers differ little from pneumatically operated grippers except that the actuator housing is somewhat bigger, see Fig. 7.3. They are also more expensive, but the extra cost can easily be justified by the higher versatility of this gripper type. The servo gripper requires a digital command specifying the wanted finger gap and grip force if force control is implemented. This command is normally transmitted via serial link from the control system.

7.4.3 Multi-finger grippers Multi-finger grippers are of three main classes: 1. Simple stiff-finger grippers with three or four fingers. 2. Many stiff fingers that partly enclose the object. 3. Human-like finger grippers with jointed fingers.

© Woodhead Publishing Limited, 2013

152 Robotics and automation in the food industry Equations [7.1], [7.2] and [7.3] are valid for grippers with any number of flat-face stiff fingers when the gripper is used with its centreline vertically oriented. These grippers operate partially as form-fit grippers when operated in other positionings. Form-fit grippers rely on partially enclosing the object with the fingers, also sometimes using grooved or contoured fingers. Form-fit gripping does not require high force from the fingers to create the friction force to hold the object. This gripping method relies on the fingers’ ability to withstand the normal force at the contact points at any position and acceleration. This will always lead to lower maximum finger grip force than force-fit gripping. The anthropomorphic (human-like) grippers should ideally mimic human handling best. Many of the grippers in this class have less than the total degrees of freedom that we find in the human hand, only a few laboratory examples come close to mechanically mimicking the human hand. Thus the dexterity of the human hand cannot be fully emulated. Furthermore, the sensitivity of the human hand, through tactile and force feedback through the human nerve system, has proven very difficult to mimic. In addition the mere complexity of the human-like finger grippers has lead to extremely high cost, making them economically unsuited for industrial use at the present stage of development. This might change in the future (Mouri et al., 2002; Thayer and Priya, 2011).

7.4.4 Enclosing grippers Enclosing grippers are designed so that they constitute a nearly closed box that embraces the whole object when closed. Thus they are able to handle all variations of a product as long as it fits inside this box. The shape of this box is normally not a rectangular prism; a ‘boat shape’ is sometimes a more appropriate description. The gripper does not exert a pinching force unless the object completely fills the box volume. In most cases the force to move the object is exerted from the box wall supporting the object in the direction opposite to the combined force vectors of gravitation and acceleration in the local movement. The enclosing gripper will not provide exact positioning and orientation unless the object fills the box completely. In most applications it will therefore be suitable only if the delivery position and orientation is not critical, or if additional alignment devices are used at the delivery station. In many cases of packaging this is no problem since either there is room that allows variation of delivery into the package, or some simple guide walls are used to lead the object into its final position. Figure 7.4 shows an example of an enclosing gripper

7.4.5 Hygienic performance of pinching and enclosing grippers All the grippers of the pinching and enclosing class have some sort of moving mechanical elements outside the gripper base housing. Moving fingers require bearings, guides and shafts passing through walls. These elements are always a concern with respect to hygienic requirements. They should be designed so that there are no narrow slits, grooves or openings to the interior of the gripper

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 153 (a)

(b)

Fig. 7.4 (a) Enclosing gripper, (b) enclosing gripper mechanism that minimizes volume swept in the gripper closing or opening action. (Source: AEW Delford.)

where small particles of the food material and moisture can collect. It is a challenge to design pinching grippers to avoid such particle-collecting areas. In particular parallel finger grippers are difficult to design according to hygienic requirements of this kind because of the guideways necessary for the finger movement. Another aspect is the geometry of the gripper fingers. They should have a smooth surface and no concave areas that will collect particles of any kind. Some gripper fingers have a serrated surface on the actual gripping face. For many food applications this is unsuitable, since the serration grooves are particle collectors that are difficult to keep clean. The mechanism of a gripper needs lubrication. Complete sealing against long-term leakage is difficult to obtain. A better solution is then to use bearing materials that are lubricated by water, or polymer-against-metal bearings that can operate without lubricants.

© Woodhead Publishing Limited, 2013

154 Robotics and automation in the food industry Finally, the materials that will be exposed to the food handled must be of hygienically acceptable quality. In practice that implies stainless steel and polymers as materials for fingers and outer housing walls of a gripper.

7.5

Penetrating (needle) grippers

Penetrating grippers can represent a viable alternative for handling food goods in many cases. They are suitable for foodstuffs that are soft and can be penetrated without harmful effect on their quality. Primarily they are usable for meat, but it is possible that they could be used for soft cakes as well. These grippers rely on the use of several needles that penetrate the surface of the object to create a mechanical grip. This punctures the surface, and the basic requirement for the usage of such grippers is that this surface penetration is allowed. Both hygienic and quality considerations are important here. Some meat processing relies on needles to inject processing fluids like salt, spices or taste enhancers. In these cases, as puncturing of the surface is necessary, additional puncturing from a gripper is then of no consequence. In other cases the puncturing leaves marks that are hardly noticeable. This gripper type is an interesting alternative because it can give a very strong and precise grip. The advantages of a penetrating gripper are: • Single side grip that makes it easy to grip and deliver in confined space. • Very strong grip due to the physical penetration of the gripper into the tissue of the object. • Precise positioning due to the physical interlock between the needles and the object. • High accelerations and transfers speeds can be obtained due to the strong grip obtained. The disadvantages are: • Puncturing of the surface can cause bacterial and other contamination. Mechanisms to ensure cleaning of the needles and the other surfaces that come in contact with the object are mandatory. • Puncturing leaves holes in the object. These holes represent visual quality degradation. In some cases such quality degradation is not accepted. • High acceleration can in some cases lead to local tearing of the tissue of the object.

7.5.1 Short needle, skin penetrating grippers Short needle grippers were developed for textile handling. Figure 7.5 shows an example of this gripper type. The gripper has many needles, typically more than ten, which are extended at a 45° to the gripper’s object face. The needles must operate pair wise in opposite directions. The needles extend 1–2 mm when

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 155

Fig. 7.5 Short needle gripper.

gripping, and retract completely behind the gripper’s face when releasing the object. All needles are operated simultaneously by the mechanical action of a single pneumatic cylinder inside the gripper. There are no reports of this type of gripper being used for gripping meat or any other foodstuff. But one could consider it as a candidate for gripping thin slices of sausage, cured meat or similar. In particular, if these slices are porous it could be a viable alternative.

7.5.2 Deep penetrating needle gripper Deep penetrating needle grippers were developed for the specific purpose of handling meat. Meat from both land animals and fish has been targeted. In particular handling of fish fillets has been investigated. The function of the gripper is to use at least two pairs of needles that are slanting towards each other, typically at an angle of about 45° to the gripper face. When fully extended these needles almost meet at a point 3–5 cm in front of the gripper surface. A typical diameter for the needles is 2 mm. Each needle is operated individually by a small pneumatic cylinder. Figure 7.6 shows a large needle gripper; Fig. 7.7 shows an example of a modular needle gripper where each module has two pairs of needles (Gjerstad et al., 2006). The individual actuation of the needles makes the gripper less sensitive to bones in the object. Any needles that meet bone inside the meat piece gripped will stop while the others continue to extend. Thus maximum possible grip force is obtained. Due to the deep penetration, these grippers will give a very strong holding force. The grip will be lost only if the local strain on the tissue exceeds its rupture

© Woodhead Publishing Limited, 2013

156 Robotics and automation in the food industry

Fig. 7.6 Long needle gripper. (Source: SINTEF.)

Fig. 7.7 Modular long needle gripper. (Source: Gjerstad et al., 2006.)

strength. For land animal meat and most fish meat this is no problem. But according to reports some fish species have a very loose meat texture in the spring season, making the use of this gripper principle less viable for these cases.

7.5.3 Lifting performance of needle grippers The main performance parameter for a penetrating gripper is its ability to hold the object during transfer. This is related to the rupture strength of the material

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 157 handled. It is very difficult to create reliable mathematical models for the holding ability since the rupture mechanism of meat in general is not well documented. Since meat is an inhomogeneous organic material, the well-known engineering models from inorganic materials cannot be applied. Some simplified models assuming homogeneous structure around the needles have not been successful in explaining observed holding strength. But some experimental data from the handling of fish fillets indicate safe regions for handling acceleration and speed. Experiments on handling of cod and salmon fillet show little difference between these species, and handling with maximum acceleration up to 2.5 m/s2 seems to create no observable damage to the meat tissues (Gjerstad et al., 2006).

7.5.4 Hygienic performance of needle grippers Hygienic requirements are a challenge for the short needle grippers. The holes for all the needles are places where residues from the handled materials will collect. A frequent purging of all needle holes with some antibacterial fluid seems to be necessary to obtain safe hygienic operation. For long needle grippers the hygienic challenges are met by building a special flushing system into the gripper. This system allows for flushing of the needles as they are retracted after each gripping cycle. Flushing can either be by disinfected water, or by some antibacterial fluid. Hygienic requirements make it mandatory to use only stainless steel needles on all types of needle grippers. The housing should be either stainless steel or polymer. Concave areas should be avoided on all external surfaces to avoid collection of residues from the handled materials.

7.6

Suction grippers

Suction grippers have the advantage of being able to grip on one surface only. In addition the physical size of the gripper at the robot’s arm end can be made relatively small. But adding vacuum generating devices and tubing to the gripper does increase the volume and the complexity of these grippers. This is only a moderate complexity increase in comparison to pneumatically operated pinching grippers or needle grippers. The overall assessment is then that suction grippers can give the most compact total gripper solution. Suction grippers are well established in the handling of solid inorganic materials. They are reliable, non-intrusive and work well in a confined space. There are also several examples of suction grippers used to handle chocolate, cookies and other small food objects. In particular packaged goods are easily handled. The main drawback of standard suction grippers is their sensitivity to air leakage into the suction cups. If the flow of air from this leakage is higher than the flow capacity of the vacuum pump or ejector, the suction underpressure will disappear and the grip is lost. This is a well-known phenomenon and limits the usage of traditional suction grippers to non-porous materials.

© Woodhead Publishing Limited, 2013

158 Robotics and automation in the food industry Newer development has partially solved this problem. Grippers based on moderate underpressure but high airflows have the ability to create sufficient gripping force for handling textiles and similar porous materials. These grippers normally use the Coanda effect ejector as a vacuum generator since this ejector will handle large airflows (Lien and Davis, 2008). Tests of gripper solutions using standard suction cups and vacuum generators for gripping fish fillet pieces have given disappointing results. This is due to the leakage because of the uneven surfaces of the fish fillets. Tests with Coanda grippers on the other hand are more promising. The leakage seems to be within the range that these grippers will handle. Another variation of the suction principle is to use the underpressure created by high-speed fluid flow. In this gripper an airflow directed towards the object to grip is deflected by the object to flow in a small gap between the object and the front face of the gripper. According to the Bernoulli equation this high-speed flow creates a decrease in the local pressure in this gap. The pressure difference to the ambient pressure can be substantial and will create a good grip force perpendicular to the gripper’s front surface. But since there has to be a continuous airflow in the gap, the coefficient of friction between the object and the gripper is virtually zero here. The gripper cannot exert any force parallel to the gripper’s front surface. This is a drawback since additional contact points have to be established to ensure proper control of the gripped object’s position and orientation.

7.6.1 Strong underpressure grippers Figure 7.8a shows the suction cup end of a conventional gripper system. Figure 7.8b shows a typical circuit diagram of the same system. In most cases several suction cups are connected to one vacuum device. But if there is high risk of air leakage into one or more of the suction cups, separate ejectors for each suction cup can be applied. Another option is to use flow-sensitive cut off valves. These valves sense the flow in the line and close if the flow becomes too high. A valve in each vacuum line will then isolate a leaking suction cup from the rest of the system. The gripping force is calculated by Equation [7.6]: [7.6]

FS = nA nAS p

where FS is the force from the suction gripper, n is the number of suction cups, AS is the area of one suction cup, Δp is the pressure difference (underpressure) in the suction cup.

7.6.2 Moderate underpressure, high-flow grippers In the Coanda effect gripper a Coanda ejector is used to create a moderate underpressure. This ejector will convey a large volume of air at this moderate pressure difference. The ejector can be combined with the suction cup so that there is no

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 159 (a)

(b)

Air supply control valve

Control signal Vacuum generator

Suction cups

Fig. 7.8

(a) A single suction cup, (b) diagram of a high underpressure suction gripper control system.

need for more than a supply line and a control valve to control the airflow to the gripper (Lien and Davis, 2008). The underpressure obtained with a Coanda effect gripper is one order of magnitude less than what can be obtained in standard suction grippers. But since it accepts a leakage flow and can operate on a large area suction cup it will still be able to lift objects handled in the normal food production. Figure 7.9a shows a planar Coanda effect gripper holding a piece of leather, Fig. 7.9 shows the same gripper holding a piece of textile. Note that the unused suction cups do not influence the lifting action of the active suction cups. The lifting capacity of a Coanda gripper can be estimated from the suction curve of the specific device. The suction curve is dependent on the ejector geometry and size and has to be determined experimentally. Figure 7.10 shows this curve for a planar device similar to the one shown in Fig. 7.9, but with a thickness four times as large (Lien and Davis, 2008).

7.6.3 Bernoulli grippers The Bernoulli gripper has the principal shape shown in Fig. 7.11. This is a noncontact gripper which could have important hygiene advantages. The gripping

© Woodhead Publishing Limited, 2013

160 Robotics and automation in the food industry (a)

(b)

Ejector under pressure (kPa)

Fig. 7.9 (a) A Coanda effect gripper holding a thin leather sheet, (b) a Coanda effect gripper with four suction heads holding a piece of textile, two of the suction cups are uncovered.

2.5

Single ply cotton fabric

2

Thin leather

1.5 1 0.5 0 0

20

40 60 80 Supply pressure (kPa)

100

Fig. 7.10 Obtainable underpressure for a Coanda gripper (9 × 20 mm throat size) for a non-porous and a porous material. (Source: Lien and Davis, 2008.)

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 161

Air flow

Plate

h

Nozzle

r int Gripped object

r ext

Fig. 7.11 The working principle of a Bernoulli gripper.

effect is dependent on the actual local air velocity in the gap between the gripper surface and the gripped object. The gripping force is given by integrating the differential radial area multiplied by the underpressure obtained from the simplified Bernoulli Equation [7.7] (Dini et al., 2009). The simplification is to assume that there is no change in air density and temperature as it flows through the gap.

Fg =

1 Q 2 ⎡ rext 1 ⎛ rex2 t rint2 ⎞ ⎤ ρ − ⎢ ln ⎥ 2 2 πh ⎣ rint 2 ⎜⎝ rint2 ⎟⎠ ⎦

[7.7]

The air flowing from the central supply channel does also create a repulsive force as it is deflected in radial direction by the gripped object. Equation [7.8] gives that force.

Fr

r

Q2 πrint2

[7.8]

where Fg is gripping force, Fr is repulsive force, ρ is ambient air density, Q is airflow, h is height of the air gap, rext is external radius of the Bernoulli gripper, and rint is internal radius of the Bernoulli gripper. With a large air gap the air flow of the gripper is determined mainly by the flow resistance of the supply channel. The air flow is then constant with a constant

© Woodhead Publishing Limited, 2013

162 Robotics and automation in the food industry

Gripping force Fg (N)

50 40 30 20 10 0

–10

0

1

4 2 3 Gripper gap h (mm)

5

6

Fig. 7.12 The gripping force of a Bernoulli gripper as a function of the gripper gap. (Source: Dini et al., 2009.)

supply pressure, and the gripping force increases as the gap height is decreased. At a certain small gap height the flow resistance of the air gap will dominate, and the air flow will start to decrease, reducing the gripping force. When the gap height approaches zero, the repulsive force will dominate. Therefore the gap height never becomes zero. There is thus no direct contact between the gripper and the object. This gripping principle has been demonstrated for handling fragile food like sliced cucumbers and tomatoes (Davis et al., 2006). Figure 7.12 shows a graph of typical values the sum of gripping and repulsive forces (Dini et al., 2009).

7.6.4 Lifting performance of suction grippers The lifting force depends on the obtainable underpressure. There is a marked difference between the standard suction cup systems and the Coanda ejector grippers. The Coanda ejector may not provide enough underpressure to lift heavy objects. The Bernoulli gripper comes closer to the standard suction cup with respect to lifting capacity. The Coanda effect gripper does on the other hand work with porous and uneven surfaces for light objects due to its capacity to maintain underpressure even with a considerable air leakage through the object or along the periphery of the suction cup. Typical figures for obtainable underpressure are given in Table 7.1.

7.6.5 Hygienic performance of suction grippers Hygienic issues are very relevant in most cases of handling bare foodstuff. The gripper should not introduce any risk of contamination of the product. Both materials and the gripper design play an important role. The material used in areas that come in contact with the product should be non-toxic and hygienically approved. Stainless steel and many plastic materials are acceptable.

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 163 Table 7.1 Underpressure range and lifting capacity for standard, Coanda and Bernoulli suction grippers Gripper type

Typical underpressure range (N/m2)

Lifting force with ø90 mm suction area (N)

High underpressure standard ejector Low underpressure Coanda ejector Bernoulli gripper

5000–80 000 500–5000 1000–8000

30–500 3–30 6–50

Table 7.2 Hygienic characteristics of suction grippers Gripper type

Collection of organic residues

High Will collect organic underpressure residues in suction standard device and on the ejector suction cup surface Low Will collect organic underpressure residues on suction Coanda ejector cup surface but not in ejector Bernoulli gripper No collection of residues

Risk of bacterial growth

Cleaning

Large

Difficult to clean the inside of the ejector or vacuum pump. Surface cleaning by washing Self cleaning ejector, surface cleaning by washing

Moderate

None

Not required except for general equipment cleaning

The most common source of contamination is residues left on the gripper from the handling of objects. These residues will lead to the growth of bacteria or fungi if they are not removed by regular cleaning. In particular, handling of bare meat will leave residues. Juices from the meat and small particles can easily aggregate in holes, grooves and the interior of the gripper. Suction grippers are vulnerable in this respect since they rely on air flowing into the gripper. Proper cleaning measures are therefore needed. Table 7.2 shows a summary of the hygienic characteristics of the three types of suction grippers.

7.7

Surface effect (freeze) grippers

One very interesting idea for hygienic gripping is to use the effect of almost instantaneous freezing that occurs when a cold body is brought into contact with a wet object. This freezing forms a surface bond by the ice formed on the interface area between the cold body and the wet object. The cold body has to have a temperature typically around −10°C to cause this quick freezing action. The principle is in industrial use for handling small electronic component were a drop of water is added to provide moisture. A gripper for large objects was patented in 1988 (Guse

© Woodhead Publishing Limited, 2013

164 Robotics and automation in the food industry et al., 1988). Although grippers of this type has been developed and demonstrated successfully in research laboratories there are no reports of these being used yet in the food industry. The main advantage of these grippers is the fact that they act single sided and can be designed to very high hygienic standards. But they also have a marked drawback in a somewhat unpredictable releasing action.

7.7.1 Design of freeze grippers The general principle of freeze grippers comes from the fact that thermal energy flow from the surface of an object into a contacting colder surface will cool the object surface. If the object surface is wet and the gripper surface has a sufficiently low temperature there will be an immediate formation of ice in the contact area. This ice will bind the object to the gripper. The seed of ice formation depends on the temperature of the gripper surface, the thermal capacity of both the gripper and the object and the wetting capacity of the gripper surface. The amount of moisture on the surface of fish fillets or fillets of land animals is usually enough. But pieces that have been left alone in dry areas for some time may not be moist enough for a safe grip. The freezing process is very difficult to model. Several important parameters are usually not well known at the specific gripping instant, such as the amount of water on the surface to grip, the heat transfer coefficient and the wetting property of the gripper surface. Even a good model will not give correct answers if these parameters are not known with high accuracy. Instead these types of grippers have been studied experimentally to obtain some information about the functional capability of this gripper family. One of the most important function parameters is the time to form a sufficiently strong frozen film between the object and the gripper. This parameter is independent of the release methods used, but it is dependent on the type of object to be gripped. Experiments with pork meat, beef, codfish and salmon have given the empirical data in Table 7.3.

7.7.2 Freeze grippers with mechanical release Freeze grippers with mechanical release have a set of smaller metallic gripping areas called freeze pads. These freeze pads are connected to a low-temperature

Table 7.3 Gripping attachment times and temperatures for a Peltier element freeze gripper with stainless steel gripping surface Contact time (s)

Grip at temperatures below (°C) No grip at temperatures above (°C)

0.2 0.4 1

−10.5 −7 −2

−8 −5 −0.5

Source: Lien and Gjerstad, 2008.

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 165

Fig. 7.13 Freeze gripper with mechanical release.

reservoir that keeps it cold as long as necessary for a good grip. Fluid heat transfer mediums are usually applied, brine or types of fluid that stay unfrozen at temperatures below −10°C are used as a heat transfer medium between the gripper’s freeze surfaces and a refrigeration unit. Figure 7.13 shows a freeze gripper with mechanical release developed by SINTEF. The freeze pads are circular areas of stainless steel that are embedded in a plastic body. They are all connected to the cooling fluid in an inner chamber of the gripper. The release mechanism consists of a stainless steel frame with a knife section that moves across the surface of the freeze pads when the frame is rotated by its mounting shaft. The shaft is rotated by a small pneumatic actuator inside the gripper body. To obtain a gripping action the gripper’s freeze pads are pressed against the object. A film of ice is then formed immediately between the object and the freeze pads. This ice creates the holding force of the gripper. Release is obtained by moving the release knife across the freeze pads between the surface of the pads and the object. It will then scrape the ice film off the freeze pads, thus releasing the object from the gripper.

7.7.3 Reversible heat flow grippers The reversible heat flow gripper is built around a Peltier element which generates low temperatures on the gripper surface. The Peltier element is a semiconductor device that will generate a heat flow when an electric current passes through it. The direction of heat flow is determined by the direction of the electric current.

© Woodhead Publishing Limited, 2013

166 Robotics and automation in the food industry

Robot interface plate

Housing

Cooling fins

Water or briner

Water or brine

Q1

Insulation

Heat transfer back plate

Q0

Peltier element

Stainless steel gripper surface

Gripped meat

Fig. 7.14 The principle of the Peltier element freeze gripper. Q0 is the heat flow from the object; Q1 is the heat flow into the cooling fluid.

Fig. 7.15 A Peltier element freeze gripper holding a piece of salmon fillet.

Changing the electric current direction will also change the direction of the heat flow. This property is very useful in a freeze gripper. Freeze gripping is obtained by sending current one way through the Peltier element. Reversing the current will release the object by melting the gripping ice (Guse et al., 1988). Figure 7.14 shows a cross section of a reversible heat flow gripper. It needs an inner chamber with a fluid that serves as a stable temperature reference for the Peltier element. Cold water is a useful fluid. It must be pumped through the gripper continuously to remove heat from the warm side of the Peltier element when the gripper is operated (Lien and Gjerstad, 2008). Figure 7.15 shows a reversible thermal flow gripper holding a piece of salmon fillet. The gripper has the advantage of a very clean gripping surface and no mechanical moving parts. An alternative solution for release by reversed heat flow is to inject a small amount of steam into the frozen area. In this gripper solution a central hole is used

© Woodhead Publishing Limited, 2013

Gripper technologies for food industry robots 167 to lead the steam into the frozen region. A slight overpressure of the steam will force it out in the gap between the gripper surface and the gripped object as the ice melts away. The freeze action of the Peltier element must be turned off during the release (Seliger et al., 2000). The holding force of a freeze gripper is sufficient to hold pieces of meat of fish or land animals. Experiments have shown that the holding force for fish fillet is larger than the rupture force of the fish meat tissue. Experiments with gripping of textiles showed that considerable gripping force could be obtained even with only partial ice formation on the gripper face. In this experiment ice covered 10–30% of the gripper–textile contact area. The measured rupture force for the holding ice was in the range 9–15 N/cm2. The major problem with this type of gripper is the release time. While a typical release time for a pinching gripper is in the order of 0.1–0.2 s, the Peltier element gripper with current reversal has a release time in the order of 0.5–1.0 s. The time delay challenge can be overcome though by starting the release heat flow reversal a fraction of a second before arrival at the delivery place (Lien and Gjerstad, 2008).

7.7.4 Hygienic performance of freeze grippers The freeze grippers with mechanical release face the same hygienic challenges as the pinching grippers. There are guide ways and actuating mechanisms that have to be designed so that residue cannot collect. In particular, in the narrow gap between the knife and the freeze surface there will be a great risk of residue collection. This area has to be flushed with a cleaning fluid frequently to avoid bacteria and fungus growth. The reversible heat flow grippers on the other hand have a much better potential for extremely hygienic operation. The gripping surface is a smooth stainless steel surface. There are no moving parts, and the housing can be made completely flat or convex to make exterior cleaning very easy. The grippers with steam release will expel a small amount of steam through holes in the grip surface. Since steam is a very good cleaning fluid, it is self cleaning. The Peltier element gripper has no holes in the housing facing the gripped object, no moving part and no fluid flow in the gripper action. From a hygienic viewpoint it is the perfect gripper.

7.8

Selection of the appropriate gripping technology

In any automation task the requirements for the handling of the object has to be mapped carefully to select the optimal gripper solution. In soft handling this becomes particularly important. Soft handling of food products does pose challenges with respect to mechanical handling and to hygienic operation. These challenges lead to conflicting requirement specifications. Some requirements will favour one gripping technology while others will rule out the same technology.

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

−

+/0

+/0

+

+

+: Favorable according to this parameter 0: Neutral for this parameter –: Unfavorable or unsuitable according to this parameter.

+

+

+

+

−

+

+

+

+

+

+

−

+

+

+

+

+

+

+

−

+

+

+

+

+

+

+

−

+

+ +

+

+

−

+

+

+ −

+

+

+

0

−

+ −

+

+

+/0

−

0/−

0/− −

0/−

+

+

−

+

+ +

0/−

+/0

+/0

−

− +

+

+

+

+

− +

+

−

+

+ −

+

+

+

+ +

−

+

+ +

+

−

+ −

−

+ −

0

+ −

0/−

+/0

+

Two/three finger Enclosing Short needles Long needles Standard suction Coanda suction Bernoulli suction Freeze

0/−

Hard Uneven Irregular Soft Porous Fragile Precise Single Dry Moist Heavy Grip marks Hygienic surface surface shape structure positioning side grip surface surface object performance

Gripper type

Table 7.4 Qualitative performance parameters for different gripper technologies

Gripper technologies for food industry robots 169 A requirement list will then have to put weights on the different requirements in order to be able to satisfy the most important ones. A qualitative summary of the suitability with respect to different factors of the gripper technologies mentioned in this chapter is given in Table 7.4. The table has been composed with food handling challenges in mind. The classification is crude. A + indicates that the technology works well. The sign 0 indicates neither bad nor good, while − indicates unsuitability. For some factors, such as one sided grip and precise positioning, the suitability is a go/no go issue. But for gripping on uneven surfaces the suitability may not be so clear. There is definitely a variation in unevenness. If the surface is very uneven, Coanda and freeze grippers will become unsuitable, while enclosing and long needle are suitable no matter what degree of unevenness the task presents.

7.9

Future trends: from laboratory to industry

The grippers used in robotic handling in the food industry today (2010) are mainly the traditional pinching, enclosing or suction types. These technologies work well in applications that are similar to the ones seen in other industries. But in handling tasks that are unique for food industries it seems to be more difficult to find good gripper solutions. For that reason there has been an increased effort in research laboratories to investigate new alternative gripping technologies. All the non-traditional gripping methods described in this book have been developed based on requirements related to food handling. These new types show promising performances that can solve some of the problems that arise in robotic applications in the food industry today. One obvious problem is the hygienic challenge. The traditional gripper solutions used in industry are not the best from a hygienic perspective. The Peltier element gripper is an example of a solution to this challenge. More important though is the capability of handling soft material in single sided grips. This is a requirement that appears quite often in discussions over automation in the processing of meat, be it fish or land animals. The softness is often paired with uneven surfaces and contours. For this class of handling problems the Peltier element gripper, the Coanda gripper and the needle grippers can provide good solutions. But performance in laboratory tests is not enough. At this stage some users must be willing to experiment with the new technologies to determine their real value in the harsh industrial environments. This step is necessary to refine the new technologies and increase insight in gripping technology for food materials.

7.10 References Chua P Y, Ilschner T and Caldwell D G (2003), Robotic manipulation of food products – an overview, Industrial Robot, 30, 345–354.

© Woodhead Publishing Limited, 2013

170 Robotics and automation in the food industry Davis S, Gray J O and Caldwell D G (2006), An end-effector based on the Bernoulli principle for handling sliced food and vegetables. Robotic and computer integrated Manufacturing, 24(2), 249–257. Dini G, Fantoni G and Failli F (2009), Grasping leather plies by Bernoulli grippers, CIRP Annals – Manufacturing Technology, 58(1), 21–24. Gjerstad T B, Lien T K and Buljo J O (2006), Handle of non-rigid products using a compact needle gripper, Proceedings of the 39th CIRP International Seminar on Manufacturing Systems, Ljubljana, 145–151. Guse R, Koch W and Schulz G (1988), Festgefriergreifer und Verfahren zu seinem Betrieben, German patent application DE 3701874 A1, 4 August 1988. Lien T K and Davis P G G (2008), A novel gripper for limp materials based on lateral Coanda ejectors, CIRP Annals – Manufacturing Technology, 57(1), 33–36. Lien T K and Gjerstad T B (2008), A new reversible thermal flow gripper for non-rigid products, Transactions of the North American Manufacturing Research Institution of SME, 36, 565–572. Mouri T, Kawasaki H, Yoshikawa K, Takai J and Ito S (2002), Anthropomorphic Robot Hand: Gifu Hand III, Proceedings of ICCAS2002, Jeonbuk, Korea, 1288–1293. Seliger G, Stephan J and Lange S (2000), Hydroadhesive gripping by using the Peltier effect, Proceedings of the ASME ‘Manufacturing Engineering Division’, 11, 3–8. Thayer N and Priya S (2011), Design and implementation of a dexterous anthropomorphic robotic typing (DART) hand, Smart Materials and Structures, 20(3), 12pp. Wolf A, Steinmann R and Schunk H (2005), Grippers in Motion, Berlin, Springer.

© Woodhead Publishing Limited, 2013

8 Wireless sensor networks (WSNs) in the agricultural and food industries N. Wang, Oklahoma State University, USA and Z. Li, South China Agricultural University, P. R. China DOI: 10.1533/9780857095763.1.171 Abstract: This chapter begins by introducing wireless sensor network (WSN) technology and its current development. It then extensively reviews the applications of WSNs in the areas of environment monitoring, precision agriculture, precision livestock farming, greenhouse management, and food traceability systems. The chapter also discusses the obstacles and challenges faced in the development and deployment of WSNs in agriculture and food production. Finally, the future trends in application of WSN technology in agriculture and food production are presented. Key words: wireless sensor network, environment monitoring, animal behavior, radiofrequency identification (RFID), greenhouse.

8.1

Introduction

Wireless technologies have been under rapid development during the past 10 years. Types of wireless technologies being developed range from simple short-range, point-to-point data transmission through infrared light (Infrared Data Association, IrDA), to wireless personal area networks (WPANs) for short-range, point-to-multipoint communications, for example, Bluetooth and ZigBee, to mid-range, multi-hop wireless local area network (WLANs), and to long-distance cellular phone systems, such as GSM/GPRS and CDMA (Wang et al., 2006). Recent scientific and technological revolutions have lead to maturity in many industrial areas such as radio frequency (RF) technology, integrated circuits, microprocessors, smart sensors, and micro-electro-mechanical systems (MEMS) (Mahfuz and Ahmed, 2005). These latest developments have made it possible for mass-production of low cost, low power consumption, high reliability, multifunctional, intelligent, miniature sensor and controller nodes with networking capability. These devices are called ‘nodes’ or ‘smart dust’. The networked nodes are commonly deployed to application environments to collect real-time data

© Woodhead Publishing Limited, 2013

172 Robotics and automation in the food industry and implement simple control strategies. Through them, more detailed knowledge about the environment is acquired, especially in dangerous, hazardous, and remote areas and locations. Agricultural and food production are normally in remote and uncontrolled areas. The production outcomes depend heavily on environment and crop conditions. With the emergence and convergence of several technologies, including the global positioning system (GPS), geographic information systems (GIS), miniaturized computer components, automatic control, in-field and remote sensing, mobile computing, advanced information processing, and telecommunications, detailed information on field and production variability both spatially and temporally can be acquired and treatments can be adjusted to meet each site’s unique needs (Zhang et al., 2002). Various sensors, controllers, and actuators with intelligence have been extensively used in field data acquisition, monitoring, and control. Their major shortcomings are the requirements for extensive wiring for multiple-point measurements and controls and frequent on-site visits by operators for data downloading and maintenance due to unavoidable damages and loss. Failure to satisfy these requirements may lead to missing data and malfunction of the overall production system. A WSN system consists of multiple nodes, each of which comprises RF transceivers (motes), sensors, microcontrollers, and power sources. Their selforganizing, self-configuring, self-diagnosing and self-healing capabilities enable them to form ad hoc single/multi-hop networks. This reduces and simplifies wiring and connectors, provides installation flexibility for sensors and controllers, and reduces maintenance complexity and costs. Most wireless nodes have signal conditioning and processing units that can convert analog to digital signals. As a result, noise pick-up becomes a less significant problem. Furthermore, advanced microelectronic technology allows the nodes to have very small physical sizes and low power consumption while maintaining their functionality. In agricultural and food production, data collected from a deployed WSN can be used to optimize management strategy in a real time fashion. A remote user (e.g. farm manager) can also send commands to selected nodes in a field to assign new tasks, change configurations, and diagnose problems. With these advantages, WSN technology can play a very important role in realizing uninterrupted data acquisition, processing, and controls with fine spatial and temporal resolution.

8.2

Current state of development of WSNs

The development of WSNs has been through its emergence, slow development, and rapid adoption stages. The fast development of computer network technology and expansion of wireless communication networks bring WSNs to a new era of rapid development and deployment. Development of WSNs is mainly based on three major components: hardware, operating system (OS), and network communication. The hardware consists of:

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 173 (1) multiple nodes distributed in a target site to form a wireless communication network, each of which has one or more sensors and/or controllers, on-board signal conditioning and processing units, limited data storage, a radio transceiver, and a power unit; and (2) one or multiple sink or gateway bridging units to relay measurement data to a remote control center through wireless LAN, cellular, or other networks (Boegena et al., 2006). The OSs are tied to the hardware and run on the nodes to coordinate on-board components to complete the assigned tasks such as data acquisition, processing, transmission, and storage (Gay et al., 2003). The network communication defines network topology and communication protocols with the consideration of routing, power and resource management, and node localization, etc. Table 8.1 shows the most common commercial wireless sensing and controlling systems. They are categorized into two classes: ‘WSN hardware platforms’ and ‘Data logger platforms’. Radio frequency identification (RFID) has been widely adopted in many applications. It plays very important roles in food traceability systems. When integrated with WSN technology, RFID systems can provide more product information in addition to identification. 8.2.1 WSN hardware platforms Many general purpose commercial platforms are currently available in the WSN product market. The platforms are similar in major components, such as nodes with/without on-board sensors, data acquisition modules, sinks/gateways modules, and power unit. They are often open-source and usually have very large user groups. Technical support is provided through resources/forums shared among system developers. The commercial WSN platforms are composed of wireless transceiver modules, data acquisition interface modules, common sensor modules, and gateway modules. Users can stack the modules together to form a node based on the requirement of their applications. A wireless transceiver module consists of a low-power microprocessor and an RF transceiver. The DAQ modules provide interfaces to external sensors and controllers, such as analog I/O, general purpose digital I/O, I2C interface, and UART ports. The sensor modules include common sensors, such as temperature, humidity, light, acceleration, and sound on one circuit board with a standard connection interface to the other modules. The gateway/base station modules are used as a sink to allow the aggregation of sensor data from a WSN to a remote PC through various means such as Ethernet, cellular network, high-power radio, and satellite. The pioneer manufacturer in this category is Crossbow Technology, CA, USA who provides wireless transceivers (Mica2, MicaZ, IRIS and Imote2), data acquisition boards (MTS, MDA, ITS and IMB400), gateways (Star-gate and MIB), and development software (XMesh and XServer, Crossbow Technology Inc., California, USA). This product line provides WSN developers and users with flexibility on design application-oriented WSNs. Jennic Technology (Jennic, 2010) in the United Kingdom provides a similar family of wireless microcontrollers, modules, and development tools that have been

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

CPU

Atmega128

Intel PXA271

MSP430

MSP430

Imote2

TinyNode584

TinyNode184

Atmega128

IRIS

MicaZ

WSN hardware platforms Mica2 Atmega128

Platform

3.6V battery 22 µA, Sleep; 2.2A, Active

3.6V battery 4µA, Sleep; 77mA, Active

3.3V battery, 390µA, Sleep; 66 mA, Active

3.3V battery, 15µA, Sleep; 8mA, Active

3.3V battery, 15µA, Sleep; 8mA, Active

3.3V battery, 15µA, Sleep; 8mA, Active

Power

10 kB RAM 512 kB Flash

10 kB RAM 512 kB Flash

4 kB RAM 128 kB ROM 512 kB Flash 4 kB EEPROM 256 kB SRAM 32 MB SDRAM 32 MB Flash

128 kB ROM 512 kB Flash 4 kB EEPROM

128 kB ROM 512 kB Flash 4 kB EEPROM

Memory Regular I/Os*, 51 pin interface to other extension boards Regular I/Os*, 51 pin interface to other extension boards Regular I/Os*, 51 pin interface to other extension boards GPIO/SPI/UART/ I2S/USB/ AC’97/Camera/ IMB400 multimedia extension board Regular I/Os, factory made extension board for custom interface electronics Regular I/Os, factory made extension board for custom interface electronics

I/O interfaces

Table 8.1 Commercial WSN hardware platforms and data logger platforms

SX1211 (868/915 MHz)

OS

30 m with attached antenna

75–100 m outdoor 20–30 m indoor >300 m outdoor >50 m indoor

150 m outdoor TinyOS 50 m indoor

TinyOS

Embedded Linux or Windows support

TinyOS

TinyOS

300 m outdoor TinyOS 1.x/2.x

Max. range

XE1205 (915 MHz) Up to 2000 m

CC2420

RF230

CC2420

CC1000

Radio

Shockfish SA

Shockfish SA

Crossbow Technology

Crossbow Technology

Crossbow Technology

Crossbow Technology

Manufacturer

© Woodhead Publishing Limited, 2013

V/SG/G-Link N/A nodes+ 2.4 GHz base station

Wireless data logger platforms CR1000+RF4xx Renesas H8S 2322 for CR1000

eyesIFXv2.1

Atmega128L

BTnode rev3

3.6V battery 5.1 µA, Sleep; 23mA, Active

10 kB RAM 48 kB Flash

0.5 mA sleep 25 mA max

Up to 2 MB (approximately 1 000 000 data points)

12 V battery 4 MB SRAM Sleep, 0.6 mA CF card extension Active, 27.6 mA/100 mA for CR1000; 9-16 V battery 1 mA Stand-by 75 mA max for RF4xx

Sleep 9.9 mW 64+128 kB SRAM Active with 128 kB Flash Bluetooth, 198 4kB EEPROM mW or 102.3 max for CC1000 MSP430F1611 36 mW max 10 kB RAM 48 kB ROM 4 MB Flash

TI MSP430

Tmote Sky

Fixed on-board sensors depending on series ‘x’

16 Analog inputs, 3 analog outputs, 8 digital I/Os, communication ports

Regular I/Os, no manufacture built extension board

Regular I/Os, on-board sensors, no manufacture built extension board Regular I/Os, no manufacture built extension board

125 m outdoor TinyOS 50 m indoor

2.4 GHz, 16 nodes maximum for simultaneous streaming

RF401 (915 MHz) RF411 (922 MHz) RF416 (2.4 GHz)

Infineon Technologies

ETH Zurich

Moteiv

(Continued)

Up to 1.5 km LoggerNet 3.x, Campbell using short PC400 1.2 or Scientific distance ShortCut 2.2 antennas Longer distance using high gain antennas 70 m line of Precompiled Micro Strain sight, up to system in 300 m with VB/VC++ / optional LabView high gain antenna

ZV4002(Bluetooth) 10 m using TinyOS or or CC1000 Bluetooth BTnut 300 m outdoor using CC1000 Infineon TDA5250 200 m–600 m TinyOS (868 MHz) depending on data rates

CC2420

© Woodhead Publishing Limited, 2013

N/A

Em50R/G

AC powered or 3 AAA battery

5 AA batteries

Power

N/A

1 MB data storage

Memory

Radio

50R: 900 MHz or 2.4 GHz selectable 50G: using cellular network TMB: Temperature 2.4 GHZ SMC: Soil moisture

4 Analog inputs, 1 serial input port

I/O interfaces

*Regular I/Os include GPIO, I2C, SPI, ADC, UART. Available type and amount may vary between nodes.

HOBO W-TMB/ N/A SMC

CPU

Platform

Table 8.1 Continued

420 m with clear line of sight or 300 m typical

N/A

Max. range

Manufacturer

HOBO Node viewer

Onset

ECH2O Utility Decagon Devices and DataTrac

OS

WSNs in the agricultural and food industries 177 used in many applications (Zhou et al., 2007). Shockfish (2010) in Switzerland produces the TinyNodes series mainly for 868 MHz band applications. The extension boards provide interfaces to add external power supply, sensors, and controls. Sentilla Corporation (Sentilla, 2010) in the USA provides Tmote Sky and Tmote mini, which are ‘one-board’ module integrating sensors, a microprocessor, and a RF transceiver. They are cheap in production cost, and robust due to the integrated design, but with very low application flexibility. Ember Corporation in the USA (Ember, 2010) offers a low-power, high-performance integrated ZigBee System-on-Chip with a size of only 7 × 7 mm. Various international standards have been established for WSN applications in recent years. Among them, the standards for wireless LAN, IEEE 802.11b (‘WiFi’) (IEEE, 1999) and wireless PAN, IEEE 802.15.1 (Bluetooth) (IEEE, 2002) and IEEE 802.15.4 (ZigBee) (IEEE, 2003), are used more widely for measurement and automation applications. All these standards use the instrumentation, scientific and medical (ISM) radio bands, including the sub-GHz bands of 902–928 MHz (US), 868–870 MHz (Europe), 433.05–434.79 MHz (US and Europe) and 314–316 MHz (Japan) and the GHz bands of 2.400–2.4835 GHz (worldwide acceptable). In general, a lower frequency allows a longer transmission range and a stronger capability to penetrate through walls and glass. However, due to the fact that radio waves with lower frequencies are more easily absorbed by various materials, such as water and trees, and that radio waves with higher frequencies are easier to scatter, the effective transmission distance for signals carried by a high frequency radio wave may not necessarily be shorter than that of a lower frequency carrier of the same power rating. The 2.4 GHz band has a wider bandwidth (250 kbs) that allows more channels and frequency hopping and permits compact antennas. The ZigBee standard was established by the ZigBee Alliance that is supported by more than 70 member companies. It adds network, security and application software to the IEEE 802.15.4 standard. Owing to its low power consumption and simple networking configuration, ZigBee is considered the most promising for wireless sensors and has been embedded into many commercial products worldwide. One of the common characteristics of the WSN hardware platform is it needs expertise to program the nodes and overall system. System developers with a good background of electronic circuit design and microprocessor programming are required. Since most of the manufacturers of the WSN platforms do not provide convenient technical support, it becomes a big hurdle for application engineers in agricultural and food productions to develop and deploy the WSN systems.

8.2.2 Wireless data logger platforms In agricultural applications, stand-alone data loggers with various standard interfaces (analog I/O, digital I/O, serial ports, parallel ports, etc.) are widely used. They are easy to use, easy to program, and very rugged under various environmental conditions. For example, data loggers from Campbell Scientific Inc. (2010) have been widely used in weather stations to collect environmental data. Some

© Woodhead Publishing Limited, 2013

178 Robotics and automation in the food industry of the data loggers integrate RF modules to remotely transmit data to a control center. Recently, more and more data logger manufacturers are adding wireless communication modules to their existing data loggers. Campbell Scientific Inc. now provides communication modules of Ethernet, Spread spectrum RF, satellite, cellular (GPS and CDMA), etc. which can easily connect to their data loggers. Decagon Devices Inc in the USA (Decagon, 2010) recently provides the Em50G wireless cellular data logger and Em50R wireless radio data logger, which allow long-distance networking and data communication. MicroStrain (2010) has a production line of wireless nodes with various sensors, wireless base stations, and a wireless sensor data aggregator. Onset (2010) manufacture a series of Hobo data loggers. They have recently started to provide wireless data loggers for temperature and/or humidity for indoor and outdoor applications. These data logger platforms are tied closely with the OSs and development tools from their respective manufacturers. They can be easily networked among modules from the same manufacturer, but it can be difficult to connect with systems from different manufacturers. This reduces the flexibility on both hardware and software selections. Meanwhile, these systems are often expensive. With these shortcomings, they are not suitable for large-scale WSN applications.

8.2.3 Radio frequency identification (RFID) systems An RFID system consists of tags or transponders, readers/interrogators, and a computer with an information management system. The tags/transponders carry product IDs, which are read by the reader/interrogator. The computer system stores and manages the product IDs and other information. There are two types of RFID tags, active and passive. The active tags are powered by an on-board battery, while the passive tags derive their power from a reader or an interrogator through electrical magnetic field. Hence, active RFID tags are equipped with larger memory, more I/O ports, and on-board sensors, and a more powerful radio. They carry more information in addition to product ID. The more powerful radio in an active tag allows a longer transmission distance than a passive tag. Most passive tags operate at low frequencies between 125 and 148 kHz (RuizGarcia et al., 2009). These tags are cheap (around US$2 per tag) and have been adopted for animal tags since the 1980s. ISO 11784/11785/14223 are the standards to regulate the animal RFID applications. Recently more and more active tags operating at frequencies of 455 MHz, 2.45 GHz and 5.8 GHz have been reported. Their multifunctional features make them more similar to the nodes in WSNs and easier to be included in WSNs. Z2RFID system developed by ZigBeef (2010) consists of long range cattle RFID tags, which can transmit unique ID numbers and motion and temperature information a long distance (>60 m) using ZigBee standards and a handheld interrogator (PDA) to store the information. These tags have been used by ranchers and rodeos to track and count their animals. Another company, TekVet (2010), has been marketing a small active 418 MHz RFID tag, TekSensor, which can be fixed inside an animal’s ear canal to measure external ear canal temperature, which is an indicator of animal core body temperature.

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 179 A reader receives the IDs and temperature readings from the TekSensors every hour, 24 h a day. A software package, TekAccess, was developed to access and analyze the data in order to provide an early indication of abnormal changes in temperature, which may be associated with many common illnesses.

8.2.4 WSN operating systems All the WSN hardware platforms share features of limited on-board or on-chip resources, severe memory constraints, and limited power access. Hence, an OS running on them needs to be both very small in footprint and event driven. Originally developed by University of California-Berkeley, Intel Research, and Crossbow Technology, TinyOS is a free, open-source embedded OS for low-power WSNs. The development group has now grown to be an international consortium, the TinyOS Alliance, which makes TinyOS the de facto standard OS for WSNs. The latest released version, TinyOS 2.1.1 supports most commercial WSN hardware platforms. TinyOS (2009) is written in nesC, a specially designed C dialect, and supports many popular microcontrollers and radio chips. It provides a large library of software components related to those hardware elements and has only one stack and a single execution context. Hence, the TinyOS programs are very robust and efficient. TinyOS strongly supports low-power operations including multihop networking, network-wide sub-millisecond time synchronization, data collection to a designated root or gateway, reliable data dissemination to every node in a network, and installing new codes over the wireless network. The scheduling mechanism used in TinyOS is specifically designed to fit WSN platforms with limited system resources, low-power consumption, and high concurrency. The main objective of WSN applications is to collect, transmit and store a vast amount of data. The collected raw data is usually a mixture of useful information, unwanted information, and noise. To efficiently handle these data using limited resources, a query processing system running on TinyOS, TinyDB (2010), was developed to handle the data and extract useful information. It provides a simple, SQL-like interface to allow users to easily inquire, filter, aggregate, and route the data through power-efficient in-network processing algorithms without writing NesC code. Other small footprint and high efficiency OSs for WSNs were also reported. A finite state machine-based OS, namely SenOS, was proposed by Kim and Hong (2005). It can operate in an extremely resource-constrained environment (including power and hardware resources) while providing desired task concurrency and network re-configurability. A mote-class WSN OS, named SOS, was also proposed to achieve dynamic reprogramming. Dynamically loadable modules and a common kernel were nested to support dynamic addition, modification, and removal of network services (Han et al., 2005). Table 8.2 summarizes the OSs developed for WSN applications. Most OSs for WSNs provide barebones functionality and leave system development to application developers. This becomes a big hurdle for users with weak software development background. Many manufacturers are providing network

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

Event driven

Thread-based

Sensor OS (SOS) (Han et al., 2005)

MantisOS (Bhatti et al., 2005) SenOS (Kim and Hong, 2005). Nano-RK (Rajkumar et al., 1997)

Modular/component level

Application level

kOS (Britton et al., 2005) Hybrid (event and object based)

Instruction/variable level

Yes

Real-time

No

Yes

Not real-time

Not real-time

Yes

No

Yes

Power management

Not real-time

Not real-time

Not real-time

Levels of granularity Scheduling supported in reprogramming

Modular/component level Finite state machine based Instruction/variable level Reservation-based Instruction/variable level

Event driven

Execution model

TinyOS (2009)

Name

Table 8.2 Summary of operation systems developed for WSNs

Atmel ATMEGA128 with Chipcon CC2420 transceiver User defined

Telos, Mica2Dot, Mica2, TMote Sky, Eyes, MicaZ, iMote Cricket, imote2, Mica2, MicaZ, tmote, Protosb, emu Mica2, MicaZ, Telos, Mantis nymph Not specified

Supported platform

WSNs in the agricultural and food industries 181 (a)

(b) Source

Source Sink Sink

Obstacle

Fig. 8.1 (a) Single-hop and (b) multi-hop communication.

management tools to help users to setup, monitor, and manage a WSN. They are coded as firmware based on the operation systems and provide computer-based graphical user interfaces (GUI). XMesh from Crossbow Technology is a software library based on TinyOS which runs on motes with pre-build multi-hop, ad hoc, mesh networking protocols and limited sensor board access methods. XSniffer, also from Crossbow Technology, is a tool to monitor packet transmission, packet contents, link-quality and node health information of all motes in a WSN after deployment.

8.2.5 Network architecture The wireless standards also address the network issues for wireless sensors. A WSN can be single-hop or multiple-hop (Fig. 8.1). The single-hop networks allow direct communication between a source (wireless sensor) and a sink (gateway). Due to the power limitation of a radio, the feasible distance of transmission may not be sufficient to establish a reliable communication between a source and a sink. Multi-hop networks transmit data packets through several relays in order to extend network coverage, avoid obstacles between a source and a sink, and provide users flexibility to design data routing maps to improve performance and energy efficiency. Three types of networks: star network, hybrid network and mesh network, are commonly used. The star topologies mainly depend on single-hop communications (Fig. 8.2a). The nodes only talk with their parents or sons. It has advantages in simplicity, easy installation, easy management and maintenance, and isolation of devices. However, the network performance highly depends on the hub, and the network scale is limited by the reachable distance of the nodes to the hub. The most efficient network uses peer-to-peer, mesh networks, in which all the nodes in the network have routing capability (Fig. 8.2b). Mesh networks allow nodes to self-assemble into the network, sensor information to propagate across the network with a high reliability and over an extended range, and time synchronization and low power consumption for the ‘listeners’ in the network to extend battery life (Karl and Willig, 2005). When a large number of wireless sensors need to be networked, several levels of networking may be combined. For example, an 802.11 (WiFi) mesh network

© Woodhead Publishing Limited, 2013

182 Robotics and automation in the food industry (a)

(b)

Hub

Hub

Fig. 8.2 Examples of wireless network topology: (a) star network; (b) mesh network.

comprising of high-end nodes, such as gateway units, can be overlaid on a ZigBee sensor network to maintain a high level of network performance. A remote application server (RAS) can also be deployed in the field close to a localized sensor network to manage the network, to collect localized data, to host web-based applications, to remotely access the cellular network via a GSM/GPRS or a CDMAbased modem and, in turn, to access the Internet and remote users (Crossbow Technology Inc., 2004).

8.2.6 Energy harvesting Most WSN platforms are powered by batteries, such as alkaline, cell, and lithium batteries, which can be replaced when needed. Although many manufacturers claim long battery life in their products, users often face much shorter battery life from the power sources, especially when external sensors, high sampling rate, and high data transmitting rate are used and extreme environment conditions are encountered. To develop long-lasting and truly-autonomous wireless WSNs, consistent and stable power sources are crucial. Now the ultra-low power consumption design on WSN hardware and software lessen the need of energy for operations. Micro-scale energy harvesting (at micro- or milli-watt level) from ambient environment is becoming a feasible approach to be incorporated with WSN applications. In the 2010–2011 Energy Harvesting Report, IDTechEx, a consulting firm in printed electronics, RFID, and energy harvesting, forecasts a market of over two billion dollars in 2016 just for the harvesting elements, excluding power storage and electronic interfaces and nearly 10 billion energy harvesting devices will be sold in 2020 (IDTechEx, 2010). There are two major tasks in energy harvesting, that is, energy capture and storage. The common targeted ambient energy sources include photovoltaics, temperature gradients, vibrations, pressure variations, flow of air/liquid, and electromagnetic (RF). Solar energy harvesting systems for WSN applications have been widely marketed and deployed. Other energy harvesting systems for WSNs are relatively new. Some manufacturers started to market their first generation power harvesters from spring 2010. Advanced Cerametrics Inc. in the USA provides Energy Harvestor™ which uses piezoelectric fiber composite material to harvest vibration and provide milliwatts energy. Powercast and Perpetuum in

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 183 the USA have developed power harvester receivers to hunt for energy in the RF field. Micropelt GmbH in Germany have marketed their battery-free TE-POWER NODE, which harvests thermal energy based on a heat flux through a thermogenerator. The company claimed that the thermogenerators can produce voltages in the range of 0.5–5 V on just a few square millimeters of footprint and are able to drive low power wireless sensors. Meanwhile, other venders are providing rechargeable energy storage device that will replace currently-used batteries. Cymbet Corp in the USA is developing a thin-film battery technology, EnerChip, which can be used to store energy collected from eight different sources, including vibration, thermal, RF, flow and solar-based systems. Infinite Power Solutions in the USA produces a family of solid-state, rechargeable, micro-energy cell (MEC) products named Thinergy. There is no doubt that harvesting energy from ambient environment offers an exciting future for long-term WSN deployments. However, more research and development and field testing on energy harvesting and storage technologies are still needed. In agricultural and food production applications, many sensors and actuators need higher current to drive, hence consume relatively more power during operations in comparison with those on-board semiconductor sensors. Hence, efficient power harvesting, management, and storage strategies need to be developed.

8.2.7 Data management The success of WSNs is determined by the quality of data generated, which need to be processed, filtered, interpreted, stored, and displayed to end users. On the one hand, WSNs provide users with detailed knowledge of target environments; on the other hand, they create a huge load on data processing and handling. The limited resources in terms of WSN components also challenges traditional methods of data management. In WSN applications, the data are commonly used in two ways: (1) queries on current data; (2) queries on historical data (Diao et al., 2005). The current data are often used for decision-making to determine control operations. For example, if a soil moisture level is below a predefined threshold, a water pump should be powered on. To save energy on communication, some push-down filter methods are used to preprocess raw data before transmitting (Ganesan et al., 2004; Diao et al., 2005). This is especially useful for multimedia WSNs with large data sets including images, audio and/or video streams. TinyDB, BBQ (Deshpande et al., 2004) and Direct Diffusion (Intanagonwiwat et al., 2003) provide tools for continuous queries to the current data. Another method, acquisitional query processing (AQP), offers functionality to determine which nodes, which parameters, and at what time to collect data (Ganesan et al., 2004). In many WSN applications, data collected by WSNs are streamed to a remote traditional database through various long-distance communication networks. End users can query stored ‘historical’ data any time. Data mining, artificial intelligence, and multivariate statistical analysis methods can be used to process, analyze, and query the data. Some new energy-efficient query methods and database

© Woodhead Publishing Limited, 2013

184 Robotics and automation in the food industry management systems are still under development that view the WSNs as a database that supports archival query processing (Diao et al., 2005).

8.3

WSN applications in agriculture and food production

WSN technology has rapidly penetrated into many applications including industrial process, medical equipment, environment monitoring, construction, agriculture, military, food tracking, merchandise management, inventory handling and recording, etc. It has been used in many new applications that were not previously considered possible. Agriculture and food production are very different from other industries. For example, crop fields are often remote and large but with variability in soil properties, crop conditions, and environment. Lack of electricity access is very common in the fields, while the crop canopy can often generate interference to radio propagation. In animal monitoring applications, wireless nodes installed on animals are movable. Hence, the WSNs have a dynamic network topology. Most importantly, except for value-added produce, technologies adopted by agricultural production need to be low cost, reliable, easy-to-use, and easy-to-maintain. Hence, WSN applications in agriculture and food production are behind the progress in other industries and areas.

8.3.1 Environment monitoring Many field measurements of environment variables, such as weather data and geo-referenced water quality data, still depend on stationary sensors and data loggers. They need regular visits by operators to download the data and change configurations and operations, if needed. Hence, they are labor-intensive and susceptible to recording errors during transcription (Vivoni and Camilli, 2003). WSNs are designed and developed to overcome these disadvantages by providing large-scale, unsupervised, real time, short-interval environment monitoring. Wireless devices are not new in environment monitoring. Many data loggers for remote environment monitoring have been equipped with radios or land-line telephone modems to bring users field measurement data. In recent years, with the expansion of cellular networks, Internet, and general packet radio service (GPRS), more and more data loggers have an internal cellular modem installed in order to load the data to a web server and be accessed by users from the Internet. The Oklahoma Mesonet is an environmental monitoring network with 120 automated stations covering 77 counties in Oklahoma (Mesonet, 2010). All the measurement data are transmitted to a central facility every 5 min, 24 h per day year-round, and can be accessed from Internet. Each station has a data logger with radio communication capability and sensors to take measurements including air and soil temperatures, wind speed and direction, rain fall, relative humidity, solar radiation, soil properties, etc. With an on-line tool, AgWeather, the Mesonet has been providing the latest weather data to help agriculture producers to make decisions on operations (Agweather, 2010).

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 185 In Italy, an advanced vineyard network system (NAV) was developed and deployed for remote, real time monitoring and micro-meteorological data collection in a vineyard (Matese et al., 2009). The system included a base agrometeorological station (master unit) and several wireless nodes (slave units). The master unit was a gateway that linked the slave units to a remote server through cellular services. Various sensors, including temperature, pressure humidity, solar radiation, rain fall, and anemometer, were attached to the master unit and each of the slave units through data acquisition boards. A GUI was developed to allow users to easily access the data and monitor the field conditions. Test results showed that the NAV system was able to provide continuous and real time environmental monitoring with flexibility on network design, configuration, and maintenance. Crossbow technology announced the eKoTM Pro series wireless crop monitoring system in January 2008 that consisted of multiple sensor nodes with mesh network architecture. The features of self-configuration and self-healing capability made the system deployment much easier for users. The system could collect information such as soil moisture, ambient temperature and humidity, leaf wetness, soil water content, solar radiation. A Web-interface, eKoView, provided a tool for users to visualize and analyze the collected data and make relevant decision on agricultural operations. Vivoni and Camilli (2003) developed a wireless prototype system to acquire, store, display and transmit real-time, geo-referenced environmental data between multiple field teams and remote locations. Each field team had a handheld data collection unit which could communicate to other teams or a field station server through a WLAN. The field station server combined information received from all the teams and periodically reported to a remote web/data server through a dual-frequency mobile phone (GSM/GPRS service at 900 MHz and 1.9 GHz). Field tests conducted in Maryland in the USA and New South Wales in Australia demonstrated a great potential to improve efficiency and precision on field measurement. A WSN system was reported by Zhang et al. (2010) for real-time remote monitoring of sediment runoff at a low-water crossing. The optical soil sediment sensor was submerged into the water while the rest of the system was located on the bank. The in-site network used simple but reliable star topology. Data was first transmitted from sensor nodes to a gateway based on ZigBee protocol and then from the gateway to the Internet server through a cellular network.

8.3.2 Precision agriculture Precision agriculture (PA) technology has been promoted and implemented around the world in recent decades. The factual base of PA is the spatial and temporal variability of soil and crop factors between and within fields. Traditionally, crops have been treated under the assumptions of ‘uniform’ soil, nutrient, moisture, weed, insect, and growth conditions. This has led to over- or under-applications of herbicides, pesticides, irrigation, fertilizers, and other treatments. The utmost

© Woodhead Publishing Limited, 2013

186 Robotics and automation in the food industry goal of PA is to apply appropriate treatments at right place and right time to pursue a low-input, high-efficiency, sustainable agricultural production. Spatial data collection A mobile field data acquisition system was developed by Gomide et al. (2001) to collect data for crop management and spatial-variability studies. The system consisted of a data collection vehicle, a manager vehicle and data acquisition and control systems on farm machines. The system was able to conduct local field surveys and to collect data on soil water availability, soil compaction, soil fertility, biomass yield, leaf area index, leaf temperature, leaf chlorophyll content, plant water status, local climate data, insect-disease-weed infestation, grain yield, etc. The data collection vehicle retrieved data from farm machines via a WLAN and analyzed, stored and transmitted the data to the manager vehicle wirelessly. The manager and engineers in the manager vehicle monitored the performances of the farm machines and the data acquisition systems, and troubleshot problems based on received data. Lee et al. (2002) developed a silage yield mapping system, which included a GPS, load cells, a moisture sensor and a Bluetooth wireless communication module. The moisture sensor and the Bluetooth transmitter were installed on a testing truck. Signals from the moisture sensor were sent to a Bluetooth receiver on a host PC at a data rate of 115 kbps and were used to correct the yield data. Li et al. (2009) reported a hybrid soil sensor network (HSSN) designed and deployed for in-situ, real-time soil property monitoring (Fig. 8.3). The HSSN included a local wireless sensor network (LWSN), formed by multiple sensor nodes installed at pre-selected locations in the field to acquire readings from soil property sensors buried underground at four depths and transmit the data wirelessly to a data sink installed on the edge of the field; and a long-distance cellular communication network (LCCN). The field data were transmitted to a remote web server through a GPRS data transfer service provided by a commercial cellular provider. The data sink functioned as a gateway which received data from all sensor nodes; repacked the data, buffered the data according to cellular communication schedule, and transmitted the data packets to LCCN. A web server was implemented on a PC to receive, store, process and display the real time field data. Data packets were transmitted based on an energy-aware self-organized routing algorithm. The data packet delivery rate was above 90% for most of the nodes. Precision irrigation Damas et al. (2001) developed and tested a distributed, remotely controlled, automatic irrigation system to control a 1500 ha irrigated area in Spain. The area was divided into seven sub-regions with 1850 hydrants installed. Each sub-region was monitored and controlled by a control sector. The seven control sectors communicated with each other and a central controller through a WLAN network. Field tests showed 30–60% saving in water usage. Evans and Bergman (2003) led a USDA research group to study precision irrigation control of self-propelled, linear-move and center-pivot irrigation systems.

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 187

PC Sensor node

Network user

Field Central node

Gate way

Web server (data storage)

Internet Cellular modem Cellular tower Local wireless sensor network

Long range cellular network

Fig. 8.3 WSN for soil property monitoring. (Source: Li et al., 2009.)

Wireless sensors were used in the system to assist irrigation scheduling based on on-site weather data, remotely sensed data and grower preferences. O’Shaughnessy and Evett (2010) established a test platform including two WSNs of infrared thermometers to monitor canopy temperature, which was commonly used as an indicator of crop water stress. One of the WSNs was installed on the lateral arm of a center-pivot irrigation system and used to monitor crop canopy temperatures while moving. Another WSN formed by stationary nodes was installed in the field below the pivot arm to provide the stationary reference canopy temperature. RF modules, XBee (MaxStream, Logan, Utah, USA), were used for communications. An embedded computer was used for recording and analyzing the collected data and making irrigation decisions. Kim et al. (2006) developed and tested a closed-loop automated irrigation system. The system consisted of in-field sensing stations, an irrigation control station, a weather station, and a base station. The sensing station and the weather station provided in-field status readings. The base station collected and processed the in-field data, made decisions on irrigation scheduling, and sending the control commands to the irrigation control station to control the operations of sprinkler nozzles. The irrigation control station also updated GPS locations of the linear irrigation system regularly and sent them to the base station. The wireless communications among the stations were through Bluetooth technology. King et al. (2005) developed a closed-loop, distributed control and data acquisition system for site-specific irrigation management for a center pivots system. The system was formed by a group of stationary field sensing nodes deployed in a field and a group of nodes, called P2K nodes, installed on the lateral arms of the central pivot system. The field sensor nodes communicated with the P2K nodes

© Woodhead Publishing Limited, 2013

188 Robotics and automation in the food industry using low power RF signals. The P2K nodes collected the data from the field and the central pivot system and transmitted them to a master controller through a power-line carrier communication. The system performance was compared with a conventional uniform irrigation treatment in a potato field. Results showed that, with essentially equal water consumption, the tuber yield under site-specific irrigation management was significantly greater, by about 4%, than the yield from uniform irrigation.

8.3.3 Machine and process control More and more technologies are being developed to support wired or wireless (WPAN, WLAN, cellular networks, etc.) communications from machine to machine, from machine to mobile or from mobile to machine, and from human to machine or machine to human. These technologies greatly enhance the automation of machines, the efficiency and effectiveness of machine management systems through an integration of discrete components within the system. Yang et al. (2010) developed a farm agriculture machinery information management system, that integrated tasks including information management, farm machinery job scheduling, machinery performance evaluation, work-load distribution, fuel and spare parts supply information, maintenance scheduling, statistical analysis on technical and economic indicators, financial management and on-line meeting support systems. The system was deployed in farms in Heilongjiang, China. Each agri-machine in the tested farm was equipped with a GPS unit and a cellular device. It reported its locations and performance parameters regularly to a central management center through GPRS service provided by cellular carriers. Some machinery had camera systems on board, which allowed the operator to communicate with the central manager through video conferencing services. The field tests showed a great improvement on task scheduling for agri-machinery uses. More and more agri-machinery manufacturers integrate communication modules to their systems to allow the machinery to remotely report current information such as geographical location, fuel condition, system conditions, accidents, etc. The JDLink product line from John Deere allows the equipment manager to access daily data via the Internet to a central server confidentially. Malfunctioning alarm messages can be delivered to the manager through emails and short text messages. Some New Holland combine-harvesters have network cards developed by EIA Electronics in Belgium that allow a manager to remotely monitor their operations via cellular network or satellite communication. A German company, RTS Rieger, have developed a system to record various operating parameters of machinery and transmit the data to a PDA (without service subscription fee) through Bluetooth or to a webserver via cellular network. A WLAN-based, real-time, vehicle-to-vehicle data communication system was established by Guo and Zhang (2002) to exchange information between vehicles on vehicle state and operation control variables. Laboratory and field tests demonstrated the feasibility of real-time, wireless data communications between vehicles in autonomous, master-slave vehicle guidance. Charles and Stenz (2003)

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 189 implemented an autonomous tractor for spray operations in fields. During spraying, the tractor drove fully autonomously at least 90% of the time. This tractor could also be precisely controlled by a supervisor through a radio link. Ribeiro et al. (2003) developed an autonomous guidance tractor for spray operations in citric and olive trees fields in Spain. A user-friendly visualization agent was developed for human operators to remotely control and supervise unmanned tractors in a field through WLAN. Stentz et al. (2002) developed a wireless link between tractors and a human supervisor in a fleet of semi-autonomous tractors. Each tractor had the capability to detect people, animals and other vehicles in its predefined path and to stop before hitting such obstacles until it received control commands from a supervisor over a wireless link.

8.3.4 Precision livestock farming Precision livestock farming aims to improve animal productivity and production sustainability by feeding the right animal at the right place with right amount at the right time. To achieve that goal, animals need to be carefully monitored and managed with specific temporal and spatial resolution. However, animals are moving objects, which makes wired sensing methods infeasible. WSN technology and miniature electronic devices open a door for precision livestock management. In recent years, much research on the development of animal monitoring systems have been reported. Animal behavior monitoring A WSN system was developed to study the lengths of time cows spent near a water trough (Kwong et al., 2008). Each cow in the study wore a collar with a sensor node including a GPS receiver, a microprocessor, and a radio transceiver. The sensor node recorded GPS data at predefined time intervals and uploaded the stored data to a base station wirelessly, using 802.15.4 standard protocol, whenever it entered the communication range of the base station. A cattle management WSN system was demonstrated to monitor cows’ presence and pasture time in a strip of new grass in real-time (Nadimi et al., 2008). Micaz sensor motes from Crossbow Technology were used and enclosed in collars worn by cows. A gateway was installed in the area of new strip of grass. It received and stored data packets when a cow entered to its communication area. The animal pasture time was then calculated. Monitoring cattle grazing activity is not only important to know their eating behavior, but is also a benefit to grassland management. Reed (2008) developed a wireless sensor to monitor free-range grazing activity. The sensor consisted of a microcontroller, signal conditioning, and a radio transceiver integrated on a PCB board with dimensions of 19.6 mm × 71.8 mm × 11 mm. A miniature GPS logger was also included in the system. Data processing algorithms were developed to characterize grazing motions against normal movements. The field tests showed that the system was able to provide geo-referenced cattle grazing maps.

© Woodhead Publishing Limited, 2013

190 Robotics and automation in the food industry Owing to concerns regarding animal welfare and food safety, non-cage housing systems are becoming more and more popular in the poultry industry. In the United States, many egg producers are in the process of transition from traditional operations to non-cage operations. The dynamics of hens’ behaviors under such noncage environment need to be studied and carefully monitored in order to make an appropriate management control strategy and discover unhealthy animals at an early stage. Quwaider et al. (2010) developed a platform to monitor hens’ activity. The platform included wireless sensors mounted on hens and stationary wireless sensors. The platform provided information including hens’ relative locations to stationary sensors and other hens, speed of movements, modeling of locomotion (flying or ambulating), and specific behaviors (eating, drinking, and egging). The Mica2Dot mote from Crossbow Technology was used as the wireless sensor with on-board light, acceleration, temperature, and sound sensors. The relative location of the hens was determined based on radio-signal-strength-indicator (RSSI). Preliminary results showed that the platform was able to locate hens with an accuracy of 80–90%. This result demonstrated that the developed system was able to detect when and where hens were active. This information could be invaluable to the producers to monitor hens in real time, record activity patterns, such as egg laying, eating, drinking, and dust bathing, etc., and adjust operations in real time. Animal health monitoring Pastell et al. (2009) reported a wireless sensing system developed to measure gait features in dairy cows. Each sensor node with a three-dimensional accelerometer was attached to a limb proximal to fetlock joint of a cow. Three-dimensional movements of limbs were logged to the sensor node and transmitted to a computer wirelessly using the 869 MHz radio channel to determine the lameness level of cows. A wireless floating base sensor network system was developed in Brazil to monitor animal physiological responses (de Sousa Silva et al., 2005). The system included two modules, animal module and base module. The animal modules were installed on tested animals to collect EEG signals and transmit the signals to the base module using 433 MHz carrier frequency. The base module was used to receive and store the EEG signals which could be used to monitor bovine brain electrical activity. Deep body temperature (DBT) was commonly used as an indicator of heat stress of broilers. Wireless sensors were developed to continuously measure DBT of broilers (Yang et al., 2007). Each bird was implanted with a wireless sensor at a depth of 2 cm into the abdominal cavity, which transmitted DBT measurements to a stationary receiver in a broiler house. The results showed that the wireless sensors functioned well and were responsive the changes of DBT. Animal environment monitoring Precise controls of climate-related variables within an animal house can improve animal welfare and maintain good health. Pessel and Denzer (2003) developed a portable, mobile instrument to measure temperature, relative humidity, noise, brightness and ammonia content in the air within an animal house and transferred

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 191 the data wirelessly to a PC through an infrared data link. Darr and Zhao (2008) developed a wireless sensor and control system to monitor temperature variations and control ventilation in a swine house. Nadimi et al. (2008) developed a WLAN sensor network and text-message delivery system to observe and alert climate changes in a broiler house. Wireless temperature sensor nodes and web-cameras were installed in multiple broiler houses and networked through WLAN. The data were streamed into a remote server which sent text messages when abnormal conditions were detected.

8.3.5 Greenhouse management Greenhouse production needs many sensors and controls to ensure productivity. They have been used to monitor indoor climate changes, soil conditions, crop health, and pest infestations and control the use of resources such as light, water, chemicals, etc. Due to the large amount of wired sensing and control systems used, greenhouses have a reputation as ‘houses full of wires and cables.’ A big effort is needed to handle the wires and cables during system installation and maintenance. When wireless technology appeared, greenhouse producers were pioneers in adopting it. They replaced the old systems with cleaner, more reliable, more flexible, and more efficient wireless systems. Most greenhouses have power and Internet accesses, which also make the deployment of WSNs much easier than outdoor applications. Morais et al. (1996) implemented a wireless data acquisition network to collect outdoor and indoor climate data for greenhouses in Portugal. Several solar-powered data acquisition stations (SPWAS) were installed indoors and outdoors to measure and monitor the climate data. RF links were established among multiple (up to 32) SPWASs and a base station, which was used to control the SPWASs and to store the data. Serôdio et al. (1998, 2001) developed and tested a distributed data acquisition and control system for managing a set of greenhouses. Several communication techniques were used for data communications. At a lower supervisory level, inside each greenhouse, a WLAN network with a RF of 433.92 MHz was used to link a sensor network to a local controller. A controller area network (CAN) was provided to link an actuator network to the local controller. Through another RF link (458 MHz), several local controllers were connected to a central PC. High level data communication was provided through Ethernet to connect the central PC to a remote network. Liu and Ying (2003) reported a greenhouse monitoring and control system using the Bluetooth technology. The system collected environmental data from a sensor network in a greenhouse and transmitted this data to a central control system. Mizunuma et al. (2003) deployed a WLAN in a farm field and greenhouse to monitor plant growth and implemented remote control for the production system. They believed that this type of remote control strategy could greatly improve productivity and reduce labor requirement. Systems based on ZigBee technology are widely used in greenhouse management to form mesh networks, which makes the system configuration and maintenance more convenient. Zhu et al. (2006) established a WSN in a greenhouse to

© Woodhead Publishing Limited, 2013

192 Robotics and automation in the food industry collect temperature, humidity, and light intensity periodically. The sensor nodes communicated with a gateway using ZigBee protocol, while the gateway node communicated with a web server through Ethernet. Park et al. (2009) developed an automatic control system for greenhouse use to collect temperature and humidity of crop leaves. Sensor nodes were installed inside a greenhouse to collect air temperature, humidity, leaf temperature, and leaf humidity. Through ZigBee compliance radio in the sensor nodes, the collected data were transmitted to a base station and stored in a database. The base station periodically sent the data to a remote control system via the Internet, which analyzed the data and made decisions on the control strategy. The control commands were then transmitted to a control station in the greenhouse to implement relevant operations on ventilation, irrigation, and heating, etc. Lea-Cox et al. (2007) established a WSN to measure substrate water, temperature, electrical conductivity, daily photosynthetic radiation and leaf wetness in real-time in nursery houses. These data could be used for the control of irrigation. A web-based graphical reporting system was also developed to provide easy access of the data via the Internet. Growers who adopted the technology could benefit from improved plant health, more efficient water and fertilizer applications, and a reduction in disease problems related to over-watering.

8.3.6 Food traceability systems A traceability system is a record-keeping system that is able to trace the flow of a particular product from its origin, through all intermediate steps (e.g. processing and supply chain), to consumers. Traceability is not new to the agricultural and food communities. Various methods to record product information have been in use for decades, for example, Tattoos, brands, barcodes, tags, labels, forms and tables. The recent RFID technology brings in new functions beyond ‘record-keeping’. Wentworth (2003) conducted a study aimed at inexpensive, disposable RFID biosensor tags used on food products for history checking and contamination and inventory control. The biosensor was based on an acoustic wave platform and used antigen–antibody reaction to detect bacteria. Gebresenbet et al. (2003) proposed an on-the-road monitoring system for animals during transportation. The system included sensors installed in the animal compartment to identify the animals and to monitor the air-quality, vibration and animal behaviors. A GPS provided the location of the vehicle. A data transfer unit regularly sent data to a service center via the GSM network. They reported that the system greatly improved animal welfare during handling and transportation. Abad et al. (2009) tested a refrigerated truck loaded with fish from Frankfurt Airport in Germany to Vitoria Airport in Spain. The truck was equipped with a RFID-based system including smart tags attached to fish packages and a reader/writer installed on the truck. The smart tags measured and transmitted light, humidity, and temperature data every 2 min at 13.56 MHz frequency. The reader collected and stored the data and set an alarm on abnormal conditions.

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 193 Obstacles and special issues Due to the nature of agricultural production, wireless communication which links ‘fields’ to ‘office(s)’ and allows in-time data exchanges can be used to improve management and decision-making processes. With the significant reduction in the cost of electronics, the use of WSN technology becomes feasible in agricultural applications. Yet, despite of its promising advantages, WSN technology is still facing numerous challenges during the deployment for agriculture and food applications. To be used in agricultural and food production, WSNs need to maintain reliable and stable performance under various environmental conditions, many of which are unpredictable. Most wireless sensor nodes are deployed in outdoor, remote areas. The temperature variations can be very significant which often make the systems run in extreme conditions. Sunshine, rain, snow, and wind can also significantly affect the system performance. Wild animals and theft can cause damage to the systems. Regular field operations, such as tillage, chemical applications, and harvesting may interrupt the system running. Hence, the design considerations of WSNs used in crop fields includes not only the performance of electronic systems but also system packaging, protection schemes, and field installation issues. In contrast to indoor applications, most field applications have no access to a fixed power supply. The wireless sensor nodes have to depend on their own power source, commonly batteries, to complete their tasks and operate for a long period of time. There are two solutions to accomplish the goals: (1) to minimize power consumption of each sensor node and (2) to keep a stable power source on the sensor nodes. Minimizing the power consumption can be achieved by carefully designing hardware and software systems and communication protocols. For example, most wireless sensor nodes have adopted low-power consumption hardware components and small footprints of software. The scheduling of ‘wake up’ and ‘sleep’ states has also been carefully developed. These methods can be seen as passive solutions. A more active approach is to maintain a stable power source by applying efficient power harvesting techniques. An agricultural field environment is full of potential natural energy sources, such as solar, wind, hydraulic movements, thermal variations, RF signals, and vibrations. Energy harvesters can collect these energies from surrounding environment and use them to charge the power source for the sensor node. The combinations of two or three forms of energy harvesters (e.g. solar and wind, or solar and hydraulic, etc.) can provide a weather-proof energy source to charge and maintain a stable power supply for a sensor node. Recently, miniature energy harvesters have been commercialized for WSN applications. They can be a driving-force to speed up the deployment of WSN technology. Wireless technology, especially cellular networks, has been growing rapidly in past decades. Many technical methods, algorithms, and models are mature and widely adopted. However, to apply them in agricultural applications, many adjustments and corrections are needed. For example, the radio antennas on sensor nodes installed in fields are much closer to the ground and canopy than

© Woodhead Publishing Limited, 2013

194 Robotics and automation in the food industry those in cellular towers. The interference from the canopy can be much more significant. The canopy variations during growth stages can strengthen the impact. The sensor nodes installed in animal houses or transportation trucks can be easily affected by humidity and dust. Hence, the models used for wireless communication need to be recalibrated and corrected based on actual situations. The WSNs for agricultural production are often used to acquire a wide-range of data, from soil properties, and plant status, to weather conditions. Sensors, controllers and actuators from different venders are often used in one system. Hence, the data are typically multi-source and heterogeneous. To integrate all the components, technical standards on system design and integration becomes very important. The establishment of standards to define physical layer, data interface, and testing procedures are urgently needed. The massive amounts of data generated by the WSNs are more difficult to handle in terms of hardware, communication, and processing in comparison to traditional methods. Manual analysis of the data becomes infeasible. Higher hardware cost will be induced to store and process the data. Wider bandwidth is needed to transmit the data within the network. Hence, more efficient data management strategies need to be developed to maximize the use of resources while maintaining minimum cost. For example, preprocessing raw data on sensor nodes can save memory space to store the data and bandwidth to transmit the data. But the sensor nodes may need more calculating power in order to implement preprocessing procedures. How to balance the calculation load, transmission load, power consumption, and available resources on the sensor nodes within WSNs will be a big challenge for WSN developers. Agriculture and food production is a traditional industry with a long history of technology adoption. Many producers have been seeking new technologies to improve their narrow profit-margins. Meanwhile, they are very cautious about adopting them to avoid potential negative impacts. Hence, to build customers’ confidence, WSNs need to perform consistently and reliably. Another important aspect is that WSN systems need to be easy-to-learn, easy-to-use, and easy-to-maintain. Users often have limited knowledge on electronics, programming, and networking. The WSN systems are expected to work in the form of ‘plug-and-play’ to allow autonomous network establishment, configuration, operation, and maintenance. As mentioned before, WSNs are still only beginning to be widely adopted by the agricultural industry. The majority of WSNs have been developed for research purposes. Although many research outcomes have demonstrated their strengths in improving the management of production, there is no quantitative evaluation on how the WSN technology is beneficial to the industry in terms of productivity and efficiency. This evaluation needs long-term, systematic experiments, which should be researched soon.

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 195

8.4

Future trends in WSN technology in agriculture and food production

In the past, excessive wiring was considered a sign of technological advancement in many sensing and control systems. In 1955, the total length of electrical wires on a state-of-the-art automobile was about 45 m. This number grew to more than 5 km on a high-end vehicle in recent years. The thick wiring harness is often hard to manage and difficult in diagnosing problems. In field applications, a lot of efforts are given to protect the wiring harness from damages of weather, animals, and theft. Many greenhouses are full of wires on the floor and ceilings which needs to be carefully protected to avoid shortage due to high moisture and humidity. WSN technology shows great potential to overcome these shortcomings and be economically viable to replace wired networks. The autonomous massive data collection provides user detailed information in a target environment. Combining historical data and agronomic knowledge, more effective management strategy can be made to improve the production. As the demand for food quality, health benefits, and safety increases, more stringent scrutiny on the inspection of agricultural and food products have become mandatory. Also being increasingly demanded is ‘traceability’, which requires not only rigorous inspections, but also systematic detection, labeling and recording of quality and safety parameters while archiving the entire agricultural production chain, from farms to consumers’ tables. RFID has shown a booming trend with adoption by producers, food processing and handling industry, and merchants to establish an effective ‘traceability system’. RFID allows an ‘intelligent tag’ assigned to each individual product to be read at any check point. The intelligent tag can be updated along the entire supply chain to provide complete archives of information on the growth, processing, packaging, transportation, distribution, storage, shelving, and recycling. Active RFID systems can also record environmental parameters and specific quality/safety attributes of the product along the chain. Communication has changed significantly over the past decade. Various new devices including smart cell phones, Internet, Tablet PCs, etc., become imperative in human life. These new technologies greatly shorten the distances between people, between nature and human beings, and between goods and consumers. There is no doubt that WSNs will soon penetrate all aspects of agriculture production, food process, handling, and storage, transportation, and marketing and trading.

8.5

References

Abad, E., F. Palacio, M. Nuin, A. González de Zárate, A. Juarros, J.M. Gómez and S. Marco. 2009. RFID smart tag for traceability and cold chain monitoring of foods: Demonstration in an intercontinental fresh fish logistic chain. Journal of Food Engineering. 93(4): 394–399. Agweather, 2010. http://agweather.mesonet.org/. Accessed on 20 October 2010. Bhatti, S., J. Carlson, H. Dai, J. Deng, J. Rose, A. Sheth, B. Shucker, C. Gruenwald., A. Torgerson and R. Han. 2005. Mantis OS: An embedded multithreaded operating

© Woodhead Publishing Limited, 2013

196 Robotics and automation in the food industry system for wireless micro sensor platforms. ACMK Mobile Network Journal, Special Issue on Wireless Sensor Networks. Boegena H., K. Schulz and H. Verrechken. 2006. Towards a network of Observatories in Terrestrial Environmental Research. Advances in Geosciences, 9: 109–114. Britton, M., V. Shum, L. Sacks, and H. Haddadi. 2005. A biologically inspired approach to designing wireless sensor networks, in Proceedings of the Second European Workshop on Wireless Sensor Networks, Istanbul, Turkey. Campbell Scientific, 2010. http://www.campbellsci.com/. Accessed on 10 October 2010. Charles, K. and A. Stenz. 2003. Automatic Spraying for Nurseries. USDA Annual Report. Project Number: 3607–21620-006-03. 22 September 2000–31 August 2003. USDA, USA. Darr, M.J. and L. Zhao. 2008. A wireless data acquisition system for monitoring temperature variations in swine barns. Proceedings of Livestock Environment VIII, The Eighth ASABE International Symposium. Iguassu Falls, Brazil, 31 August–September 2008. ASABE Publication Number 701P0408. Decagon, 2010. http://decagon.com/, Accessed on 10 October 2010. Deshpande, A., C. Guestrin, S. Madden, J. Hellerstein and W. Hong. 2004. Model-driven data acquisition in sensor networks. Proceedings of the Thirtieth International Conference on Very Large Databases, 30: 588–599. Diao, Y., D. Ganesan, G. Mathur and P. Shenoy. 2005. Rethinking data management for storage-centric sensor networks. The 2005 SIAM International Conference on Data Mining, Newport Beach, CA, 21–23 April 2005. Ember, 2010. http://www.ember.com/index.html, Accessed on 10 October 2010. Evans, R. and J. Bergman. 2003. Relationships between cropping sequences and irrigation frequency under self-propelled irrigation systems in the northern great plains (Ngp). USDA Annual Report. Project Number: 5436–13210-003-02. 11 June 2003–31 December 2007. Ganesan, D., D. Estrin and J. Heidemann. 2004. Dimensions: Why do we need a new Data Handling architecture for Sensor Networks? The Proceedings of Information Processing in Sensor Networks, Berkeley, California, USA, 2004. Gay, D., P. Levis, R.V. Behren and M. Welsh. 2003. The nesC Language: A holistic approach to networked embedded system. http://nescc.sourceforge.net/papers/nesc-pldi-2003. pdf. Gebresenbet, G., D. Ljungberg, G. Van de Water and R. Geers. 2003. Information monitoring system for surveillance of animal welfare during transport. The Proceedings of the 4th European Conference in Precision Agriculture, Berlin, Germany, 14–19 June 2003. Gomide, R.L., R.Y. Inamasu, D.M. Queiroz, E.C. Mantovani and W.F. Santos. 2001. An automatic data acquisition and control mobile laboratory network for crop production systems data management and spatial variability studies in the Brazilian center-west region. ASAE Paper No.: 01-1046. The American Society of Agriculture Engineers, St. Joseph, Michigan, USA. Guo, L.S. and Q. Zhang. 2002. A wireless LAN for collaborative off-road vehicle automation. The Proceedings of Automation Technology for Off-Road Equipment Conference, Chicago, Illinois, USA, 26–27 July, 51–58. Han, C.C., R. Kumar, R. Shea, E. Kohler and M.B. Srivastava. 2005. A dynamic operating system for sensor nodes. Proceedings of the 3rd international conference on Mobile systems, applications, and services, ACM New York, NY, USA. ISBN:1-931971-31-5, 163–176. IDTechEx. 2010. http://www.idtechex.com/research/reports/energy_harvesting_and_storage_for_electronic_devices_2010_2020_000243.asp IEEE, 1999. Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Higher-Speed Physical Layer Extension in the 2.4 GHz Band. IEEE

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 197 Standard 802.11b. The Institute of Electrical and Electronics Engineers Inc., 345 East 47th Street, New York, USA. IEEE, 2002. Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Wireless Personal Area Networks (WPANs). IEEE Standard 802.15.1. The Institute of Electrical and Electronics Engineers Inc., 345 East 47th Street, New York USA. IEEE, 2003. Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (LR-WPANs). IEEE Standard 802.15.4. The Institute of Electrical and Electronics Engineers Inc., 345 East 47th Street, New York, USA. Intanagonwiwat, C., R. Govindan, D. Estrin and J. Heidemann. 2003. Directed diffusion: A scalable and robust communication paradigm for sensor networks. The IEEE/ ACM Transactions on Networking (TON), 11(1): 2–16. Jennic, 2010. http://www.jennic.com/index.php. Accessed on 10 October 2010. Karl, H. and Willig, A., 2005. Protocols and Architectures for Wireless Sensor Networks. Wiley. ISBN: 978-0-470-09510-2. Kim, H. and S. Hong. 2005. State machine based operating system architecture for wireless sensor networks. Parallel and Distributed Computing: Applications and Technologies, 3320: 803–806. Kim, Y., R.G. Evans, W. Iversen and F.J. Pierce. 2006. Instrumentation and control for wireless sensor network for automated irrigation, The 2006 ASABE Annual International Meeting, Paper No. 061105, Portland, Oregon. 16–19 July 2006. King, B.A., R.W. Wall and L.R. Wall. 2005. Distributed control and data acquisition system for closed-loop site-specific irrigation management with center pivots. Applied Engineering in Agriculture, 21(5): 871–878. Kwong, K.H., H.G. Goh, C. Michie, I. Andonovic, B. Stephen, T. Mottram and D. Ross. 2008. Wireless sensor networks for beef and dairy herd management, The 2008 ASABE Annual International Meeting, Paper No. 084587, Providence, Rhode Island, 2008. Lea-Cox, J.D, G. Kantor, J. Anhalt, A. Ristvey and D.S. Ross. 2007. A wireless sensor network for the nursery and greenhouse industry. Proceedings of the 52 Annual Southern Nursery Association Research Conference, August 8-9, 2007, 52:454-458, Atlanta, GA, USA. Lee, W.S., T.F. Burks and J.K. Schueller. 2002. Silage yield monitoring system. ASAE Paper No.: 02-1165. The American Society of Agriculture Engineers, St. Joseph, Michigan, USA. de Sousa Silva, A.C., A.I.C. Arce, S. Souto and E.J.X. Costa. 2005. A wireless floating base sensor network for physiological responses of livestock. Computers and Electronics in Agriculture, 49(2): 246–254. Li, Z., N. Wang, A. Franzen, P. Taher and X. Li. 2009. In-field soil profile property monitoring system based on a hybrid sensor network. ASABE Paper No: 096191. The 2009 ASABE Annual Meeting, 21–24 June 2009, Reno, Nevada. Liu, G. and Y. Ying. 2003. Application of Bluetooth technology in greenhouse environment, monitor and control. Journal of Zhejiang University – Agricultural and Life Science, 29: 329–334. Mahfuz, M. and K. Ahmed. 2005. A review of micro-nano-scale wireless sensor networks for environmental protection: prospects and challenges. Science and Technology of Advanced Materials, 3–4: 302–306. Matese, A., S.F. Di Gennaro, A. Zaldei, L. Genesio and F.P. Vaccari. 2009. A wireless sensor network for precision viticulture: The NAV system. Computers and Electronics in Agriculture, 69(1): 51–58. Mesonet, 2010. http://www.mesonet.org/, Accessed on 20 October 2010. Microstrain, 2010. http://www.microstrain.com/wireless, Accessed on 30 August, 2012.

© Woodhead Publishing Limited, 2013

198 Robotics and automation in the food industry Nadimi, E.S., H.T. Sogaard, T. Bak and F.W. Oudshoorn. 2008. ZigBee-based wireless sensor networks for monitoring animal presence and pasture time in a strip of new grass, Computer and Electronics in Agriculture, 61(2): 79–87. Mizunuma, M., T. Katoh and S. Hata. 2003. Applying IT to farm fields – A Wireless LAN. NTT Technical Review, 1: 56–60. Morais, R., J.B. Cunha, M. Cordeiro, C. Serodio, P. Salgado and C. Couto. 1996. Solar data acquisition wireless network for agricultural applications. The Proceedings of the 19th IEEE Convention in Israel, Jerusalem, Israeli, 5–6 November, 527–530. Onset, 2010. http://www.onsetcomp.com/index.php, Accessed on 10 October 2010. Park, D., B. Kang, K. Cho, C. Shin, S. Cho, J. Park and W. Yang. 2009. A study on greenhouse automatic control system based on wireless sensor network. Wireless Personal Communications. DOI: 10.1007/s11277-009-9881-2. Springer. Rajkumar, R., K. Juvva, A. Molano and S. Oikawa. 1997. Resource kernels: A resource-centric approach to real-time and multimedia systems. Journal of SPIE, 3310: 150–164. O’Shaughnessy, S.O. and S.R. Evett. 2010. Developing wireless sensor networks for monitoring crop canopy temperature using a moving sprinkler system as a platform. Applied Engineering in Agriculture, 26(2): 331–341. Pastell, M., J. Tiusanen, M. Hakojarvi and L. Hanninen. 2009. A wireless accelerometer system with wavelet analysis for assessing lameness in cattle. Biosystems Engineering, 104: 545–551. Pessel, G.J. and H. Denzer. 2003. Portable and mobile instrument for continuous stable climate measurement. The Proceedings of the 4th European Conference in Precision Agriculture and the 1st European Conference on Precision Livestock Farming, Berlin, German, 14–19 June 2003. Quwaider,, M., C.L., Daigle, S.K., Biswas, J.M., Siegford and J.C., Swanson. 2010. Development of a wireless body-mounted sensor to monitor activity and location of laying hens in a Non-cage housing system. Applied Engineering in Agriculture. 53(5): 1705–1713. Reed, S. 2008. Custom wireless sensor for monitoring grazing of free-range cattle. PhD Thesis. Oklahoma State University, USA. Ribeiro, A., L. Garcia–Perez, Garcia-Alegre and M.C. Guinea. 2003. A friendly man-machine visualization agent for remote control of an autonomous tractor GPS guided. The Proceedings of the 4th European Conference in Precision Agriculture, Berlin, Germany, 14–19 June 2003. Ruiz-Garcia, L., L. Lunadei, P. Barreiro and J. I. Robia. 2009. A review of wireless sensor technologies and applications in agriculture and food industry: State of art and current trend. Sensors, 9: 4728–4750. Sentilla, 2010. http://www.sentilla.com/, Accessed on 10 October 2010. Serôdio, C., J.L. Monteiro and C.A. Couto. 1998. Integrated network for agricultural management applications. The Proceedings of the 1998 IEEE International Symposium on Industrial Electronics, Pretoria, South Africa, 7–10 July 1998, 2: 679–683. Serôdio, C., J.B. Cunha, R. Morais, C.A. Couto and J.L. Monteiro. 2001. A networked platform for agricultural management systems. Computers and Electronics in Agriculture, 31: 75–90. Shockfish, 2010. http://tinynode.com/index.php?id=99, Accessed on 10 October 2010. Stentz, A., C. Dima, C. Wellington, H. Herman and D. Stager. 2002. A system for semi-autonomous tractor operations. Autonomous Robots, 13: 87–104. TekVet, 2010. http://www.tekvet.com/, Accessed on 10 October 2010. TinyOS, 2009. http://en.wikipedia.org/wiki/TinyOS, Accessed on 9 June 2009. TinyDB. 2010. http://telegraph.cs.berkeley.edu/tinydb/, Accessed on 10 October 2009. Vivoni, E.R. and R. Camilli. 2003. Real-time streaming of environmental field data. Computers & Geosciences. 29(2003): 457–468.

© Woodhead Publishing Limited, 2013

WSNs in the agricultural and food industries 199 Wang, N., N.Q. Zhang and M.H. Wang. 2006. Wireless sensors in agriculture and food industry – recent development and future perspective. Computers and Electronics in Agriculture, 50: 1–14. Wentworth, S.M. 2003. Microbial sensor tags. The 2003 IFT (The Institute of Food Engineering) Annual Meeting Book of Abstracts, Geosciences. Chicago, Illinois, USA, 12–16 July 2003, 29: 457–468. Yang, H., X. Wang and W. Zhuang. 2010. Case analysis of farm agriculture machinery informatization management network system. Computer and Computing Technologies in Agriculture III: IFIP Advances in Information and Communication Technology, 317: 65–76. Yang, H.H., Y.H. Bae and W. Min. 2007. Implantable wireless sensor network to monitor the deep body temperature of boilers. Software Engineering Research, Management & Applications, 2007. The 5th ACIS International Conference, 513–517. Zhang, N., M. Wang and N. Wang. 2002. Precision agriculture – a worldwide overview. Computers and Electronics in Agriculture, 36(2–3): 113–132. Zhang, N., W. Han, N. Wang, J. Dvorak, X. Wang, C. Johnson, D. Oard, D. Bigham, T. Peyman, S. McClung and J. Steichen. 2010. Large-scale, real-time monitoring of sediment concentration and sediment movement using three-tier wireless sensor networks. Paper No. 1009515. The 2010 ASABE Annual Meeting, 21–24 June 2010, Pittsburg, Pennsylvania. ZIGBEEF, 2010. http://www.zigbeef.com/index.html, Accessed on 10 October 2010. Zhou, Y.M., X.L. Yang, X.S. Guo, M.G. Zhou and L.R. Wang. 2007. A design of greenhouse monitoring & control system based on ZigBee Wireless Sensor Network. The Proceedings of the 2007 International Conference of WiCom (Wireless Communications, Networking and Mobile Computing), 21–25 September 2007. Shanghai, China, 2563–2567. Zhu, Y.W., X.X. Zhong and J.F. Shi. 2006. The design of wireless sensor network system based on ZigBee technology for greenhouse. Journal of Physics, 48: 1195–1199.

© Woodhead Publishing Limited, 2013

9 Intelligent quality control systems in food processing based on fuzzy logic N. Perrot and C. Baudrit, INRA, France

DOI: 10.1533/9780857095763.1.200 Abstract: In the food industry, end-products must achieve a compromise between several properties, including sensory (texture or color for example), sanitary, and technological properties (dimensions of a biscuit or mass loss of a cheese for example). Managing these properties right from the fabrication stage with the aim of controlling them is no easy task. At present, many production processes are largely reliant on the skill and experience of the operator, something that no system will be capable of replacing in the foreseeable future. The upshot is that the quality of a food product cannot be described exclusively through strict transitions between the crisp qualifiers ‘good’ or ‘bad’, but instead requires an infinite series of potentially vague grades to be properly defined and, ultimately, controlled. This chapter outlines the key functions of fuzzy sets and proposes possibility theories to integrate skill and experience for control purposes within the food industry. Key words: fuzzy logic, possibility theory, uncertainty, food control systems, expert knowledge.

9.1

Introduction

In the food industry, end-products must achieve a compromise between several properties, including sensory, sanitary, and technological properties. Sensory and sanitary properties are essential as they influence consumer choice and preference. However, managing these properties right from the fabrication stage for subsequent control is no easy task, for several reasons (see Mittal, 1997): 1. The food industry works within many parameters that must be taken into account in parallel. A single sensory property, such as color or texture, can be linked individually to several dimensions registered by the human brain. For example the color of a biscuit evaluated on line by operators using their senses is the synthesis of three instrumental measurements represented as L, a, b.

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 201 2. The food industry works with non-uniform variable raw materials that, when processed, need to lead to a product that satisfies a fixed standard. 3. The phenomena involved in food processing are highly non-linear, and variables are coupled. For example, the texture evolution of a milk gel, such as cream cheese during processing, is a non-linear and indirect result of the structure evolution of the gel at a nanoscopic scale. 4. The food industry operates with a diverse set of products and processes and has requirements in terms of the portability and adaptability of the systems developed. 5. Little data is available in traditional manufacturing plants producing, for example, sausage or cheese, and this situation is mirrored throughout the food industry. Furthermore, even when databases do exist, it is not always possible to use them for controlling food product quality. In this context, despite the fact that the design of standards and reliable procedures for controlling product quality is a major objective for the food industry, automation is limited: • Few sensors are available to carry out the measurements needed like sensory texture or odor. Although new sensors such as artificial noses have been developed (Ruan and Zeng, 2004), the road is long, difficult, and unfeasible for SMEs. • For several processes, it is difficult to establish models that are sufficiently representative of the phenomena involved, even for control purposes. • Classical automated approaches remain limited, for the reasons mentioned above. At present, many production processes are largely reliant on the skill and experience of the operator, something that no system will be capable of replacing in the foreseeable future (Perrot et al., 2006). Consequently, in practice, operators often play an important role and cooperate with automation so as to: (1) make on-line evaluations of the sensory properties of the product, and/or (2) adjust the on-line process accordingly. Moreover, experienced operators make macroscopic interpretations of physicochemical phenomena occurring during processing, which can act in synergy with classical engineering knowledge on the process. Integrating operator and expert skill in a control framework is a relevant direction, especially for traditional processes. Nevertheless, it leads to designing mathematical tools that have to integrate: (1) reasoning based on the use of linguistic symbols such as ‘over-coated,’ ‘good color,’ etc., which are expressed not on a numerical scale but on a discontinuous graduated scale referring to an evaluation of a deviation in comparison to a setpoint; (2) uncertainty on these symbols, which is translated after fusion in a specific action; and (3) an action that is the result of an implicit or explicit interpolation between two specific states recorded by the operator over time. Decisions related to the management of food processes increasingly rely on mathematical models that (1) represent the available state-of-the-art knowledge about the phenomena involved, and (2) are able to simulate the different transient and

© Woodhead Publishing Limited, 2013

202 Robotics and automation in the food industry equilibrium states over time. However, due to time limits, financial constraints, and scientific and technological obstacles, information for model parameters is often incomplete and imprecise. This leads to uncertainty, which needs to be accounted for in the decision-making process. Integrating the uncertainties inherent to real levels of knowledge available on the food processes is a crucial step in the development of process representation models. Given that experts have to contend with a lack of useful data for building such models, they may sometimes opt to make their own estimates of poorly-known quantities given in confidence interval format by applying their own experience and intuition. Available knowledge about food processing may display (1) randomness (also known as ‘stochastic uncertainty’) due to the natural variability of observations, or (2) imprecision (also known as ‘epistemic uncertainty’) due to a lack of information (Ferson and Ginzburg, 1996). Randomness and imprecision may coexist, especially in a background of several heterogeneous knowledge sources, such as statistical data and expert opinions. As a rule, partial ignorance is represented by a uniform distribution based on what is known as the ‘indifference principle’ (all that is equally possible is equally probable). This choice, even though guided by principles, remains highly debatable, since the analyst is bringing further information (equiprobability) to knowledge that was not sourced from experts or data and does not, therefore, present any cognitive value. In practice, the choice is dictated by the fact that the only viable option is a classical probabilistic model such as uniform, normal, Beta or log-normal probability distributions for example, which can be argued (Walley, 1991) incapable of accounting for epistemological uncertainty and imprecision in general. While information regarding variability is best conveyed using probability distributions, information tied to imprecision may be more accurately conveyed using families of probability distributions encoded by possibility distributions (also known as ‘fuzzy intervals,’ see Section 9.2.2, second subheading, for details) (Dubois et al., 2000). Fuzzy sets and possibility theories were introduced by Zadeh in 1965 and further in 1978 as an extension of the set theory by replacing the characteristic function of a set by a membership function whose values range from 0 to 1 (Zadeh, 1965, 1978). Possibility theory equips fuzzy sets with the full settings required for a sound comparison with probability, and is relevant for representing consonant imprecise knowledge, such as confidence intervals given by experts pertaining to ill-known quantities. A possibility distribution can model imprecise information regarding a fixed unknown parameter, but can also serve as an approximate representation of incomplete observations of a random variable. The core concept is the possibility distribution describing the more or less plausible values of some uncertain variable. Possibility theory is now a wide field of study that, over the last 20 years, has seen the development of a range of tools. Applied to food engineering, it has been considered as pertinent by several authors for different applications (see Perrot et al., 2006), especially for taking into account the reasoning process, expressed in linguistic terms, of operators and experts where its focuses include: (1) representation of the descriptive sensory evaluation performed by a quality team, an operator, or a consumer; (2) indirect measurement of the properties of a food product, such as the modeling of the expert sensory sausage quality

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 203 evaluation; (3) diagnosis, supervision, and control of food quality, such as the control of processes, for example roasting peanuts, the building of decision-support systems in browning processes, or cheese ripening processes. Examples will be described in this chapter.

9.2

Principles of intelligent control systems using fuzzy logic

Although Boolean logic found widespread applications in the twentieth century (from binary logic to logic integrated circuits), it remains underequipped to model or employ human reasoning in situations of uncertainty or imprecision, since it only deals in binary information that cannot translate the symbol language or the complexity of our perceptions (Dubois et al., 1980). Mathematical tools, such as fuzzy logic and possibility theory, appear better adapted to dealing with or employing human reasoning, alongside other candidate approaches such as qualitative physics (Price et al., 2006). In traditional logic, the ‘universe of discourse’ (U) is formed of elements (u) characterized by total membership (characteristic function: μE(u), equal to 1) or total exclusion (characteristic function: μE(u), equal to 0) of a given set E. This leads to strict between-set transitions imposing a framework that can prove too brittle to capture the complexity of different event processes (Dubois et al., 1999). The upshot is that the quality of a food product cannot be described exclusively through strict transitions between the crisp qualifiers ‘good’ or ‘bad’, but instead requires an infinite series of potentially vague grades to be properly defined and, ultimately, controlled. Fuzzy logic theory and possibility theory take a different stance grounded in progressive continuous between-set transitions. The membership function in fuzzy logic is a good illustration of these transitions, as it extends to the conventional characteristic function. Any element u found in the universe of discourse U can be characterized by its degree of membership of a set E (μE(u)), valued in the real continuous interval [0,1]. A fuzzy set E in the universe of discourse U can be defined by:

{(u,µ

( u )) µE :U → [ ] E

E

u

E

}

[9.1]

μE is thus the membership function of set E. It represents the set of membership degrees (μE(u)) of a variable u mapped to fuzzy set E. To illustrate, a membership function can be used to describe the relationship between the numerical variable moistness and the linguistic tag employed by an operator for a biscuit that is ‘too moist’ (Fig. 9.1). In practice, biscuits of measured moistness (×) lower than 3.5 g/100 g of dry matter (gDM) are not considered moist, whereas biscuits with a moisture content

© Woodhead Publishing Limited, 2013

204 Robotics and automation in the food industry µE

‘Moisture too high’

1 0.5

3.5

4

4.5

Moisture content (g/100 gDM)

Fig. 9.1 Example of a fuzzy membership function applied to qualify the moisture of a biscuit.

ranging from 3.5 to 4.5 g/100 gDM are moist to an increasing degree that is given by the membership function. Thus, a biscuit of moisture 4 g/100 gDM will be considered ‘too moist’ with a degree (μ ‘‘too moist’’ (4)) of 0.5. Biscuits presenting a moisture greater than 4.5 g/100 gDM will be considered totally moist with a degree of membership to linguistic tag ‘‘too moist’’ of 1. ⎧0 ⎪ ⎪ x a1 ⎪⎪ a2 a1 µ (x) = ⎨ ⎪ a3 x ⎪ a3 a2 ⎪ ⎪⎩0

(x

a

)

(a

x

a

)

(a

x

a

)

(a

x)

[9.2]

Membership functions can be expressed through various representations. The representations most widely used are triangular (Equation [9.2]) for a given triplet series a1, a2, a3, graphed in Fig. 9.2, trapezoidal, or sigmoidal. Certain applications (including command controls) also employ what is termed a fuzzy singleton, where μ(x) is 1 for a given x value and 0 in all other points. A fuzzy set can be characterized by its boundary parameters – a1, a2, a3, for example, characterizing a triangular function. This means a range of influences and noises can be integrated by making the right adjustments to the set characteristics – for example, the membership function can be made to integrate a degree of sensitivity or precision, that is, trueness and accuracy, of the measurement. Based on this key definition of fuzzy subsets and fuzzy membership functions, different tools can be developed for the aims of the control system. The following sections will explore these tools further. This chapter will develop two lines of approach. The first addresses the question ‘how can fuzzy logic theory be used to build a control system for food

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 205 µ 1

a1

a2

a3

x

Fig. 9.2 Parametric representation of a membership function.

engineering?’ while the second addresses the question ‘how can probability and possibility theory be used to represent and jointly propagate uncertainties relative to the knowledge about the processes managed?’ The aim is to estimate the uncertainty in model results and highlight that a joint treatment of uncertainties can be useful for controlling the process, even with few instruments and/or little information available.

9.2.1 Building a control system using fuzzy logic Different steps are necessary in order to build an intelligent control system using fuzzy logic (Fig. 9.3). A first level, dubbed ‘adapted systemic analysis’ is a key issue in fuzzy intelligent control systems. It is a knowledge extraction step, carried out based on a systemic analysis of the strategies implemented by experts to resolve a process control problem. A second level, dubbed mathematical knowledge formalization,’ represents the mathematical formalization of the human knowledge handled in the first step. The result reached is an algorithm that mimics the expert tackling a process, that is, a numeric model of the expert strategies for solving control problems. It is built according to the goal of the study: indirect measurement, state diagnosis, or process control. The strategies employed are highly dependent on whether the process type involved is a batch process or a continuous process. Knowledge is then organized in space or in time. The third and last level, dubbed ‘‘validation tools,’’ represents selected ways of validating these approaches and processing symbolic data. Adapted systemic analysis (level 1) The aim is to define the limits of the system to be represented and the hypothesis, to define and formalize in a unified way the variables of interest and the type of interactions between the process studied and the variables manipulated, and to identify and formalize in a unified way the strategies experts employ to solve the control problem. It integrates the non-linear and coupled links in certain fixed organized clusters. Figure 9.4 depicts the methodology for handling expert knowledge. After interviewing the selected experts (as a rule, there will not be too many), a first step is the formalization of the measurements and the associated

© Woodhead Publishing Limited, 2013

206 Robotics and automation in the food industry A case study Adapted systemic analysis

Level 1

Mathematical knowledge formalization Indirect measurement Implicit on data basis

State diagnosis

Explicit expertise

Batch process

Algorithm

Level 2 Control or decision help

Continuous process

Algorithm

Validation tools

Algorithm

Level 3

Qualitative with the experts Quantitative on a data basis Validation at line Ready to use tool

Fig. 9.3 Steps to build a control system using fuzzy logic.

Interviews of the selected experts

If several experts, are they in accordance?

No

Confrontation

Yes

First symbolic algorithm

Formalization of the measurements

Yes Yes

Validation of the measurements

Formalization and structuring of the links

Oral validation

No

No

Fig. 9.4 Methodology for expert knowledge handling.

measurement granularity. If it is validated, then the second step, formalizing and structuring the clusters experts have in mind is carried out, and should explain the non-linear links between the inputs and outputs.

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 207 Table 9.1 Type of formats for the variables of interest Variables Variable Sensor measure Sensory measurement on a structured scale – sensory indicators (Curt, 2001) Format Numeric Numeric discontinuous continuous Example Luminance Number of spots for a browning (camera) process (Ioannou et al., 2004)

Non-existent Small Medium Large Very large

Sensory measurement on an ordinated linguistic scale Linguistic Color of a cheese: line scale from 1 to 6 with three anchors: ivory, yellow and brown (Perrot et al., 2004)

The groundwork to this step involves defining and choosing experts. The experts chosen will largely shape (1) the results obtained, (2) the proportion of empiricism versus knowledge injected in the models, and (3) the generalizability of the models built, whose properties are themselves tied to ‘(2)’. Although expertise has been extensively analyzed in psychology and in artificial intelligence, there is still no operational and universally-accepted definition of expertise per se (Shanteau, 1992). In this study, experts were chosen on the basis of consistent practice, high performance levels and peer recognition. This approach of modeling clusters established by experts on a given process control problem made it possible to handle not just numerical-format variables but also different types of symbolic-format variables expressed by the experts. Table 9.1 recaps these format types. Mathematical knowledge formalization (level 2) We have chosen to differentiate the mathematical tools available according to the type of study undertaken: indirect measurement, state diagnosis or decision help systems. Different tools adapted from fuzzy set theory can be implemented depending on the type of knowledge available. With no expert knowledge available, applicable concepts include supervised clustering, such as the fuzzy k-nearest neighbor method (Keller et al., 1985) or the so-called fuzzy multicomponent membership functions method, which generalizes the membership functions concept given by Benoit and Foulloy (2003). The principle is an approach able to automatically make the link between expert sensory evaluations and instrumental measurements. With explicit expert knowledge available, the concept of fuzzy symbolic sensor (Mauris et al., 1994), based on the fuzzy membership function concept presented above, can be developed. Fuzzy techniques are used to define a language, such as a relation between a set of words (L) and a numerical set (N). This relation is

© Woodhead Publishing Limited, 2013

208 Robotics and automation in the food industry

IS3: Humidity

Low

Middle

High

3 4 (g/100 gDM)

5

1

0

1

2

6

Fig. 9.5 Example of a fuzzy meaning describing the moisture content of a cheese.

characterized by a membership function, µRF, which represents the degree or the strength of the link between symbols and numbers. This fuzzy relation can be described by two projections that take their values from the set of fuzzy numerical subsets F(N) and from the set of the fuzzy symbolic subsets F(L): Meanings and Descriptions. Fuzzy meaning represents the symbols employed by operators in word format. To illustrate, fuzzy meaning can be used, after analysis of expert handling, to represent a projection in numeric space of the symbolic way operators qualify a cheese in terms of product moisture content: ‘low, middle, high’ (Fig. 9.5). It means for example, that between 0 and 1.5 g/100 gDM we are sure that the product is low in term of expert evaluation. Between 1.5 and 2.5 g/100 gDM, the product is between low and middle with a given linear law. Fuzzy description is a simple way of describing a measurement with words. For example, in Fig. 9.5 a moisture content of 4 g/100 gDM can be described as µD(4) (‘dry’) = 0; µD(4) (‘normal’) = 0.5; µD(4) (‘moist’) = 0.5. We can also notate the concept as D(4) = 0/dry + 0.5/normal + 0.5/moist. To complete this tool, it is necessary to build approaches able to aggregate several symbols together. This is the case in a control framework where decisions for actions are based on a diagnosis of the combined deviations of the different dimensions of product quality and the state of the process. For example, the atline operator needs four main sensory characteristics to evaluate the trajectory of a cheese during ripening and to anticipate his control actions (Perrot et al., 2004): color, coat, consistency, moisture content. The fuzzy function used for merging is created on the basis of an expert explanation of the logical links between symbols. This system, tested in the industry, was used to predict the quality evolution of the cheese production with a good accuracy and robustness by comparison to an expert evaluation (around 85% of similar answer). This system is used as a support decision system when the expert is not available. These between-symbol links incorporate the global aim of the merging. The inference is made by Zadeh’s compositional rule of inference (Zadeh, 1965), applied to the fuzzy symbolic descriptions of the inputs, as explained below. Process control or decision-aid systems For process control or decision-aid system purposes, the output of the fuzzy controller should generally be a numerical output, the setpoint of a physical regulator.

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 209 The most widely used fuzzy controller is a constant-type Takagi–Sugeno fuzzy controller (Takagi and Sugeno, 1985), where the output gi,j of each rule is assimilated to a constant (e.g. a temperature setpoint of a ripening chamber for a given rule). Numerical outputs are directly calculated using the result of Equation [9.3], membership grade µExF(Ai, Bj) for each rule l annotate αl, and the associated constant output annotated Plk, for each rule 1 and each constant K. Dlk =

∑αP ∑α * l

l

l

lk

[9.3]

l

Validation tools (level 3) The third and last level is dedicated to validating the tools that could viably be used for the algorithms developed (see Fig. 9.3). A first step involves working with experts to secure the qualitative validation. This first step requires an expert–observer feedback exchange to verify that the knowledge expressed by the expert matches the knowledge implemented in the algorithm. In this step, the only possible validation is qualitative. Quantitative validation has to be achieved on test points of a custom-compiled database. This quantitative validation is the minimum baseline validation required to be able to consider the tool as compatible with the objective targeted, which is either (1) to achieve an indirect measure or state diagnosis matching the operator’s responses, or (2) to achieve a level of process control that is at least as good as the operator’s and just as robust, reliable, and stable. In case (1), the cogency of the approach is checked based on test points that give the best possible representativity of the range of different system states. The model’s output is compared with the operator’s response. The number of compatible responses (C) is given by Equation [9.4], where Rm is the model response corresponding to the input measurements logged in the validation database, Ro is the response of the corresponding expert operator, and d is incremented by 1 when D is less than or equal to the tolerance threshold (St). A percent compatibility value is produced by calculating the ratio of number of compatible responses to number of products making up the validation database. n

C

∑

d

i =1,2,3

d

if

Di ≤ St with Di = Rm − Ro

d

if

Di > St

[9.4]

A key variable in this step is the tolerance threshold. It is generally specified in tandem with the experts for sensory measurements (Ioannou et al., 2002). In cases where the outputs are expressed in the form of linguistic values, the validation indicators are based on equations developed in Bouchon-Meunier et al. (1996).

© Woodhead Publishing Limited, 2013

210 Robotics and automation in the food industry If the model is validated based on recorded data, the validation phase continues under real conditions; otherwise, the only solution is to go back to the model design drawing board: either the knowledge was badly formulated or the interviewer missed some important information. Every step in this phase will need to be repeated start-to-finish. For validation of a process control, the criteria used are the same as those employed under classic automated processing. Underpinning this step is the need to compile a salient database. This database needs to be as representative as possible of the full set of potential system states and the most common system disturbances. If the variables integrated in the database are assessed by experts, it is important to control the quality of these measurements. This can be achieved using methodologies adapted from standard NF X 07-001 to test the reproducibility and repeatability of human measurements expressed on a symbolic scale (numeric discontinuous or linguistic), thus generalizing earlier works led by Curt et al. (2004) on sensory indicators.

9.2.2

Representation and propagation of uncertain and imprecise knowledge Let T be a mathematical model representing a phenomenon that occurs during food processing whose arguments are represented by the appropriate formalism (probability–possibility) according to the character of available knowledge (variability, imprecision, incompleteness, etc.). Knowledge representation by means of probability distributions Faced with substantial data, it may be possible to identify a probability distribution describing the variability. For example, Fig. 9.6 shows the distribution of the cumulated relative frequencies of heavy rain (describing the probability that the real value of heavy rain is lower than a certain threshold) estimated from 30 years of data collected from a weather station. This distribution shows, for example, that there is 95% of chance that the heavy rain does not exceed 300 mm but exceeds 150 mm. In a risk assessment setting, the risk manager could use the probability law fitted from data (see the example in Fig. 9.6 showing the cumulative normal distribution with mean 222.2 mm and standard deviation 39.5 mm fitted from raw frequencies). Probability distributions can be also used for theoretical reasons, or where there is a priori knowledge on the variability involved. Knowledge representation by means of possibility distributions Faced with information along the lines of ‘I am sure that quantity x lies within an interval [a, b] but values located within the interval [c, d] (included inside [a, b]) are more likely,’ it is coherent to propose a trapezoidal possibility distribution (formally equivalent to a fuzzy interval – see Fig. 9.7). For example, the possibility distribution π defined in Fig. 9.7 means that the most likely (risky but informative) heat mass transfer values are located within the interval [3, 3.2] (referred to as the ‘core’ of π) with no preference inside this interval. We find the conservative

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 211 1 0.9 Cumulative frequencies

0.8

Data Probability law from data

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 100

150

250 200 Rough rain by year (mm)

300

350

Fig. 9.6 Cumulated relative frequencies and probability law from data about annual heavy rain. Rain history over 30 years.

but not very informative values within [2.86, 3.33], also known as the ‘support’ of π. Numerical possibility distribution may also be viewed as a nested set of confidence intervals (information able to be given partially by experts), known as cuts of π noted πθ = {e/π(e) ≥ θ}. The degree of certainty that interval πθ contains the real value of heat mass transfer is 1 − θ, meaning that the probability that the heat mass transfer lies within πθ is greater than 1 − θ. Possibility distribution enables the representation of sets of probabilities (Dubois et al., 2000) limited by an upper probability bound known as the possibility measure (see the upper cumulative probability bound on Fig. 9.7) and a lower probability bound known as the necessity measure (see the lower cumulative probability bound on Fig. 9.7). Possibility distribution thus makes it possible to represent incomplete probabilistic knowledge and obtain a bracketing of probabilities.

9.3

Current applications in the food industry

Fuzzy logic has already been applied in the food industry for quality estimation, state diagnosis and process control. It will be developed in this section.

9.3.1

Indirect measurement of symbolic variables employed by experts for estimation of food product quality Indirect measurement can be used alone for inspection purposes, as in Chao et al. (1999) for post-mortem poultry classification of color images of viscera

© Woodhead Publishing Limited, 2013

212 Robotics and automation in the food industry Possibility distribution π

Upper probability bound

Lower probability bound

1.0

0

0.8

0.2

0.6

πθ: θ – cut of π

0.4

0.4

P(h∈πθ) ≥ 1– θ

0.6

0.2

Values considered not possible

Values considered certain

Values considered not possible

Certainty

Likelihood, θ

Values considered most likely

0.8 1.0

0 2.8

3 3.2 Heat transfer coefficient h (W.m–2.K –1 )

3.4

Fig. 9.7 Example of a possibility distribution induced from a knowledge expert on a heat transfer coefficient used as a parameter in a cheese ripening model. (Source: Baudrit et al., 2009.)

to distinguish air sacculitis livers from normal livers, or as in Nielsen and Paul (1997) for tomato quality grading. It can also be coupled with a controller or embedded in a decision-aid system to control food product quality, as in Perrot et al. (1996) to control the quality of biscuits in an industrial tunnel oven. The requirements in terms of measurement precision and reliability will therefore differ. In both cases, it is most often a ‘feature level fusion’ as defined in Valet et al. (2000). The fusion is performed on specific features extracted from raw information that is generally provided by sensors such as cameras, electronic noses, NMR sensors, etc. For example Yea et al. (1994) used three commercial gas sensors to detect four kinds of fragrant smells with a discrimination rate of 99.2%. In the first step, fuzzy reasoning detects the fragrant smells, then in the second step, odor is discriminated by a neural network. De Silva et al. (1996) used PCA-extracted features from acoustic video images to determine the firmness of herring roe. A fuzzy decision-making system tested on 160 samples gave about 84% good discrimination. Sundic et al. (2000) used features extracted from an electronic tongue coupled with an electronic nose to mimic human sensory perception of potato chips and cream. An illustration of such an approach based on the concepts presented in Section 9.2 is developed in the example below applied to the evaluation of the sensory properties of a crusting sausage. It is a sausage defect that appears under specific conditions of drying. The defect of crusting is one of the two important defects that can appear on the product and should be avoided. The evaluation of this defect in the industry is made close to the manufacturing line (Mauris et al., 2000). The degree of sausage crusting evaluated by the operators is then reproduced using fuzzy meanings (Section 9.2.1) based on adapted image processing

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 213 Camera

Processing image

Sausage

Each pixel is associated with a luminance Threshold

LB and LD in pixel

Threshold

Center color in linguistic term

Meaning of center color

Center color in pixel

Meanings of LB and LD

LB and LD in linguistic term

Aggregation by fuzzy rules

Symbolic information

Defuzzification

Crusting degree

Fig. 9.8 Evaluation of the crusting degree of a sausage using the fuzzy meaning concept and an adapted image analysis.

and explicit expert rules about crusting (Fig. 9.8). Crusting is evaluated from the human visual perception of the color of the center of the slice, length of black areas (LB), and length of dark red areas (LD) on sausage slices. Figure 9.9 gives an example of fuzzy meaning for LB. Crusting classes are defined according to expert knowledge by a set of fuzzy symbolic rules linking the terms of black lengths and dark red lengths, and classes of crusting, together with expert rules such as ‘if LB is non-existent and LD is non-existent, crusting degree is nonexistent, class 1.’ The approach was implemented with a very limited number of rules: 16 rules processed in this application. It takes 2 months to handle those rules. The result is symbolic information in the form of a degree of membership to crusting classes 1 to 4. The result for a threshold of tolerance (see Section 9.2.1) bounded in a first approximation by the value 0.5 is 88% for compatible answers tested on two databases. This system is useful as regards to the sausage crusting evaluation that is difficult to achieve for the operators. This tool is used as a decision-support system in the industry to help to achieve the right evaluations. In applications where technical or economic reasons dictate that no sensors are available, the only inputs to the indirect measurement module are human operators. In this case, the symbolic space employed by the operators to evaluate the product is used directly by the system, as in Ioannou et al. (2004) for overall browning appearance during a heating process. This application is representative of situations commonly encountered in the food industry, where it is important to propose mathematical approaches capable of directly coping with the symbolic data handled by the operators. The browning of food products is obtained through heating the product surface, which leads to biochemical reactions called Maillard reactions. In this application, we have developed a diagnosis model to

© Woodhead Publishing Limited, 2013

214 Robotics and automation in the food industry (a)

(b) Membership from 0 to 1

Non-existent

A bit long

Very long

Long

1

0

150

200

500

700

1000

LB

Number of pixels of length of black areas (LB)

Fig. 9.9 Meaning of LB: (a) pixels of the image that represent LB, (b) fuzzy meaning of the variable LB.

help the evaluation of the browning variable. It is an important sensory variable assessed by an operator to control the browning process, which influences consumer choice and preferences. Nevertheless, in the industry, it is difficult to measure and is achieved by the quality team of the factory through a visual quality criterion so-called the browning global appearance (BGA). The diagnosis model consists in aggregating the descriptive variables assessed by the operators into the BGA (V1, percentage of spotted areas; V2, color of the spotted area; V3, color of the product surface excepting the spot areas), see Table 9.2 for V2. Fuzzy rules are in the form, for example, of the descriptive variables being assessed as follows: V3 is beige, V2 is dark brown, and V1 is non-existent with a membership degree to 0.75 and is small with a membership degree of 0.25. In this way, the rules having a membership degree different from 0 are R1 and R2. R1: if V1 is non-existent, V2 is dark brown and V3 is beige then the BGA is Low browning LB. R2: if V1 is small, V2 is dark brown and V3 is beige then the BGA is No browning NB. The diagnosis was validated on 40 products with a percentage of compatibility of 92.5% at a threshold of one half symbol (Fig. 9.10). The selected number of fuzzy membership functions, and consequently the number of symbols chosen to characterize a linguistic variable, is a crucial factor in this approach. It gives an indication of the importance of each variable in the expert’s strategy. It is also linked to the sensitivity of the operator’s measurement on this variable, which is itself tied to its importance, if not to difficulties in

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 215 Table 9.2 Sensory indicator percentage of spotted area (V2) The percentage of spot areas Proportion of spots in comparison with the total area Observe the product area at the output of the browning oven On the product area, observe the areas which are different from the background color of the product Assess the rapport between these areas in comparison with the total area Non-existent, small, medium, large, very large

Name Definition Protocol

Assessment scale References

Non-existent

Small

Medium

Large

Very large

At the output of the browning oven The center of a set of products

Time Location

Source: Ioannou et al., 2004.

Burnt

5

Browning global appearance

Operator BGA Model BGA High 4 browning

Target 3 browning

Low 2 browning

No 1 browning

1

2

3

4

5

6

7 8 9 10 Image numbers

11

12

13

14

15

Fig. 9.10 Comparison between the operator and the diagnosis model for the evaluation of the BGA. (Source: Ioannou et al., 2004.)

performing the measurement. This is what can be termed the ‘granularity’ of the measurement, which is linked to the graduality selected for the fuzzy meanings and descriptions. For example, in the application developed below on the global appearance of browning (BGA). A test was run on the repeatability of the expert measurements for these three variables (Table 9.3), revealing that repeatability can be linked to the number of symbols associated to each of these variables (V1 and V3 were explained by 5 symbols, and V2 by 4 symbols).

© Woodhead Publishing Limited, 2013

216 Robotics and automation in the food industry Table 9.3 Repeatability for the expert measurements involved to predict the global browning appearance during a heating process.

Operator repeatability

Tolerance threshold BGA (%)

V1 (%)

V2 (%)

V3 (%)

0.5 0.25

98 80

92 62

90 83

91.67 82

This graduality induced by the fuzzy meanings is key to managing fuzzy control, as it offers a way to explain part of the tolerance of the system built. Going back to the sausage crusting application, if we double the slope of the line linking the symbol ‘a bit long’ and ‘long’ for LB for a given LD fixed to a membership degree of 1 for the ‘very long’ symbol and 0 elsewhere, it leads to a change of product classification from class 3 to class 4, which means the product is then classified as defective. This graduality could be explicitly set up when adjusting the parameters of implicit classification methods. A result is presented in Perrot et al. (1996) for a problem involving an indirect measurement of cookie color.

9.3.2 State diagnosis and food process control This subject has been the focus of a great deal of scholarship in the ‘fuzzy and food’ field (51% of the total papers for a query on FSTA; Perrot et al. (2006)). Most of these papers are classical applications of the Takagi–Sugeno controller (see section Mathematical knowledge formalization (level 2), such as Alvarez et al. (1999) for controlling isomerized hop-pellet production, Honda et al. (1998) for controlling the sake brewing process, or O’Connor et al. (2002) for controlling the beer brewing process. The authors use two specific approaches to develop the diagnosis, control and supervision modules: (1) ‘data-driven’ approaches, and (2) ‘expert knowledge-driven’ approaches. For ‘data-driven’ approaches (around 30% of papers), the system is identified automatically using a database and an optimization algorithm. For example, Honda et al. (1998) developed a fuzzy neural network to control the temperature of the Ginjo sake mashing process, O’Connor et al. (2002) developed a fuzzy PID to control the brewing process, Guillaume and Charnomordic (2001) optimized a fuzzy rule basis using a genetic algorithm to establish a decision-aid system for the cheese-making process. In this case, the risk is that the models developed will have to conform with a large number of rules (50–100 or more) and thus have a large number of parameters to identify. For example, in Honda et al. (1998), 283 data points are used to identify the model. Handling such large quantities of data can lead to difficulties in setting all the parameters, and is a major limitation for applications used in food processes. Moreover, the tools built can only be used to interpolate but not extrapolate new cases. Nevertheless, in these cases, the added-value brought by using fuzzy logic hinges on (1) a user-friendly operator interface that handles data in a space close to the symbolic space employed

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 217 by the operators; (2) the possibility of automatically adapting the module to new processing conditions (such as a new formula for a biscuit) if a model or a set with enough data to characterize the process become available. For ‘expert-knowledge-driven’ approaches, fuzzy logic is used as a way of providing a mathematical formalization of expert knowledge and embedding it in decision-aid algorithms and controllers. For example, Ioannou et al. (2004) proposed a fuzzy logic approach to control the browning process during industrial heating, Curt et al. (2002) developed five Takagi–Sugeno modules to control sausage quality during ripening. Fuzzy logic is also used for supervisory tasks, as in Acosta-Lazo et al. (2001) for the supervision of a sugar factory, or Goelzera et al. (2009) for decision-support software including a physiological model of yeast activity and fuzzy logic to provide fast and efficient solutions to facilitate winery management on an industrial scale in various winemaking conditions. In this case, the modules are used directly, without database adaptation, and the datasets are only used to validate the approaches. The rule bases are generally compact, featuring only a limited number of rules (around 20 on average). It is very interesting for two reasons: firstly, the building of such a fuzzy expert system is less time-consuming than for classical expert systems building; secondly, it is very easy to correct and adapt from one process to another on the same product. We shall illustrate this chapter using the results developed in Perrot et al. (2000). The principle was to develop a decision-aid system for the baking process. The entire process was non-linear and time-variant (the operating conditions of an oven change after it has been used for extended periods), with strong interactions between variables, such as the impact of the temperature all along an industrial tunnel oven on the moisture content, color and thickness of the biscuit. The operator plays an important role in this process, and the quality of the biscuits baked hinges on the operator’s measurements and actions. Working in this application, we built and validated a closed-loop quality control feedback system on an indirect continuous pilot tunnel oven (Fig. 9.11). Numerical inputs for moisture and thickness are formatted in a linguistic space using the fuzzy meanings described in Section 9.2.1. The space formatted in this way can be aggregated at the same level as the operators’ symbolic description of the color. The second step is to compute the fusion between the symbols expressed by the operators according to the fuzzy formalism. For example: ‘oven too hot in section c of the oven = (moisture normal OR moisture dry) AND color over-cooked’ or ‘oven not hot enough in section c of the oven = (moisture normal OR too much moisture) AND color slightly under-cooked.’ Finally, a constant-type Takagi–Sugeno fuzzy controller is implemented to aid decisions on action on the output variables, that is, temperature setpoints for the air in the different chambers of a continuous oven. This controller is based on the membership grades calculated for system diagnosis. Figure 9.12 illustrates an example of an experiment on an indirect continuous pilot oven, giving the response of the feedback controller using the outputs of the support system and the process state diagnosis. Good regulation performances were achieved for all biscuit quality characteristics during baking, despite the disturbances as

© Woodhead Publishing Limited, 2013

218 Robotics and automation in the food industry Knowledge of the operator on the evaluation and action Color Operator or sensor

Fuzzy symbolic approach Computing meanings and fusion of symbols

Biscuit Moisture and thickness Sensor

Diagnosis on the state of the process Feedback decision of action on the process

Fig. 9.11 The fuzzy symbolic approach applied to biscuit baking.

confirmed after strong disturbance applied to the process at 9000 and 1200 s. The operator detects the quality drift after about 8 min and gives his/her evaluation to the symbolic module. In parallel, the same change of color was checked by the sensor. Finally, the symbolic module aggregates the operator measurement with the instrumental measurements of biscuit moisture content and thickness and establishes a diagnosis (presented Fig. 9.12b) that is used by the automatic control system to adjust its action on the oven. Thus, the three color deviations are expressed in terms of process deviation, that is, ‘oven too hot in section C’ or ‘oven not enough hot in section C,’ with a grading between [0, 100]%. These diagnoses are transmitted to the automatic controller, which then fixes new setpoints to maintain biscuit quality. The same results are reached simultaneously for product moisture content and thickness. Both control accuracy and stability are validated as satisfactory. Generally, knowledge is clustered on the type of process encountered. For batch processes, knowledge clusters are organized by time clusters. For example in Perrot et al. (2004), where the authors deal with an application of cheese ripening, knowledge is organized in four stages of ripening linked to the ripening time. For continuous processes, such as the example presented above, knowledge clusters are organized in space. What is interesting to underline is the coherence of the experts’ knowledge structuring expressed under fuzzy rules, with the scientific knowledge that is sometimes available. For example, in the biscuit baking application described above, there were mechanistic models available (SavoyeBarbotteau et al., 1992). In these mechanistic models, the phenomena were guided by the temperature reached at the biscuit surface. Below 100°C, drying velocity increased, whereas above 100°C it remained stable, and water activity equaled 1. For the continuous oven considered in this study, this key temperature

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 219 (a) Classes of color scaled in [0–5]

6 5

Off-line evaluation of color by a sensor

Classes of color and diagnosis [0–100%] scaled in [0–5]

Over cooked

4

Well cooked

3 2

Under cooked

1 0 780

(b)

Operator evaluation (input of the automatic controller)

880

6

980

1080 Time (s ⫻ 10)

1180

1280

Diagnosis – oven too hot

5 4 3

Diagnosis – oven not hot enough Color

2 1 0 780

880

980

1080

1180

1280

Time (s ⫻ 10)

Fig. 9.12 Feedback control (a) and diagnosis of the color of biscuits (b) using the fuzzy symbolic module during an experiment for the cooking process.

was reached in the second third of the oven. The knowledge clusters gained after expert handling were coherent with this, and organized, for moisture diagnosis, in the so-called ‘section C’ of the oven which corresponds to the second third of the oven. For the kinetics of product coloration, both knowledge representations led to action in the last third of the oven. It has induced a more regular control that has led to avoiding sensory defects on the biscuits.

© Woodhead Publishing Limited, 2013

220 Robotics and automation in the food industry 9.3.3

Joint processing of imprecision and variability in the modeling of cheese mass loss during ripening The dynamics of cheese mass loss is a key factor in the ripening process, with consequences for productivity and a risk that the resulting product may be dropped in status. These dynamics are dependent on climatic conditions in the ripening chambers where ventilation is used to evacuate heat and humidity generated by cheeses. In an industrial context, it is not feasible to install sensors at every point in the ripening chamber in order to control atmospheric conditions. Due to confidentiality issues, the literature contains few studies on the interaction between climatic conditions and airflow. Heat and mass transfers have been extensively studied in relation to cooking and drying processes, but few data have been published in the field of cheese ripening, and transfer coefficients between cheeses and the atmosphere are not precisely described. The aim here is to take into account the imprecise and incomplete knowledge of ripening chamber characteristics in the cheese mass loss model (Baudrit et al., 2009). Each input variable and model parameter is represented by means of probability and possibility distributions according to the nature of the available knowledge, and we implemented the joint propagation scheme for each time-step. Figure 9.13 (left) shows the lower and upper cumulative probability bounds at day 14 before wrapping for an initial mass of 0.33 kg resulting from uncertainty propagation. The gap between these two limits is due primarily to the imprecise nature of available information. We can thus summarize the total uncertainty on cheese weight before the wrapping step, using the interval [0.260, 0.284]. That means we are 95% sure that cheeses will exceed 0.26 kg at day 14. Figure 9.13 (right) presents the evolutions of uncertainty margins of 5% and 95% percentiles for the mass loss at each timestep during the two ripening processes. On the one hand, we are 95% certain that the mass loss of cheeses throughout the ripening room will not exceed 70 g for the first trial at day 14 of ripening. On the other hand, we are 95% sure that cheeses have lost at least 46 g before they are wrapped. On the basis of a model calibrated to pilot scale, our approach could provide key information to improve the control of cheese mass losses in industrial conditions.

9.4

Advances in research and future trends

As we have seen, fuzzy logic could have a promising future in food applications to (1) capture and formalize the descriptive sensory evaluation performed by a quality team, an operator or a consumer, (2) deliver a state diagnosis, and (3) control or aid decisions in food engineering. The state of the art of knowledge available for modeling food processes induces uncertainty on certain phenomena and, consequently, on certain model input variables and parameters. Possibility distributions are well-geared to representations of expert knowledge as the experts can be expected to be consistent

© Woodhead Publishing Limited, 2013

(a)

Probability (cheese mass≤ threshold) at day 14

Intelligent quality control systems based on fuzzy logic 221 1 Imprecision + dependence

0.9

Variability + imprecision + dependence

0.8 0.7 0.6 0.5 0.4

Measurement at day 14 Lower probability Upper probability Lower 5% percentile Upper 95% percentile

0.3 0.2 0.1 0 0.255

0.26

0.265

0.27

0.275

0.28

0.285

0.29

Cheese mass (kg)

(b)

0.33 Upper 95% percentile Lower 5% percentile Online measurement

Cheese mass (kg)

0.32 0.31

46g

0.3 70g

0.29 0.28 0.27 0.26

Gap due to variability + imprecision + dependencies

0

2

4

6

8

10

12

14

Time (days)

Fig. 9.13 (a) Lower and upper probabilities that cheese mass is lower than a certain weight at day 14. (b) Uncertainty margins of the 5th and 95th percentiles pertaining to the cheese mass loss through the ripening process from an initial mass of 0.33 kg.

with their own decisions: the interval of values that they consider most likely is necessarily included in the interval outside which they consider values are not possible. Integrating partial ignorance in the mathematical model established at pilot scale can help transpose this knowledge to an industrial scale. Propagating imprecision can help improve the control process. To round up, it will become possible to analyze the contribution of imprecision and/or incompleteness to modeled response-outputs in order to elucidate whether an ambiguous response is due to a lack of information or to a source of unpredictable variability, and thus determine the key variables and/or effect events where further insights are critically needed. Several questions arise when building a control system using expert knowledge translated into fuzzy functions. Question 1 centers on expert knowledge and the difficulties in handling it, even though cognitive science has developed some useful tools for this purposes. One point to underline is the importance of choosing ‘good’ experts, which will

© Woodhead Publishing Limited, 2013

222 Robotics and automation in the food industry largely shape the results obtained. More in-depth work with the cognitive science community could prove valuable in progressing this question. Question 2 raises a series of questions centered on the choice of graduality for fuzzy functions and fuzzy meanings. How should the right graduality be chosen according to objectives targeted? How should this graduality be validated? What is the best comparative method according to objective targeted? How is it possible to validate model results that go beyond human perception of graduality but are still relevant when used for process control purposes? Which mathematical tools need to be deployed so as to implement a theoretical approach on a given indirect measurements model in order to assess the graduality achieved on the output variable according to curves established on input variables? Question 3 is centered on the optimality of the models built and their capacity for optimization through structural change without destroying the model-embedded semantics. Linked to this question is the issue of the generalizability of the models built. Question 4 concerns the dynamic properties of the models based on the expertise handled. The cognitive nature of the experts’ resources (Wickens, 1991) makes it extremely difficult to extract mental representations of the dynamics of a system. This is a significant bottleneck for the purposes of controlling complex systems. Nevertheless, it is essential to understand the dynamics of the complex systems involved in food processes. The development structure, as demonstrated by Datta (2008), is not just a function of temperature and moisture but also a function of their history, where the complex physical structure that develops changes attributes such as porosity and transport properties. It is true for all the structured food produced by the industry like cream cheese, milk gels, etc. One of the major interests stems from the development of confident and reliable models integrating different knowledge sources and formats. The guiding principle is to deal with the different pieces of the puzzle of knowledge represented under different formalisms, that is, data, models, expertise. In this sense, the ability to model expert knowledge using fuzzy functions is valuable progress and should be coupled to other forms of modeling. To conclude, the benefits that fuzzy logic can provide to the food industry are: • An easy way to formalize and capitalize on the expert knowledge that represents the memory of the factory. • Improvements in the management and traceability of the quality of production with systems that help the operators in their measurement or control tasks. • Management of the uncertainty with tools that can help to cope with this uncertainty and propose process management improvements.

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 223

9.5

References

Acosta-Lazo, G. G., Alonso-Gonzales, C. J. and Pulido-Junquera, B. (2001). ‘Knowledge based diagnosis of a sugar process with teknolid.’ International Sugar Journal 103(1225), 44–51. Alvarez, E., Cancela, M. A., Correa, J. M., Navaza, J. M. and Riverol, C. (1999). ‘Fuzzy logic control for the isomerized hop pellets production.’ Journal of Food Engineering 39, 145–150. Baudrit, C. Hélias, A. Perrot, N. (2009). ‘Joint treatment of imprecision and variability in food engineering: application to cheese mass loss during ripening.’ Journal of Food Engineering, 93(3), 284–292. Benoit, E. and Foulloy, L. (2003). ‘Towards fuzzy nominal scales.’ Measurement 34(1), 49–55. Bouchon-Meunier, B., Rifqi, M. and Bothorel, S. (1996). ‘Towards general measures of comparison of objects.’ Fuzzy Sets and Systems 84, 143–153. Chao, K., Chen, Y. R., Early, H. and Park, B. (1999). ‘Color image classification systems for poultry viscera inspection.’ Applied Engineering in Agriculture 15(4), 363–369. Curt, C., Hossenlopp, J., Perrot, N. and Trystram, G. (2002). ‘Dry sausage ripening control. Integration of sensory related properties.’ Food Control 13(3), 151–159. Curt, C., Perrot, N., Allais, I., Agioux, L., Ioannou, I., Edoura-Gaena, B., Trystram, G. and Hossenlopp, J. (2004). ‘Formalization of at-line human evaluations to monitor product changes during processing: the concept of sensory indicator,’ in D. Ruan and X. Zeng, Intelligent Sensory Evaluation, Springer Verlag, 157–174. Datta, A. K. (2008). ‘Status of physics-based models in the design of food products, processes, and equipment.’ Comprehensive Reviews in Food Science and Food Safety 7(1), 121–129. De Silva, C. W., Gamage, L. B. and Gosine, R. G. (1996). ‘An intelligent firmness sensor for an automated herring roe grader.’ Intelligent Automation and Soft Computing 1(1), 99–114. Dubois,, D. and Prade,, H. (1980). Fuzzy Sets and Systems: Theory and Applications, Academic Press, Inc Dubois, D., Foulloy, L., Galichet, S. and Prade, H. (1999). ‘Performing approximate reasoning with words,’ in Computing with Words in Information/Intelligent Systems 1, Springer Verlag, 24–29. Dubois, D., Nguyen, H.T. and Prade, H. (2000). ‘Possibility theory, probability and fuzzy sets: misunderstandings, bridges and gaps.’ Fundamentals of Fuzzy Sets, Dubois, D. Prade, H., Eds: Kluwer, Boston, Mass, 343–438. FERSON, S. and GINZBURG, L. R. (1996) ‘Different methods are needed to propagate ignorance and variability.’ Reliability Engineering and Systems Safety 54, 133–144. Goelzera, A., Charnomordica, B., Colombiéa, S., Fromiona, V. and Sablayrolles, J. M. (2009). ‘Simulation and optimization software for alcoholic fermentation in winemaking conditions.’ Food Control 20(7), 635–642. Guillaume, S. and Charnomordic, B. (2001). ‘Knowledge discovery for control purposes in food industry databases.’ Fuzzy Sets and Systems 122(3), 487–497. Honda, H., Hanai, T., Katayama, A., Tohyama, H. and Kobayashi, T. (1998). ‘Temperature control of Ginjo sake mashing process by automatic fuzzy modeling using fuzzy neural networks.’ Journal of Fermentation and Bioengineering 85(1), 107–112. Ioannou, I., Perrot, N., Curt, C., Mauris, G. and Trystram, G. (2004). ‘Development of a control system using the fuzzy set theory applied to a browning process – a fuzzy symbolic approach for the measurement of product browning: Development of a diagnosis model – part I.’ Journal of Food Engineering 64, 497–506.

© Woodhead Publishing Limited, 2013

224 Robotics and automation in the food industry Ioannou, I., Perrot, N., Hossenlopp, J., Mauris, G. and Trystram, G. (2002). ‘The fuzzy set theory: A helpful tool for the estimation of sensory properties of crusting sausage appearance by a single expert.’ Food Quality and Preference 13(7–8), 589–595. Ioannou, I., Perrot, N., Mauris, G. and Trystram, G. (2004). ‘Development of a control system using the fuzzy set theory applied to a browning process – towards a control system of the browning process combining a diagnosis model and a decision model – part II.’ Journal of Food Engineering 64, 507–514. Keller, J., Gray, M. and Givens, J. (1985). ‘A fuzzy k-nearest neighbor algorithm.’ IEEE Transactions on Systems, Man and Cybernetics 15(4), 580–585. Mauris, G., Benoit, E. and Foulloy, L. (1994). ‘Fuzzy symbolic sensors – from concept to applications.’ Measurement 12, 357–384. Mauris, G., Perrot, N., Lambert, P. and Philippe, J. (2000). ‘Fuzzy techniques evaluate sausage quality.’ IEEE Instrumentation and Measurement 3(4), 14–17. Mittal, G.S. (1997). Computerized Control Systems in the Food Industry. Marcel Dekker, New York, 597 pp. Nielsen, H. M. and Paul, W. (1997). ‘Modelling image processing parameters and consumer aspects for tomato quality grading’. IFAC Mathematical and Control Applications in Agriculture and Horticulture, Hannover, Germany. O’Connor, B., Riverol, C., Kelleher, P., Plant, N., Bevan, R., Hinchy, E. and D’Arcy, J. (2002). ‘Integration of fuzzy logic based control procedures in brewing.’ Food Control 13, 23–31. Perrot, N., Agioux, L., Ioannou, I., Mauris, G., Corrieu, G. and Trystram, G. (2004). ‘Decision support system design using the operator skill to control cheese ripening – application of the fuzzy symbolic approach.’ Journal of Food Engineering 64, 321–333. Perrot, N., Ioannou, I., Allais, I., Curt, C., Hossenlopp, J. and Trystram, G. (2006). ‘Fuzzy concepts applied to food product quality control: A review.’ Fuzzy Sets and Systems 157, 1145–1154. Perrot, N., Trystram, G., Guely, F., Chevrie, F., Schoesetters, N. and Dugre, E. (2000). ‘Feed-back quality control in the baking industry using fuzzy sets.’ Journal of Food Process Engineering 23, 249–279. Perrot, N., Trystram, G., LeGennec, D. and Guely, F. (1996). ‘Sensor fusion for real time quality evaluation of biscuit during baking. Comparison between Bayesian and Fuzzy approaches.’ Journal of Food Engineering 29, 301–315. Price, C. J., Trave-Massuyes, L., Milne, R., Ironi, L., Forbus, K., Bredeweg, B., Lee, M. H., Struss, P., Snooke, N., Lucas, P., Cavazza, M. and Coghill, G. M. (2006). ‘Qualitative futures.’ Knowledge Engineering Review 21(4), 317–334. Ruan, D. and Zeng, X. (2004). Intelligent sensory evaluation. In D. Ruan and X. Zeng, Eds., Methodologies and Applications. Springer, Berlin, 443 pp. Savoye-Barbotteau, I., Trystram, G., Duquenoy, A. and Brunet, P. (1992). ‘Heat and mass transfer dynamic modelling of an indirect biscuit baking tunnel oven.’ Journal of Food Engineering 16, 173–196. Shanteau, J. (1992). ‘Competence in experts: the role of task characteristics.’ Organizational Behavior and Human Decision Processes 53, 252–266. Sundic, T., Marco, S., Samitier, J. and Wide, P. (2000). ‘Electronic tongue and electronic nose data fusion in classification with neural networks and fuzzy logic based models’. Proceedings of the 17th IEEE Instrumentation and Measurement Technology Conference. Takagi, T. and Sugeno, M. (1985). ‘Fuzzy identification of systems and its application to modelling and control.’ IEEE Trans. on Systems, Man and Cybernetics 15(1), 116–132. Valet, L., Mauris, G. and Bolon, P. (2000). A statistical overview of recent literature in information fusion. ISIF, Paris.

© Woodhead Publishing Limited, 2013

Intelligent quality control systems based on fuzzy logic 225 Walley, P. (1991). Statistical Reasoning with Imprecise Probabilities. Chapman and Hall, London. Wickens, C. (1991). ‘Processing resources and attention,’ in T. A. Francis, ed., Multiple Task Performance. London. Yea, B., Konishi, R., Osaki, T. and Sugahara, K. (1994). ‘The discrimination of many kinds of odor species using fuzzy reasoning and neural networks.’ Sensors and Actuators 45, 159–165. Zadeh, L. (1965). ‘Fuzzy Sets.’ Information and Control 8, 338–353.

© Woodhead Publishing Limited, 2013

10 Advanced methods for the control of food processes: the case of bioconversion in a fed-batch reactor D. Dochain, Université catholique de Louvain, Belgium

DOI: 10.1533/9780857095763.1.226 Abstract: Novel approaches to guarantee complete and intelligent control of entire food processing plants are required in order to ensure product quality, flexible and efficient operation, and to reduce the impact of processes on the environment. The objective of this chapter is to present several issues associated with, and advanced methods for, modelling, monitoring and control that may be appropriate for the handling of on-line process optimization in the context of food processes. The chapter mainly concentrates on a bioconversion process in a fed-batch reactor. Key words: food engineering, population balance models, state observers, parameter estimation, extremum-seeking control, adaptive control.

10.1 Introduction The food industry is well established and many processes in operation nowadays are the subject of intensive work with regard to ways of devising better operation modes in terms of product quality and safety (how to operate in order to ensure quality and comply with safety constraints) as well as in terms of operation costs and environmental impact. There is also intensive development work aimed at responding to consumer demands by designing new products and designing and operating the most appropriate combination of unit operations needed to produce them. However, and despite the fact that the essential physical, biochemical and microbiological principles are reasonably well understood, foods are complex systems with properties that, because they are connected with quality and safety, are usually

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

227

very difficult to measure, estimate or even represent through reliable models. Such properties may include physico-chemical parameters associated with quality, such as nutrient content, texture, colour or rheology, or microbiological characteristics usually connected with food safety. In addition, and from a process engineering perspective, the food industry integrates a rich variety of apparently very diverse processes and technologies, thus hampering the search for unifying paradigms that would be useful for dealing with different yet analogous processes. Such processes have only recently been classified into a reasonably small number of categories, namely bioconversion, separation, preservation and structuring.1 Bioconversion includes probably the largest class of processes. It is central in the production of fermented foods, with many well-known examples from the dairy or winemaking industry, but it is also the purpose of cooking in its more general conception. Firstly, it is based on the use of a biologically active element, such as bacteria or yeast. Secondly, the biochemical transformation of food is most often performed through cooking. In both cases, with the appropriate combination of ingredients and additives, a number of relevant reactions activated by temperature or bioactive agents take place. These reactions not only modify the structure of proteins (denaturalization) to make them more digestible, but also add flavours and promote combinations between lipids, proteins and carbohydrates that increase the quality of food in terms of appearance, texture and taste, even making some nutrients or biochemical compounds of interest improve the bioavailability. Understanding the underlying kinetic mechanisms and developing models of use in sensing technologies and control are crucial aspects of these types of processes. Sensing technologies would allow the accurate estimation of quality parameters that need to be under control. Models and simulation tools will also be required to develop efficient control algorithms to maximize quality for this type of process. Separation and purification operations are widely employed in the food industries, mostly as downstream processing after bioconversion, or as a means of extending the shelf-life of the product by reducing the water available in the foodstuff. Typical examples include microfiltration, evaporation and drying. Extraction operations (such as in gelatine or pectin production) are also separation operations widely used in the food industry to recover high-value biocompounds. Other separation processes, such as distillation in its versions of flash, molecular or vacuum distillation, are also examples characteristic of, for instance, the oil-refining industry. Freeze-drying, cryoconcentration of juices or other fluids of interest in the food industry, or super-critical extraction processes, which are typically employed to recover highly valuable bioactive compounds, complete the long list of operations involved in separation. Preservation processes and technologies are those aimed at increasing the product life and ensuring health and safety by reducing the growth of microorganisms (in freezing or refrigeration, for instance) or by eliminating them (as in pasteurisation or sterilization) with thermal treatments such as steam heating, microwave or ohmic heating. The so-called emerging technologies, based on high-voltage electric pulses or high pressures, fall into the same category. Other preservation

© Woodhead Publishing Limited, 2013

228 Robotics and automation in the food industry technologies are those having a chemical or biochemical basis, such as smoking, salting, or those aimed at reducing water activity and thus reducing degradation kinetics or growth of microorganisms, as in the case of freeze-drying. Structuring processes integrate a number of technologies aimed at modifying the food product microstructure, such as in extrusion, emulsification or crystallization. Numerous operations are concerned, in which the structure of the product is modified or promoted. Most of them are thermo-mechanical operations. For instance, the introduction of extrusion, during the1960s, promoted the design of new-properties-based products. Also heating (or cooling, freezing), because of a better understanding of combined thermal and mechanical mechanisms, became a way to control the structure at different levels. On the other hand, foaming and emulsion operations are the basis of new products, even if control is still difficult due to lack of sensors. From a control and automation point of view, the following weaknesses have been detected in most food companies. Most plant control schemes reduce to local and decentralized control loops acting on a usually very small number of states (typically temperature or pressure) not directly connected with product quality, and in many cases not connected either with critical aspects of the operation such as water or energy consumption. Despite the fact that the performance of PID control can be in many instances more than acceptable (as is also the case in other process industries), the control loops should be combined among the different processes to highlight synergies and not to cancel them. In addition, this regulatory layer is not commanded or integrated on higher supervisory levels. Although presently many food plants benefit from advances in data acquisition and monitoring of the full production line to gather and store huge amounts of data, the use of such information is quite limited, usually not efficiently employed, and reduced to configure alarms (often handled at a very low level) or to help in producing simple production decision rules and off-line control of inventories. Much more effective use of such information could be made possible, when properly combined with process models and prognosis tools, to be able to estimate unmeasured yet relevant plant states and to predict future scenarios, even in a real-time context. Often recognized as a specificity of food processes, the lack of sensors for relevant product characteristics is still a problem. Even if numerous algorithms are available for advanced control purposes, it is obvious that on-line, real-time reliable information is necessary. Common sensors (temperature, pressure pH, flowrates,) give information only very indirectly related to the product properties of interest, such as texture, aroma content, biomass, contaminants, vitamins, etc. Sensors for such product properties are either missing completely or very expensive and not robust or reliable enough to be used in everyday industrial practice. Improving the reliability of sensing devices and developing new hardware–software sensing techniques for on-line estimation of difficult quality product parameters, are critical in developing smart-control applications for food factories. It is important to note that, generally speaking, there is a lack of control design and operation paradigms to optimally operate plants, either learnt or inherited

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

229

from diverse yet similar processes or scenarios. One such paradigm is that pursued by the engineering community focusing on the same concepts, such as simulation, optimality or optimization as those containing a real systematic approach to the problems of devising smart control for food processes. In order to overcome the present limitations detected in food production plants, it is important to offer novel and original integral approaches to guarantee complete and intelligent control of the whole plant in response to changes in the demand and the supply of materials, in order to ensure product quality, flexibility and efficient operation, compliant with the environment.2 In this chapter, for obvious space reasons, we shall concentrate on one class of food processes (bioconversion) and present several issues and advanced methods in modelling, monitoring and control that may be appropriate to handle the on-line process optimization in the context of food processes. The chapter is organized as follows. We shall start by briefly introducing in Section 10.2 the mass balance model that will serve as a guideline throughout this chapter. In Section 10.3 we shall concentrate on modelling issues, by considering the importance today of considering the particulate dimension (via the distribution of size, mass or age) in food processes, briefly introducing the use of population balance to complement mass balance modelling and address some of the complexity related to these partial differential equation models. In Sections 10.4, 10.5 and 10.6, we shall mainly concentrate on several key issues related to fed-batch processes, which are playing a central role in the bioconversion processes of the food industry today. Fed-batch bioreactors represent an important class of bioprocesses, mainly in the food industry (e.g. yeast production or alcoholic fermentation) and in the pharmaceutical industry (such as the production of the vaccine against hepatitis B) but also, for example, for biopolymer applications (poly-beta-hydroxybutyric acid, PHB). It is also very much involved in the field of enzyme production, which has been developed over the past decade due to the recombinant DNA technology and via the use of filamentous microorganisms. In a fed-batch reactor, the tank is only partially filled before the start-up of the process operation, and the reactive matter is progressively added until the tank is full (this is the feeding phase summarized by the word fed in the denomination of the operation). One waits for the different reactions to be completed (batch phase) before emptying the tank. In Section 10.4, we shall address the monitoring of fed-batch processes by considering the tuning of observer-based estimators for unknown parameters that can handle the time-varying nature of the fed-batch process dynamics. This issue is also the object of Section 10.5, which will concentrate on the specificities of the design of PID controllers for fed-batch reactors. Section 10.6 will finally concentrate on the design of real-time optimization algorithms, namely adaptive extremum-seeking controllers, for fed-batch processes.

10.2 The basic dynamical model Consider the following dynamical model of a simple microbial growth process:

© Woodhead Publishing Limited, 2013

230 Robotics and automation in the food industry 0.1

Specific growth rate µ (h–1)

µ max 0.08 0.06

µ 0 = 0.53 h –1 KS = 1.2 g/L

0.04

KI = 0.22 g/L 0.02 (KS KI)1/2 0 0

0.5 1 1.5 Substrate concentration S (g/L)

2

Fig. 10.1 Haldane model.

S→X

[10.1]

in a completely mixed fed-batch reactor: dX = µX dt

[10.2]

DX

dS = k1 X + D (Sin − S ) dt

[10.3]

dV = DV dt

[10.4]

where X (g/L) and S (g/L) are the biomass and substrate concentrations, respectively. µ (h−1) is the specific growth rate, D (h−1) is the dilution rate, Sin (g/L) denotes the concentration of the substrate in the feed, k1 is a yield coefficient, and V (L) is the volume of liquid medium in the tank. A typical situation where bioprocesses are operated in fed-batch is when substrate inhibition is possibly present. The Haldane kinetic model is then typically used to characterize such inhibition phenomenon. This model (see Fig. 10.1) is given by:

µ=

µ0S KS + S +

(

)

[10.5]

where µ0 is a parameter related to the maximum value of the specific growth rate as follows: µ 0 = µ max 1 2 K S K I . The coefficients KS and KI denote the saturation constant and the inhibition constant, respectively.

( (

))

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

231

10.3 Modelling issues: population balance modelling in food processes One of the key issues presently in process control is to appropriately address the fact that production quality is not just a global property of the mass produced within the process but also of the distribution of different variables, such as size, mass or age. In crystallization processes, it is obvious that the main objective is to produce crystals of a specified size.3 As mentioned in the introduction, crystallization plays a major role in the food industry. As an example, the EC project CAFE2 deals in particular with the control of ice cream crystallization. A population balance model has been considered to emphasize the process dynamics.4,5 Such a model has been identified6 and the model properties have been studied,7 in order to develop and implement an appropriate controller of the process. The same concern applies also to the production of additives such as anti-foam particles for washing powders, where the size of additives needs to be similar to the powder to ensure that the mixture is sufficiently homogeneous.8 When considering living organisms, it is clear that the age of the cells plays an important role in the efficiency of the process, and therefore the knowledge of the age distribution (typically via mass distribution measurements) is a key issue to improving the efficiency of biological processes.9 The natural way to characterize the particulate dimension of industrial processes is to consider population balance.10 There has been in recent years a great deal of research activities concerning the modelling and analysis of the properties of the cell population balance model. Such models describe the dynamics of cell growth and take into account the biological fact that the cell properties (e.g. mass or age) are distributed among the cells of a population. It consists of a non-linear partial integro-differential equation with a non-linear boundary condition coupled with an ordinary integro-differential equation.11–13 Another difficulty with such a model is due to the intrinsic physiological functions, namely the growth rate function, the cell division probability density function, and the partitioning probability density function, whose selection may appear to be a complex task in many instances. As an example, let us consider a cell population growing in a continuous stirred tank reactor. The cells are distinguishable from each other in terms of their mass or some other property of the cell that obeys the conservation law. Let N(m, t) be the number of cells that have a mass between m and m + dm at time t. The cells are considered to grow at a rate r(m, S) that depends on their mass and on the concentration of the limiting substrate S. We also assume that the value of the mass is standardized and that m ϵ [0,1]. The cell division and the birth processes of the cell population are described by the division rate Γ(m, S) defined as follows: Γ ( m, S ) =

f ( m) 1

∫

m

0

f ( m′ ) dm ′

r ( m,S S)

γ ( m ) r ( m, S )

© Woodhead Publishing Limited, 2013

[10.6]

232 Robotics and automation in the food industry where f(m) is the division probability density function, which is assumed to depend only on the cell mass, and is taken to be a left hand side truncated Gaussian distribution with mean of µf and standard deviation of σf. We also assume that the probability p that a mother cell of mass will give birth to a daughter cell of mass m is independent of substrate concentration. This function should further satisfy the following normalization conditions: m′

∫ p ( m, m′ ) dm = 1 0

and should also be such that biomass is conserved at cell division, that is: p ( m, m′ m)

p ( m ′ m, m )

We finally assume that the probability function is a symmetrical Beta distribution with a parameter of q defined by the following equation: p ( m, m ′ ) =

1 1 ⎛ m⎞ ⎜ ⎟ B ( q, q ) m ′ ⎝ m ′ ⎠

q 1

m⎞ ⎛ ⎜⎝1 − ⎟⎠ m′

q 1

[10.7]

We assume also that no cell death occurs and that cells grow in a continuous propagator from which they exit with a dilution rate of D. Under these assumptions, the cell population dynamics are described by the following integro-differential equations, see example:11 ∂N ( ∂t

)+

∂ ⎡r ( ∂m ⎣

= −DN (

)N (

)⎤⎦ + Γ (

1

) + 2∫ Γ ( m ′,′ S ) p (

)N (

′) N ( ′

,

)

) dm ′

[10.8]

m

subject to the initial condition: N (m, 0) = N0

[10.9]

System [10.8]–[10.9] is completed by the following boundary conditions: r(1, S) N(1, t) = r(0, t) N(0, t) = 0

[10.10]

The cell population Equation [10.8] consists of five terms: the accumulation term, the growth term, the division term, the dilution term and the birth term.

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

233

The behaviour of the cell population depends on the substrate concentration. The mass balance equation for the substrate expresses in particular that the substrate consumption is proportional to the total biomass production r(m, S) N(m, t)dm. The substrate concentration mass balance then reads as follows: 1

dS = D (Siin − S ) k1 ∫ r ( m,S ) N ( m,t ) dm dt 0

[10.11]

The main feature of the above population balance model as compared to the basic mass balance model [10.2]–[10.3] is that the model now includes a partial differential equation and integral terms in both mass balance equations. Such a class of partial differential equations has not so far been very extensively studied in the control community. Due to the complex structure of the dynamical model, the existing results on the analysis of the system, including the analysis of the system solution, the study of the existence, multiplicity and stability of equilibrium profile or on the control of this system, have mostly been obtained numerically, see references 11–13. However in reference 14, the problem of analysing and stabilizing of the steady state of the cell mass distribution in a continuous bioreactor is studied in an infinite dimensional system framework. The development of control laws for a cell population balance model has become an active area of research during the past decades. Several non-linear controllers or non-linear adaptive controllers have been proposed for this class of systems, for example, to ensure the control of the process productivity (for a bioprocess with two cell cycle stage13), or to control the different moments (zeroth moment (representative of the cell density), first moment (representative of the biomass concentration) and the second moment) by using a non-linear linearizing controller.11

10.4 Monitoring issues: tuning of observer-based estimators Let us consider parameter estimation via the use of the observer-based estimator15 and emphasize the specificities of its tuning for batch and fed-batch processes. Consider that the process dynamics are described by the following equations: dx = F1 ( x ) θ + F2 ( x ) dt

[10.12]

where x is the state vector (dim(x) = n), θ is the vector of (unknown) parameters (dim(θ) = p), and F1(x) and F2(x) are (matrix), generally non-linear, functions of the state vector x. A typical example is the general dynamical model:15

© Woodhead Publishing Limited, 2013

234 Robotics and automation in the food industry dx = Kr ( x ) F − Q Dx dt

[10.13]

where x, K, r(x), F, Q hold for the component concentration vector, the yield coefficient matrix, the reaction rate vector, the feedrate vector, and the gaseous outflow rate vector, respectively. In the present context, the above model can be rewritten as follows: dx = KG ( x ) θ F − Q Dx dt

[10.14]

The reaction rate r(x) is now rewritten as the product of a function of the state (the diagonal matrix G(x)) and the unknown parameter θ. In the basic dynamical model example [10.2]–[10.3], a typical choice for G(x) and θ could be: G (x) =

SX KS

(

S + S KI

)

,

θ = µ0

Assume that: • H1: the p parameters θ are unknown and possibly time-varying (with bounded time variations dθ d < M ; • H2: p state variables are available for on-line measurement. From the assumption H2, we can define a state partition: ⎡ x1 ⎤ x=⎢ ⎥ ⎣ x2 ⎦

[10.15]

with x1 the measured variables, and x2 the unmeasured ones. The dynamical equations can then be rewritten as follows: dx1 = F11 ( x ) θ + F21 ( x ) dt

[10.16]

dx2 = F12 ( x ) θ + F22 ( x ) dt

[10.17]

Further assume that: H3. F11 can be written as the product of two p × p matrices: F11 = F3F4(x) with F4 a diagonal matrix (F4 = diag{f4,i}, i = 1 top) and F3 being full rank for all admissible values of x (for (bio)chemical processes for instance, only positive

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

235

values of the state variables (i.e. concentrations and possibly temperature) are considered);H4. F11(x) and F21(x) are known functions of x. Under assumption H3, it may be possible to build an asymptotic observer15 to reconstruct the time evolution of x2 independently of the unknown parameters θ. In the following we assume that states included in x2 are either accessible for on-line measurement or available via such observers, or that the dynamics of x1 are independent of x2. In the context of stirred tank reactors (STR), F3 = K1, F4 = G (x) with K1 the yield coefficient matrix associated to x1. Note that by constructing and assuming H2, the matrix G(x) is a p × p matrix. Assumption H4 then means that the feedrates F1 (associated to x1), the gaseous outflow rates Q1 (associated to x1), the influent flow rate q and the volume V are known (via on-line measurements or user’s choice), as well as the yield coefficients in K1 and the function G(x). The design of the observer-based estimator is based on Equation [10.16] and follows indeed the line of reasoning for the design of Luenberger observers. This gives the following estimator equations: dxˆ1 = F11 ( x ) θ + F21 ( x ) − Ω ( x1 dt T dθ1 = ⎡ F11 ( x )⎤⎦ Γ ( x1 dt ⎣

xˆ1 )

x1 )

[10.18]

[10.19]

The basic motivation of the above structure for the estimator is the following. As for a classical observer, the estimator equations are the combination of the process model (F11(x)θ ̂ + F21(x)) and correction terms (−Ω(x1 − x̂1)) and ([F11(x)]T Γ(x1 − x̂1)) on the measured variables. In the above observer-based estimator, the parameters θ are assimilated as states without dynamics. The weighting factor [F11(x)]T in Equation [10.19] is indeed the term multiplying the unknown parameter in the model equation: its introduction in the estimator equation derives from classical estimator design and it is usually called the regressor. An important difference with respect to the extended Luenberger or Kalman observers is that the measured variables do not appear as estimates in the observer equations but with their ‘true’ (measured) values. The theoretical stability analysis of the above observer-based estimator is available in reference 15: the main requirements are the negative definiteness of ΩT Γ + ΓΩ and the persistence of excitation of F11(x). However, its tuning may be difficult and intricate in practice, because of the close interaction of the unknown parameters in the estimator equations, and because its dynamics depends on the process variables. The latter may be of minor importance if the system is operated around a steady-state, but it will become crucial if the system covers a large range of operating conditions (as for fed-batch and batch reactors, or process start-ups and grade changes) possibly with large variations of the state variables, and as a consequence, the matrix F11(x). Good tracking capabilities of the parameters’

© Woodhead Publishing Limited, 2013

236 Robotics and automation in the food industry variations are particularly essential in these circumstances. However, in the above form Equations [10.18] and [10.19] of the observer-based estimator, tuning may give very conservative values for the design parameters Γ and Ω, and result in bad tracking performance in some of the operating regions. The above objectives for the tuning of the observer-based estimator can be achieved by considering first the following two steps in the reformulation of the algorithm: 1. a state transformation, 2. the rearrangement of the estimator state vector entries. Consider the following state transformation: F3−1x

[10.20]

Then the dynamic equations of the system can be rewritten as follows: dz = F4 ( x ) θ + F3 1F221 ( x ) dt

[10.21]

Due to the above transformation, only one state variable is associated with each unknown parameter θ. In the specific case of STR, the invertibility of F3 (=K1) results from the independence of the p reactions and of the p measured variables, as already mentioned before. The observer-based estimator can now be redesigned on the basis of Equation [10.21]: dzˆ = F4 ( x ) θ + F3 1F221 ( x ) Ω ( z − zˆ ) dt

[10.22]

dθ = Γ (z z) dt

[10.23]

Due to the transformation Equation [10.20], the observer-based estimator is now reformulated in a decoupled format for the unknown parameters θi (i = 1 to p). Because of the decoupled estimation formulation, an obvious choice for the matrices Ω and Γ are diagonal matrices: Ω

diag {

}

Γ = diag di {

0, γ i } , ωi > 0,

0, i = 1 to p

[10.24]

In the above formulation of the estimation scheme, we have removed the regressor term F4(x) (G(x) in the STR example) from the estimation equation of θ (Equation [10.23]); since one of its main roles is to explicitly transfer the coupling between the unknown parameters and the measured variables, its presence is no longer essential.

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

237

The final step before the formulation of the tuning rule consists of a rearrangement of the estimator’s equations. Let us gather each variable zi with its related parameter θi and rearrange the entries of the vector [z, θ]T in the following order in a vector ζ: ⎡ z1 ⎤ ⎢θ ⎥ ⎢ 1⎥ ⎢ z2 ⎥ ⎢ ⎥ ς = ⎢ θ2 ⎥ ⎢…⎥ ⎢ ⎥ ⎢zp ⎥ ⎢θ ⎥ ⎣ p⎦

[10.25]

Let us first define the estimation error e = ς − ςˆ

[10.26]

The estimation error dynamics are readily derived from Equations [10.21], [10.22] and [10.23]: de = Ae + b dt

[10.27]

with a block diagonal matrix A with 2 × 2 blocks: A

⎡ −ω diag {Ai } Ai = ⎢ i ⎣ −γ i

f

,i

( x )⎤ ,

0

⎥ ⎦

i

1

p

[10.28]

and b equal to: ⎡ 0 ⎤ ⎢ dθ ⎥ ⎢ 1⎥ ⎢ dt ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ dθ2 ⎥ b=⎢ dt ⎥ ⎢ ⎥ ⎢…⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ dθ p ⎥ ⎢ ⎥ ⎣ dt ⎦

[10.29]

© Woodhead Publishing Limited, 2013

238 Robotics and automation in the food industry The characteristic equation of the matrix A, det(λI − A), is equal to: p

) ∏ ( λ 2 + ωi λ + γ i f4,i ( x ))

det ( λ

[10.30]

i =1

The key idea of the tuning rule consists of choosing each γi inversely proportional to the corresponding term f4,i(x):

γi =

γi

f4,i ( x )

, γ i > 0, i = 1 to p

[10.31]

With the choice above, the characteristic Equation [10.30] is rewritten as follows: d t (λ

p

) ∏ ( λ 2 + ωi λ + γ i ) i =1

[10.32]

and the observer-based estimator dynamics are now independent of the state variables. Such a choice corresponds to a Lyapunov transformation.16 It is obviously valid for values of f4,i(x) = 0: this condition is usually met easily in (bio)process applications, as will be illustrated in the following section. The values of the design parameters can then be set to arbitrarily fix the estimator’s dynamics for each unknown parameter θi. Since the estimator reduces via the transformations to a set of independent second-order linear systems, the classical rules for assigning the dynamics of second-order linear systems apply straightforwardly here. The reader is therefore referred to classical automatic control textbooks for further information on the subject. However the following basic guidelines are suggested. One important guideline is to choose real poles:

ωi2

4γ i

0

[10.33]

The objective is then to avoid inducing oscillations in the estimation of the parameters that do not correspond to any physical phenomenon related to the estimated reaction rates. Pomerleau and Perrier17 suggest choosing double poles, that is:

γi =

ωi2 4

[10.34]

Oliveira et al.18 propose as an alternative to choose complex poles with a damping factor equal to 0.7 in order to increase the speed of convergence of the estimator with a reduced overshoot. (Generally speaking, the damping factor can

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

239

be freely chosen; the choice may then depend on the type and nature of the application, of the time variations of the parameter to be estimated, and of the noise on the measured data. This means that then there are two design parameters per estimated parameters.) Then the tuning of the estimation algorithm reduces to the choice of one design parameter, ωi, per estimated parameter. This allows having a design procedure that has the double advantage of being simple (one design parameter) and flexible (each parameter estimation can be tuned differently if needed, for example, if the time variations of the parameters are different). As a matter of illustration, let us show real-life results of the observer-based estimator implemented on a baker’s yeast production process.17 The dynamical model considered for the observer-based estimator is based on the three reaction scheme proposed in reference 19: Respirative growth on glucose : S + C → X

P

[10.35]

Fermentative growth on glucose : S → X + E

P

[10.36]

Respirative growth on ethanol : E + C → X

P

[10.37]

where S, C, X, P and E hold for glucose, oxygen, yeast, CO2 and ethanol, respectively. The observer-base has been designed by two sub-models (one in an oxygen reductive regime with first two growth reactions, and one in a respirative regime with the first and the third reactions) and the on-line measurements of dissolved oxygen and CO2. Some results are given in Fig. 10.2. Figure 10.2a gives the estimation of the yeast concentration via an asymptotic observer. The three following panels (Fig. 10.2b, 10.2c and 10.2d) give the estimation results for the specific growth of the three growth reactions, namely µo, µr and µe. The validation of the values of the estimated specific growth rates has been performed by comparing the measured value of the respiratory quotient (RQ) with the value computed from the estimation of µo, µr and µe (Fig. 10.2e).

10.5 Design of PID controllers for fed-batch processes The design and calibration of classical controllers such as PID are typically based on a linear time-invariant model of the system. Yet most processes are characterized by a non-linear dynamical behaviour. The underlying assumption is that the non-linear dynamics can be fairly well approximated by a linear model. If this can often largely be justified in the case of regulation (i.e. when the objective is to maintain the process close enough to some defined operating conditions) when a linearized model around the process steady-state can give a good approximation dynamics, it is most likely that this will not be the case when the process

© Woodhead Publishing Limited, 2013

240 Robotics and automation in the food industry (a) X (g/L)

(b)

40 30 20 10 0.20

µo 0.10 (1/h)

(c)

0.00 0.20

µo 0.10 (1/h)

(d)

0.00 0.02

µr 0.01 (1/h) 0.00 (e) RQ

2.0 1.0 0.0

0

1

2

3

4

5 6 Time (h)

7

8

9

Fig. 10.2 Observer-based estimator: experimental results on a baker’s yeast production process. For explanation of panels (a) to (e), see text.

is operated in non-steady state conditions as in batch and fed-batch operating modes. The objective of this section is to illustrate with a simple fed-batch example how to handle this specific aspect of the process dynamics. Let us start by considering a fed-batch process whose dynamics are described by the mass balance model Equations [10.2]–[10.4] with Haldane kinetics Equation [10.5]. If the objective is to maximize the biomass production, intuitively it seems reasonable to believe that this objective would be reached if the substrate concentration S is kept at a value that maximizes the specific growth rate, that is, S KS K I . This is indeed what can deduced by determining the optimal control input via the use of the Pontryagin Maximum Principle.20 Let us consider two cases depending on the control input. First let us assume that the inlet substrate concentration Sin is the control input and that the inlet flow

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

241

rate q (= DV) is fixed (and e.g. constant) (this implies that the feeding period is fixed). This choice is primarily motivated by the simplicity of the optimal control computation. The optimal control problem consists of finding the feeding rate profile Sin(t) that maximizes the following objective function:

(t ) X (t ) f

[10.38]

f

It can easily be shown that the singular optimal control input is written as follows: Sin (t ) = K S K I +

µ 0 X (t ) V (t ) k1 q 1 + 2 KS K I

[10.39]

where V(t) and X(t) are solutions of the differential Equations [10.2], [10.4] in which S takes the constant value K S K I . It is obtained by first noting that the Hamiltonian of the system is equal to: H = λ1

⎛q (S ⎝V

⎞ S ) − k1µX ⎟ + λ 2 ( X ⎠

DX ) λ 3 q

[10.40]

which can be rewritten as follows: H

φ ψu

[10.41]

with u = Sin and

φ

⎛ q ⎞ λ1 − S − k1µ ⎝ V ⎠

ψ

λ1

⎛ λ2 µ ⎝

q V

⎞ ⎠

λ3q

q V

[10.42]

[10.43]

In the above relationships, λ1, λ2, λ3 are the co-states whose dynamics are given by the following equations: dλ1 ∂H =− = ( λ1 dt ∂S dλ 2 ∂H =− = ( λ1 dt ∂X

∂µ q + λ1 ∂S V

[10.44]

q V

[10.45]

1

λ2 ) X

1

λ2 ) µ λ2

© Woodhead Publishing Limited, 2013

242 Robotics and automation in the food industry dλ 3 ∂ H q ( λ1 = = dt ∂V

λ2 )

i in

V

2

[10.46]

Let us compute the optimal singular control input (for which ψ = 0). First note that:

ψ

λ1 = 0

[10.47]

The time derivative of ψ is then equal to: dψ q ∂µ = −λ 2 X dt V ∂S

[10.48]

It is equal to zero if λ2 = 0 or if ∂μ/∂S = 0. Only the second option is of interest. It corresponds to the (constant) value of the substrate concentration S that maximizes the specific growth rate µ, that is: Ssin g

K I KS

[10.49]

The second-order derivative of ψ is written as follows by considering (∂μ/∂S) = 0: d2 ψ ∂2 µ ⎛ q = −λ 2 X ⎜ (Sin 2 dt ∂S 2 ⎝ V

⎞ S ) − k1µX ⎟ ⎠

[10.50]

It is equal to zero if we consider the control input [10.39]. In this simple case, the optimal control will be either of the ‘bang-singular’ type (if the initial substrate concentration S(0) is different from K I K S ), or purely ‘singular’ (if S ( ) K I K S ). Note that the feeding rate profile Equation [10.39] is exponential. Indeed under the hypothesis that q and S (and therefore µ) are constant, the total biomass V(t) X(t) is equal to: V (t )X (t ) = V0 X 0 e

µ max t

[10.51]

with

µ max =

µ0 1 + KS K I

[10.52]

It is worth noting that this corresponds to an important dimension in the control (fed)-batch bioreactors: the optimal biomass profile is exponential, and ideally

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

243

X (g/L)

6 4 2 0

0

5

10

15

20

0

5

10

15

20

50 40 30 20 10 0 0

5

10

15

20

S (g/L)

1.5 1 0.5

Sin (g/L)

0

t (h)

Fig. 10.3 Optimal control of a fed-batch bioreactor.

the control action that allows it to have this exponential profile, and therefore the maximization of the biomass production, needs to be time-varying (exponential in the present situation). The optimal control of the fed-batch bioreactor is illustrated in Fig. 10.3 in the K I K S with the following set singular control case [10.39]; that is, when S ( ) of parameter and initial values:

µ max = 5 h

1

S ( 0 ) = 1 g/L g/ ,

S

(0)

g/L, g/L g/

I

= 0.1 g/L g/ , k1

.1 g/L, ((0) 0))

1, q

0.5277 L/h

, V ( f ) = 20 l

The optimal control stops when the tank is filled (tf = 18.95 h), then the batch phase starts (q = 0) when the substrate S is depleted. The example is of particular interest since it emphasizes that controlling the substrate concentration S at a (constant) value (that maximizes the specific growth rate) is enough to maximize the total biomass production. This means that we reduce a trajectory tracking problem to a simple regulation problem. However the issue is that this is indeed just simply a ‘regulation’. As a matter of fact, while the substrate concentration S remains constant (or is assumed to remain constant), the other process variables, and in particular the biomass, are covering a wide spectrum of values. In the above simple example, we have seen that under optimal conditions, the biomass quantity V(t) X(t) is following an exponential profile. It is therefore rather obvious that not taking this phenomenon into

© Woodhead Publishing Limited, 2013

244 Robotics and automation in the food industry 6

× 10−3

Oxygen dissolved (g/L)

5

4

3

2

1

0

0

5

10

15

20

25 t (h)

30

35

40

45

Fig. 10.4 PI control of the dissolved oxygen concentration in an industrial fed-batch bioreactor.

account may lead to poor control performance. And it is quite easy to understand that the tuning of the controller may be efficient in some operating condition range while being less efficient, if not possibly disastrous, in another operating condition range. This is illustrated in Fig. 10.4 for the real case of an industrial fed-batch reactor where the PI regulator (which controls the dissolved oxygen concentration), which has been tuned for conditions corresponding to the early part of the fermentation, gives a noisy signal after a sharp initial transient and then exhibits an oscillatory behaviour at the end of the fermentation, simply because the system dynamics are then very different from those observed at the beginning of the fermentation. This can be explained by looking at the closed-loop dynamics of the substrate concentration S. With the PI controller, this is written as follows:

S S⎡ KP dS * = − k1µX + in ⎢ K p (S − S ) + dt V ⎣ τi

=

t

(S ∫ (S 0

*

⎤ S )dτ ⎥ ⎦

⎡ K 1⎛ − k1 V0 X 0 e µt + (Sin − S ) ⎢ K p (S * − S ) + P ⎜ V⎝ τi ⎣

t

(S ∫ (S 0

© Woodhead Publishing Limited, 2013

[10.53]

*

⎤⎞ − S )dτ ⎥⎟ ⎦⎠

[10.54]

Advanced methods for the control of food processes

245

X (g/L)

6 4 2 0

0

5

10

15

20

5

10

15

20

15

20

S (g/L)

2 1.5 1 0.5 0

0

q (L/h)

3

Kp = 2, τi = 0.5, S(0) = 0.4

2 1 0

0

5

10 t (h)

Fig. 10.5 PI control of a fed-batch bioreactor: constant gain and S(0) = 0.4 g/L.

under the assumption that S remains close to S* (and that in consequence µ remains approximately constant). Note that the first term (exponential) on the right hand side is only compensated by terms that do not vary in the same way. This behaviour is illustrated (Fig. 10.5) in our case study for an initial condition S(0) = 0.4 g/L, different from the optimal value (in spite of an anti-windup action). A solution consists then to explicitly incorporate the exponential growth term into the control input.21 In a PI controller, this can be done via the proportional gain Kp (and only on the proportional action). If we denote the integral gain of the PI controller: Ki =

Kp

[10.55]

τi

where Ki is kept constant, and if we choose the proportional action gain value as follows: Kp

K p e µt

[10.56]

as suggested in the above equation which is now rewritten as follows: dS e µt ⎡ K p (S = S * − S )(S Sini − S dt V ⎣

k1µV X ⎤⎦ +

t Sin S K i ∫ (S S * − S )dτ 0 V

© Woodhead Publishing Limited, 2013

[10.57]

246 Robotics and automation in the food industry

X (g/L)

6 4 2 0

0

5

10

15

20

0

5

10

15

20

2.5 2 1.5 1 0.5 0 0

5

10

15

20

S (g/L)

1.5 1 0.5

q (L/h)

0

t (h)

Fig. 10.6 PI control of a fed-batch bioreactor: proportional gain with an exponential term.

Note that if the exponential term in the gain Kp is appropriately chosen (i.e. as close as possible to the optimal value of the specific growth rate µ) we introduce in the control input an exponential action that allows us to handle the exponential biomass growth. This is illustrated in Fig. 10.6 where the same conditions as in Fig. 10.5 have been considered.

10.6 Real-time optimization Most adaptive control schemes documented in the literature (e.g. references 22–23) are developed for regulation to known set-points or tracking known reference trajectories. In some applications, however, the control objective could be to optimize an objective function, which can be a function of unknown parameters, or to select the desired values of the state variables to keep a performance function at its extremum value. Self-optimizing control and extremum-seeking control are two methods to handle these kinds of optimization problems. The task of extremum seeking is to find the operating set-points that maximize or minimize an objective function. Since the early research work on extremum control in the 1920s,26 several applications of extremum control approaches have been reported (e.g. references 22, 27–29). Krstic et al.30,31 presented several extremum control schemes and stability analysis for extremum seeking of linear unknown systems and a class of general non-linear systems.30–32 A neural network based approach has been proposed in reference 33.

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

247

The implications for the biochemical industry are clear. In this sector, it is recognized that even small performance improvements in key process control variables may result in substantial economic gains. As an example, the potential benefits of extremum-seeking techniques in the maximization of biomass production rate in well-mixed biological processes has been demonstrated in reference 34. The extremum-seeking framework proposed by Guay and Zhang35 assumes that the objective function is explicitly known as a function of the system states and uncertain parameters from the system dynamic equations. Parametric uncertainties make the on-line reconstruction of the true cost impossible, such that only an estimated value based on parameter estimates is available. The control objective is to simultaneously identify and regulate the system to the lowest cost operating point, which depends on uncertain parameters. The main advantage of this approach is that one can guarantee some degree of transient performance while achieving the optimization objectives when a reasonable functional approximation of the objective function is available. The proposed scheme utilizes explicit structure information of the objective function that depends on system states and unknown plant parameters. This scheme is based on Lyapunov’s stability theorem. As a result, global stability is ensured during the seeking of the extremum of the non-linear stirred tank bioreactors. It is also shown that once a certain level of persistence of excitation (PE) condition is satisfied, the convergence of the extremum-seeking mechanism can be guaranteed. In this section we concentrate on the adaptive extremum-seeking control of bioreactors operating in the fed-batch mode (see also references 36 and 37) but several other alternatives of similar schemes (including the use of universal approximation such as artificial neural networks (ANN)) have been proposed in the literature for different bioprocess configurations,38–42 but also for chemical reactors,43–46 pulp and paper processes47 and biomedical systems.48

10.6.1 Adaptive extremum-seeking control of fed-batch bioprocesses Fed-batch bioreactors represent an important class of bioprocesses, mainly in the food industry and in the pharmaceutical industry but also for example for biopolymer applications (PHB). One of the key issues in the operation of fed-batch reactors is to optimize the production of a synthesis product (e.g. penicillin, enzymes, etc) or biomass (e.g. baker’s yeast). They are therefore ideal candidates for optimal control strategies. An intensive research activity has been devoted to optimal control of (fed-batch) bioreactors mainly in the 1970s and in the 1980s (see e.g. references 49–52). Yet in practice, because of the large uncertainty related to the modelling of the process dynamics,15 poor performance may be expected from such control strategies, and although a priori attractive, optimal control has not been largely applied to industrial bioprocesses. Alternative approaches have been proposed for handling the process uncertainties with an adaptive control scheme (e.g. reference 53). In the present approach, we propose to go a step further by including a static optimum search in the adaptive control scheme.

© Woodhead Publishing Limited, 2013

248 Robotics and automation in the food industry Consider the dynamical model of the simple microbial growth process in a fedbatch reactor Equations [10.2]–[10.4] with one gaseous product: y

k2 µX

[10.58]

where y (g/L/h) is the production rate of the reaction product and k2 is a yield coefficient. A typical situation in bioprocess applications is when the biomass concentration is not available for on-line measurement while the gaseous outflow rate y (e.g. CO2) is easier to measure on-line. We consider here that only S and y are measurable while the biomass concentration X is not available for feedback control. Here we consider the extremum-seeking problem for the bioprocess model Equations [10.2]–[10.4] with a specific growth rate µ expressed by the Haldane model (see Fig. 10.1). The control objective is then to design a controller with D as the control action such that the biomass production VX achieves its maximum at the end of the fed-batch operation. As explained in the preceding section, the maximization will be completed if the specific growth rate is kept at its maximum value: S*

KS K I

[10.59]

From the above considerations, we know that if the substrate concentration S can be stabilized at the set-point S* then the production of biomass is maximized. However, since the exact values of the Haldane model parameters KS, µ0 and KI, are usually unknown, the desired set-point S* is not available. In this work, an adaptive extremum-seeking algorithm is developed to search this unknown set-point such that the biomass production at the end of the reactor operation, that is, v(tf)X(tf) (with tf the final time of the fed-batch operation) is maximized. In the technical developments here below, we shall consider the following assumption for the parameters KS and KI of the Haldane model. Assumption: KS and KI are known to be bounded as follows: KS ,

KS ≤ KS ,

0

K I ≤ K I ,max .

This assumption is only important for the technical developments in order to avoid singularities in the extremum-seeking controller. It should not be interpreted as a ‘microbial’ constraint on the kinetic model. The design of the adaptive extremum-seeking controller will proceed in different steps. First of all, we shall start with the estimation equation for y, then include the controller equations and the estimation equations for the unknown parameters in a Lyapunov based derivation framework, and end up with the stability and convergence analysis arguments. Estimation equation for the gaseous outflow rate y Let us start with the parameter estimation algorithm for the unknown parameters KS, KI and µ0. The ratio of the yield coefficients (k1/k2) is assumed to be known. It follows from Equation [10.58] that:

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

µX =

y k2

249

[10.60]

Then Equations [10.2], [10.3] can be reformulated as follows: dX 1 = y DX dt k2

[10.61]

k dS = − 1 y D( S0 dt k2

S)

[10.62]

By considering Equations [10.58], [10.5] and [10.61], [10.62], the time derivative of y is equal to: dy = k2 dt

0X

KS (K KS

S 2 / KI S + S 2 K I )2

⎛ k1 ⎜⎝ − k y D(S0 2

⎞ 1 S )⎟ + k2 µ[ y DX ] k2 ⎠ [10.63]

Since the biomass concentration X is not accessible for on-line measurement, we reformulate dy/dt by replacing X with y/k2μ as follows:

(

)

Ks S KI dy = dt S K S + S S K I

(

(

))

⎛ k1 2 ⎜⎝ − k y + D(S0 2

⎞ S ) y + µ y − Dy ⎠

[10.64]

Equations [10.62] and [10.64] can be then rewritten as follows: dS = −θ k y D(S ( S0 dt

[10.65]

S)

1 − θi S 2 dy = ( D( S0 − S ) dt S (1 θ s S + i S 2 )

k

y)y +

θµ Sy 1 + θs S

i

S2

− uy

[10.66]

Let θ̂ denote the estimate of the true parameter θ, and ŷ be the prediction of y by using the estimated parameter θ̂. The predicted state ŷ is generated by the following equation: dyˆ 1 − θi S 2 =   dt S((

2

)

( ( S0 − S )

k

y)y +

θµ Sy − Dy + ky ey [10.67] 1 + θs S + θi S 2

with ky > 0 and the prediction error −ey = y − ŷ. It follows from [10.65] to [10.67] that:

© Woodhead Publishing Limited, 2013

250 Robotics and automation in the food industry dey dt

= − ky ey + −

1 − θi S 2 ( D( S0 − S ) s(1 θ s S + i S 2 )

1 − θi S 2 ( D( S0 − S ) S (1 θs S i S 2 )

k

k y)y −

y)y +

θµ Sy 1 + θ s S + θi S 2

θµ Sy 1 + θs S i S 2

[10.68]

Design of the adaptive extremum-seeking controller The desired set-point [10.59] can be re-expressed as follows: S* =

1

θi

Since the parameter θi is unknown, we design a controller to drive the substrate concentration S to: 1 θi that is, an estimate of the unknown optimum S*. An excitation signal is then designed and injected into the adaptive system such that the estimated parameter θ̂i converge to its true value. The extremum-seeking control objective can be achieved when the substrate concentration S is stabilized at the optimal operating point S*. Define the ‘control’ error zs: zs

S−

1 − d (t ) θi

[10.69]

where d(t) ϵ C1 is a dither signal that will be assigned later. The time derivative of zs is given by: dzs = D(S ( S0 − S ) dt

1  − 32 dθi  y + θi − d (t) t ) = Γ1 k 2 dt

[10.70]

We consider a Lyapunov function candidate: V=

2 2 2 zs2 1 ⎛ θ µ θ s θ i ⎞ + ⎜ + + ⎟+ 2 2 ⎝ γµ γs γi ⎠

(

+

sS

iS

ey2

)2

[10.71]

with constants γµ, γs, γi > 0. Taking the time derivative of V and substituting [10.65], [10.70] and [10.68] leads to:

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes (1 − θi S 2 ) dV θ µ dθ µ θ s dθ s θ i dθ i = z s Γ1 − − − + ey (D D( S0 dt γ µ dt γ s dt γ i dt S

251

S)

⎛ 1 − θ i S 2 − θ k ) + y θ µ Sy − e y ⎜ ( D( S0 − S ) − θ k y ) y ⎝ S ( + θ s + θ i S 2 ) +

⎞ (θ + 2θi S )(S0 − S )D ⎞ θ µ Sy (1 + θ s + θi S 2 ) − ey2 ⎡⎣ ky − s 2⎟   2(1 s S i S 2 ) ⎟⎠ 1 +θ s +θ iS ⎠

(

(1

s

S

i

1 S 2 ) + (θ s 2

⎤ 2θi S )θ k y ⎥ ⎦ [10.72]

Defining the functions: ⎛ ( + 2 i S )(S0 − S )D ⎞ Γ = −ey2 ⎜ ky − s (1 + 2(1 + s S + i S 2 ) ⎟⎠ ⎝

s

1 S + i S 2 ) − ey2 ( 2

s

2 i S )θ k y [10.73]

Ψi = Γ 3 D + Γ 4

[10.74]

Ψµ = ey Sy S

[10.75]

Ψs = Γ 5 D + Γ 6

[10.76]

where: ey S (1−  i S 2 )(S0 − S ) y 1 + θ s S +  i S 2

[10.77]

ey S (1− θ i S 2 k y 2 ey S 3θ µ y − 1 θ s S  i S 2 1 + θ s S + θ i S 2

[10.78]

Γ 3 = − e y S ( S0 − S ) y −

Γ 4 = ey S θ k y 2 +

Γ5 = −

ey (1− 1  i S 2 )(S0 − S ) y 1 + θ s S +  i S 2

ey (1− 1  i S 2 )θ k y 2 eyθ µ S 2 y Γ6 = − 1 θ s S  i S 2 1 + θ s S +  i S 2 We can write dV/dt as follows:

© Woodhead Publishing Limited, 2013

[10.79]

[10.80]

252 Robotics and automation in the food industry 1 − 3 dθ i  ⎤ ⎛ 1 dθ i ⎞  S ) − θ k y + θ i 2 − d (t )⎥ + Ψi − θi 2 dt γ i dt ⎟⎠ ⎦ ⎝

⎡ dV = z s ⎢ D( S dt ⎣

⎛ 1 dθ µ ⎞  ⎛ 1 dθ s ⎞  + ⎜ Ψµ − θµ + Ψs − θ s + Γ ⎟ γ µ dt ⎠ γ s dt ⎟⎠ ⎝ ⎝

[10.81]

For the solution of the extremum-seeking problem, we pose the dynamic state feedback: 1 − 3 dθ i d (t ) = a(t ) + θ i 2 − kd d (t ) 2 dt D=

1 S0

S

[ −k z

z s

+

k

y+a t

[10.82]

kd d t ] ,

kz > 0

[10.83]

where a(t) acts as a dither signal on the closed-loop process and kd is a strictly positive constant. Substitution of Equation [10.83] in [10.81] yields: . . . ⎛ ⎛ ⎛ i ⎞ µ ⎞ s ⎞ dV θ θ θ = − kz zs2 + ⎜ Ψi − ⎟ θ i + ⎜ Ψµ − ⎟ θ µ + ⎜ Ψs − ⎟ θ s Γ + Γ ⎜ ⎜ ⎜ dt γi ⎟ γµ ⎟ γs ⎟ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠

[10.84]

Using the definition of Ψi and Ψs, we express [10.84] as: . . . ⎛ ⎛ ⎛ i ⎞ µ ⎞ s ⎞ dV θ θ θ 2 = − kz zs + ⎜ Γ 3 D + Γ 4 − ⎟ θ i + ⎜ Ψµ − ⎟ θ µ + ⎜ Γ 5 D Γ 6 − ⎟ θ s Γ ⎜ ⎜ ⎜ dt γi ⎟ γµ ⎟ γs ⎟ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠

[10.85] We propose the following update law as follows: . ⎧γ ( θ i = ⎨ i ⎩0

),

i θ εi θ otherwisee

εi

⎧γ ( θ s = ⎨ s ⎩0

),

i θ ε s otherwisee

εs

.

 θˆ µ

θ

(

3

(

D + Γ4 ) > 0

5

D + Γ6 ) > 0

[10.86]

[10.87]

[10.88]

γ µ Ψµ

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

253

with the initial condition θ ̂s(0) ≥ εs = (1/KS,max) > 0, and θ ̂i(0) ≥ εi = (1/KS,maxKI,max) > 0. The update laws [10.86]–[10.87] are projection algorithms that ensure that θ ̂s(t) ≥ εs > 0 and θ ̂i(t) ≥ εi > 0. They also ensure that: ⎛ ⎜Γ D Γ 4 ⎜ 3 ⎝

. ⎞ θ i ⎟  θi γi ⎟ ⎠

⎛ ⎜Ψ ⎜ µ ⎝

. ⎞ θ µ ⎟  θµ γµ ⎟ ⎠

. ⎛  ⎞ ⎜ Γ D + Γ − θ s ⎟ θ ≤ 0, 6 s ⎜ 5 γs ⎟ ⎝ ⎠

[10.89]

Stability and convergence analysis Substituting the update laws [10.86]–[10.87] into [10.85], we obtain: dV ≤ − kz zs2 + Γ dt

[10.90]

We then assign the gain function ky such that the term Γ is negative. Using Equation [10.73], we consider: ky 0 +

ky

( S0 S ) | D | ε( S + )

[10.91]

with positive constants ky0 > 0 and 0 < ε 0, the dither signal is deemed sufficiently rich and its use can be justified on the closed-loop extremum-seeking control system. It is important to note that the corresponding conclusion depends entirely on the specific choice of parameter values that are used in the computations. More work is required to investigate how this assessment can be conducted in a manner that is invariant of the choice of parameter values. The use of this technique will be considered in the simulation study presented in the next section. Simulation results The aim of this section is to illustrate the performance of adaptive extremumseeking controller in a number of simulations, performed using a realistic example of a fed-batch process. The kinetic model parameters, yield coefficients and initial states used during numerical simulations are:

µ 0 = 00.53 53 h 1 , K S = 1.2 1 2 g/ g L, L, K I = 00.22 22 g /L L, k1 X( )

2 g/L, S( S ( ) = 2 g/L, S0

0 4 k2

20 g/L

1

[10.101]

For the Haldane model, from Fig. 10.1, the maximum on the growth specific . The control objective is to design a rate 1 occurs at S∗ = i = 0.52 g/L controller for the dilution rate, u, to regulate the substrate S at S*. The controller requires on-line measurements of the variables S and y, as well as the knowledge of the kinetic parameters, determining the S*. These values are obtained using the estimation algorithm previously presented, through the measurements of y. For the simulation study, we consider the following initial estimates of the kinetic parameters:

(

θˆ µ

1, θˆ S = 0.1, 1, θˆ I

)

3 ( µˆ 0 = 10, Kˆ S = 10, Kˆ I = 0..

© Woodhead Publishing Limited, 2013

)

[10.102]

256 Robotics and automation in the food industry The design parameters for the extremum-seeking controller are set to:

γµ

10 γ S = 200, 200 γ i

200 ky,0

20 kz

05 k

,0

1 ε = 0.2

The dither signal a(t) is chosen as follows: 5

a(t ) = ∑A i sin i =1

⎛⎛ . ⎝⎝

+

(

.001)i ⎞ ⎞ ⎟⎠ t ⎠ 4

5

∑A

2i

i =1

⎛⎛ (5 0.01)i ⎞ ⎞ cos ⎜ ⎜ 0.01 + ⎟⎠ t⎟ ⎝⎝ ⎠ 4 [10.103]

where A1i and A2i are normally distributed random numbers in the interval [−0.1, 0.1]. To test the richness of the dither signal, we computed the smallest eigenvalue of the matrix [10.99] over the interval [0, 100] using the initial conditions, x(0) = 7.2, s( ) 1 / 3 and parameter estimates given by Equation [10.102)] The simulation shows that the signal is sufficiently exciting in a region of these initial conditions and parameter values. The resulting dither signal was used in the subsequent simulation of the extremum-seeking control scheme. It should be noted that it is relatively easy in practice to provide a dither signal that means the matrix [10.99] is positive definite. However, the convergence of the parameter estimates can also depend on the conditioning of this matrix. The computations demonstrate that the condition number of the matrix remains around 103, a value that is relatively high. This indicates that parameter convergence may remain quite slow. In fact, a closer look at the spectral decomposition of this matrix indicates that the poor conditioning is associated with the adaptation of θs. The convergence properties of the extremum-seeking control scheme are shown in Fig. 10.7. We consider the initial conditions, x(0) = 7.2 and s(0) = 2.0. It is shown from Fig. 10.7 that the extremum-seeking scheme converges to the intended growth rate value. The substrate concentration converges to the unknown optimum as well. The dilution rate manipulation resulting from the extremumseeking control is also shown. The dilution rate is seen to reach its lower bound as a result of the dither signal, a(t). Convergence of the kinetic parameters to their true values is achieved. It is important to note that the control performance is strongly dependent on the convergence of the parameter θI, determining the setpoint for the control law, as illustrated by the substrate evolution, from Fig. 10.7. Overall, the extremum seeking is shown to perform satisfactorily for this case. In Krstic et al., an extremum-seeking scheme was proposed to optimize the production rate for a class of bioreactors governed by Monod kinetics and Haldane kinetics.34 An extremum-seeking control was developed following the design procedure proposed in reference 34. The main difference is that we use the procedure to optimize the growth rate µ(S). In this case, the extremum-seeking scheme yields the closed-loop system: x S

( µ D )x k1µx (S0 − S )D

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes 0.1

0.12

0.09

0.1

Dilution rate (h−1)

µ(S) ((g)−1h−1)

0.08 0.07 0.06 0.05 µ(S) µ(S)

0.04 0.00

0

10

20

30

40

50 60 70 Time (h)

80

257

0.08 0.06 0.04 0.02 0

90 100

0

10

20

30

40

50 60 70 Time (h)

80

90 100

6

0 2.5

Estimate θ Time θ

5

S Sπ

4 3

θ3θ1θµ

gL−1

2 1.5

2 1

1

0 0.5 0

−1 0

10

20

30

40

50 60 70 Time (h)

80

90 100

−2 0

10

20

30

40

50 60 70 Time (h)

80

90 100

Fig. 10.7 Illustration of the convergence properties loop system.

 Dˆ

k(µ

) + asi (

)

[10.104]

η ω h ( µ η) D = Dˆ + asi ( t ) where µ is given by [10.5]. Note that the structure assumes that the actual value of µ is measured in this case. This is extremely unlikely since this would require detailed structural information about the parameter values of µmax, Ks and Ki entering the expression for µ. In order to facilitate the comparison, the tuning parameters of the extremumseeking scheme are set to the following values: a

. , ω

. , k = 5,

h

= 0.1

The tuning parameters were assigned following the guidelines outlined in reference 30. Two simulations were performed. The first simulation was initiated at X(0) = 7.2, S(0) = 2.0, as above. Figure 10.8 shows the comparison between

© Woodhead Publishing Limited, 2013

0.1

0.1

0.09

0.09

0.08

0.08

µ(S) ((g)−1h−1)

µ(S) ((g)−1h−1)

258 Robotics and automation in the food industry

0.07 0.06 0.05

0.06 0.05 0.04

0.04 µ(S) µ(S)

0.02 0.00

0.07

0

20

40

60

µ(S) µ(S)

0.02 0.00

80 100 120 140 160 180 200 Time (h)

0.06

0

20

40

60

80 100 120 140 160 180 200 Time (h)

0 0

20

40

60

80 100 120 140 160 180 200 Time (h)

0.2 0.18 0.16

Dilution rate (h−1)

Dilution rate (h−1)

0.05 0.04 0.00 0.02

0.14 0.12 0.1 0.08 0.06 0.04

0.01

0.02 0

0

20

40

60

80 100 120 140 160 180 200 Time (h)

Fig. 10.8 Performance of the proposed scheme (right) and the controller proposed in reference 34 for the case X(0) = 7.2, S(0) = 2.0.

the performances of the two extremum-seeking schemes. On the left hand side, we show the dilution rate and specific growth rate µ for the proposed extremumseeking controller. On the right side, the dilution rate and specific growth rate resulting from the application of the extremum-seeking scheme proposed in reference 34. The results demonstrate that the two control schemes provide comparatively equivalent convergence properties. The scheme of reference 34 provides a slower convergence which could be improved by further adjustments of the design parameters. The values employed were the values that were found to provide the best performance for this system. In the second simulation, the initial concentration of biomass was set to X(0) = 1.2. The simulation results are shown in Fig. 10.9. As above, we show the dilution rate and specific growth rate for the proposed scheme on the right hand side and the results for the controller proposed by reference 34 on the left hand side. The scheme proposed in reference 34 provides very poor convergence properties compared to the scheme proposed here. Overall, the results indicate that the proposed scheme provides very consistent performance for this system. In the next set of simulations, we wish to illustrate the controller performance for regulation (as illustrated in Fig. 10.10) by introducing a disturbance in the S0 concentration between 60 and 70 h. As expected, the controller quickly rejects the

© Woodhead Publishing Limited, 2013

0.095 0.09

0.08

0.085

0.07

µ(S) ((g)−1h−1)

µ(S) ((g)−1h−1)

0.1 0.09

0.06 0.05 0.04

0.06

0.02

0.055

µ(S) µ(S)

0.01 0

0.045 0

50 100 150 200 250 300 350 400 450 500 Time (h)

0.06

0.12

0.05

0.1

0.04

0.08

0.03 0.02

50 100 150 200 250 300 350 400 450 500 Time (h)

0.06 0.04 0.02

0.01 0

µ(S) µ(S)

0.05

µ(S) ((g)−1h−1)

Dilution rate (h−1)

0.07

0.065

0.03

0

0.08

0.075

0

0

50 100 150 200 250 300 350 400 450 500 Time (h)

0

50 100 150 200 250 300 350 400 450 500 Time (h)

0.1

3

0.09

2.5

0.08

0.06

1.5 1

0.05

µ(S) µ(S)

0.04 0.00

s sπ

2

0.07

gL−1

µ(S) ((g)−1h−1)

Fig. 10.9 Performance of the proposed scheme (right) and the controller proposed in reference 34 for the case X(0) = 1.2, S(0) = 2.0.

0

10

20

30

40 50 60 Time (h)

70

80

0.5

90 100

0

0

10

20

40 50 60 Time (h)

70

80

30

40 50 60 Time (h)

70

80

90 100

0.2 0.18

Dilution rate (h−1)

0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0

0

10

20

30

90 100

Fig. 10.10 Illustration of the robustness properties. S0 changed from 20 to 40 g/L from t = 60 h to t = 70 h. © Woodhead Publishing Limited, 2013

260 Robotics and automation in the food industry 3

Substrate concentration (g−1)

0.1

µ(S) ((g)−1h−1)

0.09 0.08 0.07 0.06 0.05 µ(S) µ(S)

0.04 0.00

0

10

20

30

40

50 60 Time (h)

70

80

90 100

10

20

30

S Sπ

2.5 2 1.5 1 0.5 0

0

10

20

30

50 60 Time (h)

70

80

90 100

40

50 60 Time (h)

70

80

90 100

0.18 0.16

Dilution rate (h−1)

0.14 0.12 0.1 0.08 0.06 0.04 0.02 0

0

40

Fig. 10.11 Illustration of the robustness properties. The measurements of S and y are corrupted by additive noise. S0 is changed from 20 to 40 g/L from t = 60 h to t = 70 h.

disturbance while attenuating its effect on S which is accurately regulated at the set-point during the whole operation. The last simulation illustrates the performance of the extremum-seeking controller in the presence of noisy measurements. As in the previous case, we consider the operation of the bioreactor in the presence of a step in inlet substrate concentration, S0 from 20 to 40 g/L from t = 60h to t = 70 h. In addition, it is assumed that the substrate measurement, S, and the production rate measurement, y, are corrupted by additive measurement noise. The measured substrate concentration is given by Sm = S +0.1nS where nS is a unit variance normally distributed additive noise term. Similarly, the measured production rate is given by, ym = y + 0.005ny, where ny is a unit variance normally distributed additive noise term. The results of the simulation are given in Fig. 10.11. The results demonstrate that the extremum-seeking controller performs adequately in the presence of measurement noise.

10.7 Acknowledgements This chapter presents research results of the Belgian Network DYSCO (Dynamical Systems, Control, and Optimization), funded by the Interuniversity Attraction

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

261

Poles Programme, initiated by the Belgian State, Science Policy Office. The scientific responsibility rests with its author.

10.8 Conclusion Food processes present a large diversity of applications, and proposing a unified approach for the control of these processes is still clearly a challenging nowadays. This explains why this chapter has concentrated on drawing some key ideas for the design and implementation of monitoring and control of food processes on the basis of a class of food processes (bioconversion) and within this class more specifically on fed-batch processes. In Section 10.3, we have been concentrating on modelling issues by considering the importance today of considering the particulate dimension (via the distribution of size, mass or age) in food processes by briefly introducing the use of population balance to complement mass balance modelling and address some of the complex related to these partial differential equations’ models. The following sections had been dealing with several key issues related to fed-batch processes that are playing a central role in the bioconversion processes of the food industry today. In Section 10.4, the monitoring of fed-batch processes had been addressed by considering the tuning of observerbased estimators for unknown parameters that are able to handle the time-varying nature of the fed-batch process dynamics. This issue is also the object of Section 10.5, which has concentrated on the specificities of the design of PID controllers for fed-batch reactors. Section 10.6 has finally considered the design of real-time optimization algorithms, namely adaptive extremum-seeking controllers, for fedbatch processes.

10.9 References 1. Bruin, S., and Th. R.G. Jongen (2003). Food process engineering: the last 25 years and challenges ahead. Comprehensive Reviews in Food Science and Technology, 2, 42. 2. CAFE (Computer-Aided Food processes for control Engineering) (2008–2012). FP7 EC project. http://www.cafe-project.org/. 3. Braatz R.D. and S. Hasebe (2002). Particle size and shape control in crystallization processes. 6th International Conference on Chemical Process Control, AIChE Symposium. Series, 98, 307–327. 4. Benkhelifa, H., A. Haddad Amamou, G. Alvarez and D. Flick (2008). Modelling fluid flow, heat transfer and crystallization in a scraped surface heat exchanger. International Symposium on Applications of Modelling as an Innovative Technology in the Agri-Food-Chain. Model-IT 802: 163170. 5. Benkhelifa H., M. Arellano, G. Alvarez and D. FlickIce (2011). Crystals nucleation, growth and breakage modelling in a scraped surface heat exchanger. Proc. ICEF 2011, Athens, 4 pages.

© Woodhead Publishing Limited, 2013

262 Robotics and automation in the food industry 6. Vilas C., E. Balsa -Canto, M. Arellano, H. Benkhelifa, G. Alvarez, D. Flick, D. Leducq and A. Alonso (2011). Model Identification of the Ice-Cream Crystallization Process. Proc. ICEF 2011, Athens, 4 pages. 7. Casenave C., D. Dochain, G. Alvarez, H. Benkhelifa, D. Flick and D. Leducq (2012). Steady-state and stability analysis of a population balance based nonlinear ice cream crystallization model. Proc. ACC, Montreal, 6461–6466. 8. Henri C., B. Haut, D. Dochain and V. Halloin (2006). Modeling of an industrial agglomeration process. Proc. ISCRE 19, 3–6 September, Posdam/Berlin, Germany, 869–870. 9. Daoutidis P. and M.A. Henson (2002). Dynamics and control of cell populations in continuous bioreactors. 6th International Conference on Chemical Process Control, AIChE Symposium Series, 98, 274–289. 10. Ramkrishna D. (2000). Population Balances. Theory and Applications to particulate Systems in Engineering, Academic Press, San Diego. 11. Mantzaris N.V. and P. Daoutidis, (2004). Cell population balance modeling and control in continuous bioreactors. J. Process. Control, 14, 775–784. 12. Mantzaris N.V., J.J. Lio, P. Daoutidis and F. Srienc (1999). Numerical solution of a mass structured cell population balance in an environment of changing substrate concentration. J. Biotech., 71, 157–174. 13. Mantzaris N.V., F. Srienc and P. Daoutidis, (2002). Nonlinear productivity control using a multi-staged cell population balance model. Chem. Eng. Sci., 57, 1–14. 14. Beniich N. and D. Dochain (2009). Design and analysis of a nonlinear stabilizing controller for a mass structured cell population balance model with input constraints. Proc ECC09, 1884–1888. 15. Bastin G. and D. Dochain (1990). On-line Estimation and Adaptive Control of Bioreactors. Elsevier, Amsterdam. 16. Perrier M. and D. Dochain (1993). Evaluation of control strategies for anaerobic digestion processes. Int. J. Adaptive Cont. Signal Process., 7 (4), 309–321. 17. Pomerleau Y. and M. Perrier (1990). Estimation of multiple specific growth rates in bioprocesses. AIChE J., 27, 231–236. 18. Oliveira R., E.C. Ferreira, F. Oliveira and S. Feyo de Azevedo (1996). A study on the convergence of observer-based kinetics estimators in stirred tank bioreactors. J. Proc. Control, 6 (6), 367–371. 19. Sonnleitner B. and O. Kappeli (1986). Growth of Saccharomyces cerevisiae is controlled by its limited respiratory capacity: formulation and verification of a hypothesis. Biotechnol. Bioeng., 18, 927–937. 20. Bryson A.E. and Y.C. Ho (1975). Applied Optimal Control. John Wiley, New York. 21. Axelsson J.P. (1989). Modelling and Control of Fermentation Processes, PhD thesis, Lund Institute of Technology, Suède. 22. Astrom K.J. and B. Wittenmark (1995). Adaptive Control, 2nd edition, Addison-Wesley, Reading, MA. 23. Goodwin G.C. and K.S. Sin (1984). Adaptive Filtering Prediction and Control. Prentice-Hall, Englewood Cliffs, NJ. 24. Krstic M., I. Kanellakopoulos and P. Kokotovic (1995). Nonlinear and Adaptive Control Design, John Wiley, New York. 25. Narendra K.S. and A. M. Annaswamy (1989). Stable Adaptive Systems, Prentice-Hall, Englewood Cliffs, NJ. 26. Leblanc M. (1922). Sur l’´electrification des chemins de fer au moyen de courants alternatifs de fr´equence e´lev´ee, Revue Ge´ne´rale de l’Electricite´, 12(8), 275–277. 27. Drkunov S., U. Ozguner, P. Dix and B. Ashrafi (1995). ABS control using optimum search via sliding modes, IEEE Trans. Contr. Syst. Tech., 3, 79–85. 28. Sternby J. (1980). Extremum control systems: An area for adaptive control? Preprints of the Joint American Control Conference, San Francisco, CA.

© Woodhead Publishing Limited, 2013

Advanced methods for the control of food processes

263

29. Vasu G. (1957). Experiments with optimizing controls applied to rapid control of engine presses with high amplitude noise signals. Trans. ASME, 481–488. 30. Krstic M. (2000). Performance improvement and limitations in extremum seeking control, Systems & Control Letters, 5, 313–326. 31. Krstic M. and H.H. Wang (2000). Stability of extremum seeking feedback for general dynamic systems. Automatica, 4, 595–601. 32. Krstic M. and H. Deng (1998). Stabilization of Nonlinear Uncertain Systems, Springer-Verlag, New York. 33. Li, P.Y. (1995). Self Optimizing Control and Passive Velocity Filed Control of Intelligent Machines, Ph.D. Thesis, U.C. Berkeley. 34. Wang H., M. Krstic and G. Bastin (1999). Optimizing bioreactors by extremum seeking. Int. J. Adap. Control Signal Proc., 13, 651–669. 35. Guay M. and T. Zhang (2003). Adaptive extremum seeking control of nonlinear dynamic systems with parametric uncertainties. Automatica, 39(7), 1283–1293. 36. Titica M., D. Dochain and M. Guay (2003). Real-time optimisation of fed-batch bioreactors via adaptive extremum seeking control. Chem. Eng. Res. Design, 81 (A9), 1289–1295. 37. Titica M., D. Dochain and M. Guay (2003). Adaptive extremum seeking control of fedbatch bioreactors. Euro. J. Control, 9 (6), 618–631. 38. Guay M., D. Dochain and M. Perrier (2004). Adaptive extremum seeking control of stirred tank bioreactors. Automatica, 40 (5), 881–888. 39. Harmand J, D. Dochain and M. Guay (2006). Dynamical optimization of a configuration of multi-fed interconnected bioreactors by optimum seeking. Proc. CDC06, 2122–2127. 40. Marcos N., M. Guay, D. Dochain and T. Zhang (2003). Adaptive extremum seeking control of a continuous bioreactor. J. Process Control, 14 (3), 317–328. 41. Marcos N., M. Guay and D. Dochain (2004). Output feedback adaptive extremum seeking control of a continuous stirred tank bioreactor with Monods kinetics. J. Process Control, Special issue on Dynamics, Monitoring, Control and Optimization of Biological Systems, 14 (7), 807–818. 42. Zhang T., M. Guay and D. Dochain (2003). Adaptive extremum seeking control of continuous stirred tank bioreactors. AIChE J., 49 (1), 113–123. 43. Cougnon P., D. Dochain, M. Guay and M. Perrier (2006). Real-time optimization of a tubular reactor with distributed feed. AIChE Journal, 52(6), 2120–2128. 44. Guay M., D. Dochain and M. Perrier (2005). Adaptive extremum seeking control of nonisothermal CSTR. Chem. Eng. Science, 60 (13), 3671–3681. 45. Hudon N., M. Perrier, M. Guay and D. Dochain (2004). Adaptive extremum seeking control of a non-isothermal tubular reactor with unknown kinetics. Comp. Chem. Eng., 29 (4), 839–849. 46. Hudon N., M. Guay, M. Perrier and D. Dochain (2008). Adaptive extremum-seeking control of convection-reaction distributed reactor with limited actuation. Comp. Chem. Eng., 32(12), 2994–3001. 47. Favache A., D. Dochain M. Perrier and M. Guay (2008). Extremum seeking control of retention for a microparticulate system. Can. J. Chem. Eng, 86, 815–827. 48. Guay M., D. Dochain, M. Perrier and N. Hudon (2007). Flatness-based extremum seeking control over periodic orbits. IEEE Trans. Aut. Control, 52 (10), 2005–2012. 49. Ohno H., E. Nakanishi and T. Takamatsu (1976). Optimal control of a semi-batch fermentation. Biotechnol. Bioeng., 28, 847–864. 50. Cheruy A. and A. Durand (1979). Optimization of Erytromycin biosynthesis by controlling pH and temperature: theoretical aspects and practical applications, Biotechnol. Bioeng., 9, 303–320. 51. Parulekar S.J., J.M. Modak and H.C. Lim (1985). Optimal control of fed-batch bioreactors. Proc. ACC, Boston, 2, 849–854.

© Woodhead Publishing Limited, 2013

264 Robotics and automation in the food industry 52. Peringer P. and H.T. Blachere (1979). Modelling and optimal control of baker’s yeast production in repetead fed-batch culture. Biotechnol. Bioeng., 9, 205–213. 53. Dochain D. and G. Bastin (1989). Adaptive control of fedbatch bioreactors. Chemical Engineering Communications, 87, 67–85.

© Woodhead Publishing Limited, 2013

11 Automation and robotics for bulk sorting in the food industry G. Hamid, B. Deefholts, N. Reynolds, D. McCambridge, K. Mason-Palmer and C. Briggs, Buhler Sortex Limited, UK

DOI: 10.1533/9780857095763.2.267 Abstract: This chapter describes the use of optical sorting machines in the bulk food industry for ensuring food safety and food quality. The chapter describes recent technological innovations and the drivers behind these innovations. Case study examples are included from the rice, grain, nut, fruit and vegetable sectors. The chapter ends by highlighting the challenge of sustainable intensification of global agriculture that will drive future innovations in optical sorting for the bulk food industry. Key words: optical sorting, bulk foods, food safety, food quality.

11.1 Introduction The primary role of sorting is to control product quality. The sorting of food commodities such as rice requires the processing of many tonnes per hour (t/h) in a continuous round the clock operation. The sorting of such vast quantities on a commercial scale is made feasible using an automated technique known as optical sorting. Over the past few years there have been a series of step changes in the performance of optical sorting machines. This chapter contains a description of the recent technological innovations in optical sorting as well as the drivers that motivated these changes and some example case studies of the resulting benefits to the food industry. In the context of optical sorting, the quality of the food stream is defined in terms of maximum permissible levels for different categories of defects. An optical sorting machine automatically identifies these defects and removes them from the product stream. The categories of defects include gross contaminants such as glass, stones, insects, pieces of shell, rotten product or extraneous vegetable

Published by Woodhead Publishing Limited, 2013

268 Robotics and automation in the food industry matter, and more subtle contaminants such as blemished, discoloured or misshaped pieces of product. For the gross contaminants that are hazardous to health a zero tolerance is usually applied, that is, all such defects have to be removed. For the more subtle defects the aim is to remove a sufficient percentage such that the food stream meets a required grade or quality standard. Optical sorting machines use either cameras or lasers for the identification of defects and typically use air ejectors for the removal of the unwanted items. Optical sorting is well suited to the food industry because most defects can be identified by visual appearance. Furthermore, both optical sensors and air ejectors are non-contact (i.e. hygienic) and enable high rates of processing (i.e. high capacity). Optical sorting is a mature technology (Bee and Honeywood, 2002, 2007). In the early 1930s, E. H. Bickley invented an optical sorter in the USA for sorting beans (Eisinger, 1999). Bickney’s own company continued to improve the sorter for the next 30 years, adapting it to sort rice, peanuts and even ball bearings. Meanwhile in the UK, another company founded in 1947 and now known as Buhler Sortex Ltd was developing its own solution to automate the laborious process of hand-sorting seeds. Today, Buhler Sortex Ltd designs and manufactures optical sorting machines for sorting a diverse range of commodities, including rice, grain, pulses, coffee, nuts, fruit and vegetables. Recently there have been a number of technological advances in optical sorting. Many of these advances have been made possible by the transfer of technology from other industries, but the drivers for all these advances have come directly from the specific requirements of the food industry. The layout of the rest of this chapter is as follows: Section 11.2 comprises an overview of the basic principles of operation of an optical sorting machine. Section 11.3 lists the requirements of the bulk food industry that have in the past motivated advances in optical sorting – and continue to do so. Section 11.4 contains a description of recent technological innovations in optical sorting and Section 11.5 illustrates the benefits to the food industry of these recent innovations with a number of example case studies. Finally, in Section 11.6, we attempt to identify some trends in the food industry that will drive innovations in optical sorting in the future.

11.2 Principles of operation This section contains a description of the basic principles of operation of an optical sorting machine.

11.2.1 Overview There is a wide variety of optical sorting machines. Figures 11.1 and 11.2 are photographs of two different types of optical sorters designed for two different sectors of the food industry. Even though these two machines look radically different they

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 269

Fig. 11.1 An optical sorting machine designed for sorting dry products.

share the same basic principles of operation, which indeed apply to all optical sorting machines. An optical sorting machine consists of four principal systems: a feed mechanism, an optical system, an image processor and an ejection system. This section contains a description of these four systems, using the example of the SORTEX Z+ machine shown in Fig. 11.1 and the accompanying schematic layout shown in Fig. 11.3.

11.2.2 Feed mechanism The product stream of food passes through the machine in a continuous flow. The purpose of the feed mechanism is to present this product stream in a uniform manner past the optical system and then past the ejection system. The flow must be uniform both in the distribution of product and in the velocity of the product. A uniform distribution of product minimises the clumping and overlapping of the items of food. This is necessary because the optical system cannot detect defects that are occluded by overlapping items of product. Moreover, excessive clumping results in too much good product being inadvertently rejected with the defects. A uniform velocity of product is required because the timing of the ejection system is set as a constant delay after the time the product has passed the line of sight of the inspection system. In the schematic diagram shown in Fig. 11.3 the

Published by Woodhead Publishing Limited, 2013

270 Robotics and automation in the food industry

Fig. 11.2 An optical sorting machine designed for sorting wet products.

Input hopper

In-feed vibrator

Chute

IR camera

Camera

Foreground lighting

Background

Reject receptacle

Ejector Accept receptacle

Fig. 11.3 A schematic diagram of the main components of the machine shown in Fig. 11.1.

feed mechanism comprises an input hopper, an in-feed vibrator and a chute. The product is loaded into the hopper and the in-feed vibrator shakes the product onto the chute. The product stream accelerates down the chute under gravity. After the product stream has left the end of the chute it passes through the line of sight of the optical system and then past the line of fire of the ejection system.

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 271 There are many alternatives to this gravity-fed chute feed mechanism. For example, optical sorting machines for wet products often comprise a conveyer instead of a chute.

11.2.3 Optical systems Some optical sorting machines inspect the product as it is being carried by the feed mechanism. However, it is more usual to inspect the product stream after it has left the feed mechanism and is flying through the air. The advantage of inspecting the product in free flight is that the product can be inspected from two sides. Figure 11.3 illustrates an optical system with cameras that view the product from both the front (top side of the product) and the rear (underside of the product). The optical system is housed within one or more optical boxes with the cameras and lighting located behind toughened glass windows. The product stream does not come into contact with any part of the optical box. The purpose of the optical system is to capture one or more images of every item in the product stream and to ensure that each item is imaged in a similar manner, avoiding variations in illumination, shading or occlusion. These images are transmitted to the image processor. Optical sorting covers inspection systems that capture images using single or multiple parts of the light spectrum including infrared (IR) and ultra-violet (UV) wavelengths. The basic principle is to illuminate the product stream with light and capture an image of the reflected and/or transmitted light from each item. The optical system is usually characterised in terms of the wavelengths of light to which it is sensitive and the spatial resolution of the image sensor. The sorter shown in Fig. 11.1 uses a mix of fluorescent tubes and incandescent lamps for the foreground lighting to illuminate the product stream. The product is usually viewed against a background of known appearance, such as a white background. In optical sorting it is common to use line-scan cameras which capture images by concatenating successive single line images of the product stream. This approach avoids any synchronisation issues that arise with using area-based cameras. Common types of illumination include fluorescent tubes, metal halide lamps, light emitting diodes (LEDs) and lasers.

11.2.4 Image processor The role of the image processor is to identify the location of the defects from the image data and to output this information to the ejection system. A high rate of processing image data is required in order to detect the defects in a time period less than that taken for the product stream to travel between the cameras and the ejectors. This time delay is typically a few milliseconds. The image processor encodes the sorting criteria (algorithms) of the optical sorter. The image processor is typically implemented in specialised hardware, such as Digital Signal Processor (DSP) or Field Programmable Gate Array (FPGA) chips. The image processor usually includes some parameters (settings) that are adjusted by the

Published by Woodhead Publishing Limited, 2013

272 Robotics and automation in the food industry operator via a Graphical User Interface (GUI). This GUI is typically a touch screen display.

11.2.5 Ejection system The role of the ejection system is to separate the unwanted items from good product. As mentioned above, the ejection is typically performed while the product is in free fall; accept items are allowed to continue along their normal trajectory, whereas the defects are deflected into a separate reject stream. Precise aiming and timing of the ejectors is required to minimise the loss of good product inadvertently rejected while removing the defects. In Fig. 11.3, the ejection mechanism consists of an array of ejectors arranged side by side across the width of the machine (only one ejection shown in cross-section) and two separate receptacles for the accept and reject streams. Each ejector can be fired independently, emitting a short burst of compressed air through a nozzle of roughly 5 mm diameter. For small products, such as rice, the firing of one or two neighbouring ejectors is sufficient to remove a single defect. For larger items, such as carrot slices, multiple ejectors are fired simultaneously across the width of the machine to ensure that the unwanted items are removed. For very large items, such as potatoes, alternative ejection systems such as flaps (flippers) are used to deflect the defects.

11.3 Requirements The main requirements for optical sorting for the bulk food industry are high capacity, high efficiency and high yield. A prerequisite is that an optical sorting machine must be fit for purpose in situ in the food processing line. The environmental conditions of bulk food processing lines are varied, often extreme, and in some cases hazardous. The design of a sorting machine and its installation in the line must be well suited to the nature of the product stream. In any installation, the machine must minimise damage to the product. The individual items of food require the appropriate mechanical handling. In addition, the machine must be sufficiently robust to the nature of a continuous bulk flow of the product stream, whether wet, abrasive or high oil content. A requirement for the bulk food industry is high capacity. The trend in the industry has always been for increased capacity, so that each new generation of machines has greater throughput of product for the same footprint. The capacity of a sorting machine is typically measured in t/h. In addition, any new technology should be more efficient and increase the yield of the food processing line. In terms of optical sorting, efficiency is measured by the percentage of defects removed from the food stream. For example, if the contamination of defects within the food stream at the input to the sorter is 2% and the contamination of defects remaining in the final accept stream at the output of the sorter is 0.2% then the overall sorting efficiency is 90%. The yield is the net output of the machine and is often

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 273 expressed as a percentage by mass. For example the output stream maybe 96% of the input stream, that is, 4% of the input has been rejected. Another useful measure of the performance of a sorting machine is the reject ratio. The reject ratio is a simple measure of what is being rejected by the sorter. It follows the adage of assessing a process by looking at what is being thrown away. The reject ratio is the ratio of the number of good items to the number of defects in the reject stream. In practice, the reject ratio is estimated from counts of the numbers of good and defect items that have been rejected. In the above example, a rough estimate of the reject ratio would be one good item of product to each defect. This rough estimate is based on the assumption that all the defects have been removed (2% of input) and so comprising half the reject stream (4% of input). The requirements of capacity, efficiency and yield are all inter-related. In most sorting applications one can be traded against the other two. For example, decreasing the capacity will tend to increase the efficiency and yield, whereas decreasing the efficiency while keeping the capacity constant will increase the yield. Hence, the net benefit of an innovation in optical sorting should be measured by looking at the complete picture of capacity, efficiency and yield achieved in situ in the food processing line. Another main driver is cost innovation (Zeng and Williamson, 2007), that is innovation that maintains or improves the machine quality but also reduces the machine cost or cost of ownership. There are, of course, some innovations that cannot be measured directly in these terms, such as ease of use or noise reduction. But as a general rule, capacity, efficiency, yield and cost are the main drivers for innovation of optical sorting machines for the food industry.

11.4 Recent advances in technology The basic principles of optical sorting machines have remained the same over the years but the performance in terms of efficiency, yield and capacity has undergone a number of step changes. This section highlights a number of advances in technology that have enabled recent increases in performance.

11.4.1 New materials Advances in thermoplastic materials are providing opportunities for manufacturers to increase product life and recyclability without necessarily increasing product cost. Advances in material science have been used to increase both the speed of response and the lifetime of the air ejectors. New surface conversion processes and material coatings are improving the resilience of machine components to the erosion and the corrosion commonly seen within the food industry. In the future nano-technology will become more important, with the creation of adaptive surface coatings that can be modified to improve feed mechanism or be formulated to reduce friction and reduce wear. One such area is the utilisation of carbon nano-

Published by Woodhead Publishing Limited, 2013

274 Robotics and automation in the food industry tubes to form a virtually friction free surface; although there are potential health issues, linked to nano-particle size, that must be considered.

11.4.2 Advances in sensors The sensors of the first optical sorting machines were single photoelectric cells. Today, the typical sensor inside the camera of an optical sorting machine is a high resolution line-scan CCD array. The resolution of the cameras has been steadily increasing, enabling the detection of finer and finer spot defects. One of the major advances in sensing technology for the detection of defects is the use of IR. IR for optical sorting is not new. The first IR sorts were restricted to the near infrared (NIR), which is 700–1000 nm. However, Indium Gallium Arsenide (InGaAs) technology recently introduced for optical sorting, which was originally devised for space and military applications, has far superior performance. This superior performance derives from the extension of the wavelengths of IR into the short wave range, which is 1000–1700 nm. It is common to use the term SWIR to distinguish short wave IR from near infrared. Another benefit of InGaAs technology is the faster sensor response, which enables higher resolution sensors. Furthermore the original IR sensors were monochromatic, that is sensitive to just one band of IR wavelengths. Today optical sorting machines employ a technique known as bichromatic SWIR.

11.4.3 Advances in data processing The electronics industry continues to develop bigger and faster integrated circuits (chips). The specialised hardware of the image processor is typically realised in high performance DSP chips or FPGA chips. The amount of computational processing power available within each new generation of chips has continued to increase exponentially in accordance with Moore’s law. This increase in computational power has enabled the realisation of progressively more sophisticated algorithms in optical sorting machines for the detection of defects. In the rice sector, this increased computational power is used to implement multiple algorithms with each algorithm tailored specifically to the detection of a particular type of defect, such as peck (insect bite) defects or discoloured yellow grains. In the nut, fruit and vegetable sectors, this increased computational power has enabled significant improvements with the increasing sophistication of the shape sorting algorithms. Shape sorting is the sorting of product according to the shape of the individual product pieces. One example of these enhanced shape sorting algorithms is a new feature of the image processor to automatically separate a cluster of touching product. Figure 11.4a and 11.4b illustrate the basic concept. Figure 11.4a contains the silhouette (outline) of a clump of touching beans. This is the image captured by the camera of an optical sorting machine. In this figure the colour information has not been shown so that only the silhouette (or shape) of the object is displayed. Note that the beans are touching as they pass the line of sight of the cameras and so are perceived as one larger clump. Figure 11.4b shows the

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 275 (a)

(b)

Fig. 11.4 (a) The silhouette of a clump of beans. (b) The result of applying object separation to the image shown in Fig. 11.4a.

same image after applying an algorithm known as object separation to the image data. In Fig. 11.4b, automatic cuts have been made to the image, separating the individual product pieces. The advantage of object separation is that it allows both the sorting by shape of the individual objects even when the objects are touching and greater precision in aiming the ejectors at individual items rather than removing a whole clump of product.

11.5 Current applications This section provides a further illustration of recent advances in the performance of optical sorting machines. In this section the emphasis is on applications rather

Published by Woodhead Publishing Limited, 2013

276 Robotics and automation in the food industry than on the specific technology per se. The section is organised in three subsections with examples from the grain, nut and fruit and vegetable sectors of the food industry.

11.5.1 Grain Rice was the first grain that was colour sorted commercially; it is still the largest market for optical sorting machines. Figure 11.5 shows an installation of six optical sorting machines in a rice mill. The majority of rice is colour sorted after milling when the bran layer has been removed and the surface of the grain has been polished. The major colours that are considered defects are yellow grains, peck (insect damaged grains that have a black or brown spot), chalky kernels, under milled grains, paddy and mud balls. Sorters are also used to sort brown rice (un-milled rice) to remove green immature grains, dark defects and some foreign material such as mud balls or weed seeds that may be in the product. Most exporting countries have standards for different grades of rice and these are largely based on the grades published by the USDA (2009). In several countries, rice which has been colour sorted to a specified standard is classed as ‘sortexed’ rice and it is normal for a higher price to apply to that rice. The product going into a rice colour sorter will typically have a defect level in the range of between 2% and 5%. In order to increase the capacity of the sorter it is normal practice to run the rice through a first pass at a higher capacity than would allow single grains to be rejected, at settings that provide an acceptable product that meets the required standard. The rejected grains from this first pass sort are then reprocessed on another part of the same sorter running at a lower capacity

Fig. 11.5 An installation of six optical sorting machines in a rice mill.

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 277 and so with better separation of the rice grains. This ‘re-sorting’ improves the yield; a typical reject ratio at the reject output of the second pass is at least two defective grains for each good grain rejected. It is now becoming common practice to sort the reject from this second pass to reclaim the good grains. The input to this third pass comprises more defect grains than good grains; hence it is more efficient to separate the good grains from the defect grains by aiming the ejectors at the good grains – a technique known as reverse sorting. The layout of this three pass sort is shown in Fig. 11.6. Using these techniques a very high quality of accept product is provided while producing a reject with a ratio of at least ten bad grains for each good grain removed. Figure 11.7a, 11.7b and 11.7c show samples taken from sorting rice at the input, accept and final reject, respectively. The trend in the rice milling industry is for larger milling and reprocessing lines. This demand has lead to an increase in capacity for colour sorters from roughly 3 t/h in the early 1980s to sorters with capacities of 15 t/h today. This has been achieved through a combination of advances, including: improvements in detector technology; faster valves used in the air ejectors for removing the defect grains; improvements in sorting algorithms and also in improvements in the feeding systems for providing a more uniform distribution of the grains past the optical system. A rice sorter today, typically having cameras viewing grains from two sides and over a width of 30 cm, will view approximately 50 000 grains of rice every second. The first colour sorters for rice were adopted by millers in the USA, Europe and then Japan but in the 1980s the practice was also adopted by millers in Thailand and India, mainly for exported rice, which attracted a premium price if colour sorted. Today in countries such as India and China many High capacity in-feed

Elevator

Elevator

Second pass

First pass

Concentrated defect

Defect

High quality, high yield, accept

Third pass

Good product fed back to in-feed

Highly concentrated final reject

Fig. 11.6 The layout for a three-pass sort of rice.

Published by Woodhead Publishing Limited, 2013

Good product fed back to second pass

278 Robotics and automation in the food industry (a)

(b)

(c)

Fig. 11.7 (a) Sample of input to rice sort. (b) Sample of accept from rice sort. (c) Sample of reject from rice sort.

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 279

Fig. 11.8 Typical examples of the defect categories found in wheat.

of the sorters are now used to process rice for the domestic market as well as for the export market. This has seen a huge growth in the number of sorters being used worldwide. Sorters have also been used for sorting wheat and rye particularly in speciality applications. It has been common to sort rye to remove ergot and to sort wheat that would be used for breakfast cereals or similar applications, but the low capacity of sorters had previously prevented their widespread use in the cleaning sections of flour mills. Traditionally, wheat is processed through a cleaning section through a number of different mechanical cleaners, each optimised for a different type of defect, to remove foreign seeds and diseased grains before milling into flour. Increasingly stringent food regulations governing the safety and quality of end products, together with the increased capacity and reliability of colour sorters, has led to their adoption in the cleaning sections of European flour mills. This has normally resulted in a reduction of operational costs because of lower energy consumption, together with an increased yield through lower wastage. Figure 11.8 shows some typical examples of defect categories found in wheat. Optical sorters have several advantages over their mechanical alternatives: they use less energy, they need less adjustment if the physical properties of wheat change slightly, and they can leave in broken wheat grains if desired while removing weed seeds to give an increase of yield. The defect levels in wheat are lower than those in rice. Modern sorters, similar to the machines used for rice, will have capacities of over 30 t/h when sorting wheat. Another area that is being addressed by the latest colour sorters is the need to meet regulatory requirements demanded by the harsh, dusty environments in some wheat applications. These require the sorters to be certified for operation in a hazardous location (HAZLOC) or in an area with an explosive atmosphere (ATEX).

11.5.2 Nut sector This section contains case studies of two high-value dry commodities, namely almonds and walnuts. Almonds The almond has one of the largest production levels in the nut sector, with an annual harvest greater than 800 000 t. The cleaning of almonds consists of the following stages: hulling, shelling, screening, aspiration, de-stoning and colour sorting. In the past sorting was predominantly a manual task involving large teams

Published by Woodhead Publishing Limited, 2013

280 Robotics and automation in the food industry of hand pickers to remove the most obvious defects such as dark, rotten, pieces of shell, hull and foreign material. With an ever growing market hand picking alone became too costly and time consuming an exercise. The introduction of optical sorters revolutionised the industry by allowing high rates of production to be achieved with only a few hand pickers giving the final check before the product was shipped. As the market grew, so did the demand for safer, cleaner, contamination free product. The older optical sorting machines, although good at removing the obvious defects, began to fall short of the new demands for removing the more difficult to detect defects such as curly hull, in-shell and hard shell, as shown in Fig. 11.9. The introduction of IR sensors began to improve the detection of the more difficult to remove defects. The first had capacities of 500 kg/h and could usually clean the product to 98–99% leaving hand pickers to remove the last 1–2%. As production continued to increase and labour costs spiralled upwards the demand for faster and more accurate machines followed suit. The latest multi channel optical sorting machines are capable of sorting at capacities of 6–8 t/h. This increase in capacity also required faster and more accurate ejection systems so that a high yield could be maintained (ejectors with a slow response tend to remove unacceptable amounts of good product). Advances in the ejection system have produced ejectors with a response time of around 100 μs. These new ejectors are able to remove defects with a greater accuracy reducing the loss of good product. In addition, the introduction of indium gallium arsenide (InGaAs) sensor technology and shape sorting technology has greatly improved the efficiency of detection of defects.

(a)

(b)

(c)

(d)

Fig. 11.9 Examples of good almonds and defects for an almond sorting application. (a) good almonds, (b) hull pieces, (c) in-shell and (d) hard shell.

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 281

Fig. 11.10 Examples of walnuts, walnut shell pieces, and walnut septa.

Walnuts Walnuts, like many tree nuts come with an outer skin (hull) that has to be removed to gain access to the shell and then in turn the meat of the nut inside. The process required to gain access to the nut kernel involves large de-hulling, shelling (cracking), flotation tanks and high-pressure-water cleaning equipment. The nuts are passed through flotation tanks to remove any heavy orchard debris such as stones and mud ball. After this process the nuts are then passed through the hullers. The huller usually consists of a large flat conveying bed with hard brushes and knives plus plenty of water. The nuts enter one end of the equipment and are carried on a belt between brushes and knives to pull off the hull. After this process the walnuts are then passed over an agitator with high pressure wash systems to remove any remaining hull and membrane. Walnuts are sold either in-shell or with the shell removed. Removing the shell on a commercial scale has its pitfalls because the main value is in walnut halves and during the cracking process many of the nuts become broken leaving a sizable amount of product that can be between 2 and 15 mm in size. Regardless of size all of the product will contain a certain amount of shell and membrane (defects) that needs to be removed. Figure 11.10 shows examples of good walnut kernels and typical defects. Optical sorting machines using a combination of visible and InGaAs technology are able to identify pieces of shell that still remain in the product after all of the mechanical processes. This shell and membrane is removed efficiently with good product being cleaned to 99.8% in one pass on pieces as small as three millimetres. As with many nut products the kernels (and even pieces of kernels) are of high value so it is important to minimise the loss of good product. This is achieved by re-sorting the product rejected from the first pass sort. This resorted product stream contains more defects than good product so it is more efficient to aim the ejectors at the good product – this technique is known as the reverse sorting. The net result is a very concentrated final reject and an accept stream of both high quality and high yield.

11.5.3 Fruit and vegetables This section contains case studies from the fruit and vegetable sectors. Fruit and vegetables are coming under increasingly stringent food regulations that insist on clean and safe quality food for today’s markets. To achieve these

Published by Woodhead Publishing Limited, 2013

282 Robotics and automation in the food industry high standards many processing companies have had to go one step further than using the traditional mechanical cleaning and grading machines once seen as the backbone of the food industry. Consequently optical sorting machines are now used throughout the fruit and vegetable industry. Optical sorting of vegetables is normally carried out at four different stages throughout the cleaning process: fresh (pre-blanched), post freezer, bulk to bulk (freezer) and end of line (packing halls). At each stage of the optical sorting process, different techniques and technologies are applied, including high-definition visible camera technology, shape sorting and InGaAs technology. Fresh product The first stage of sorting is carried out while the product is still in the fresh state directly after pre-cleaning. In all cases of first stage optical sorting all products are pre-washed before being sorted but will also go through different mechanical operations depending on the products such as dicing, slicing and peeling of root crops or de-snibbing of flat and round beans. At this stage of sorting the purpose is to remove gross defects to avoid wasting energy on blanching and freezing defective material. The pre-cleaning of peas includes the washing of the product by means of a water flume system, a stone remover flotation tank and grading grids to remove the larger extraneous vegetable matter (EVM) such as pods, pod parts, stalks, flower heads, nightshade, as well as snails, insects and rodents, etc. The peas then pass through an air classifier to remove any smaller EVM that is still within the product flow. The product is then transported via de-watering conveyor belts to a dedicated delivery vibrator that feeds the product in a uniform manner to the optical sorter. The sorting machine is equipped with a self tracking belt system that controls the speed of the product through to the optical inspection system. The product is then viewed from both sides with high-definition visible cameras to detect any colour defects seen on the peas. These defects are removed from the product stream by an array of high performance air ejectors. Although the peas have been pre-cleaned there will still be contaminants that escape the pre-cleaning along with colour defects present on the product. In many cases the EVM will be the same colour as the pea and can be difficult to remove using colour information as the visual cue. In these cases shape sorting is used. Shape sorting is a technology that uses algorithms that detect the differences in the shape of each piece of product. For example, the pods are often the same colour as the peas but can be detected on basis of shape alone because pods are either a triangular or rectangular shape whereas peas are round. In the case of high capacity commodities such as peas, which are processed at 12 t/h, the shape algorithms use object separation to separate clusters of touching product thus reducing the rejection of good product. Figure 11.11 shows an example of the reject achieved while sorting peas at a high capacity. Frozen product The first stage cleaning removes most of the gross defects but there will still be a small percentage of contaminants left in the product. This is where the second

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 283

Fig. 11.11 An example of the reject from sorting fresh peas.

stage optical sorting is carried out after the product has been frozen. This can be done either post freezer where the product is optically sorted and then collected in plastic-lined cardboard cartons ranging from 600 to 1000 kg or collected into cardboard cartons post freezer without being sorted but to be optically sorted at a later date using a process known as bulk to bulk sorting. Second stage sorting is typically carried out on chute machines where the feed system is a high performance ultra-high molecular weight (UHMW) chute. When optically sorting post freezer the subtle colour defects on the product and any remaining defective materials such as EVM that were not removed at the first stage of cleaning are now removed using the visible cameras and shape technology. When the product is in its frozen state the colour of the product changes, which may present cases where it becomes harder to differentiate the colour of the product from the colour of the defect. Examples of this are flower heads or nightshade in peas. In these cases a combination of colour and shape algorithms are used to detect the defects with minimal removal of good product. Bulk to bulk Bulk to bulk uses product that has been previously stored in a freezer in 600– 1000 kg plastic-lined bulk cartons and is brought out of the storage freezer, tipped into a holding hopper, sorted and then returned to another bulk carton. When optically sorting bulk to bulk additional contaminants can be introduced to the products in the form of packaging materials such as plastics wood and cardboard. These are easily detected and removed with the high performance InGaAs technology. End of line The final stage of sorting is at the end of line in the packing halls. Here the optical sorter acts as the last check on the finished mono and pre-mixed products

Published by Woodhead Publishing Limited, 2013

284 Robotics and automation in the food industry

Fig. 11.12 An optical sorting machine designed for bulk to bulk sorting.

(up to a 15 product mix) ranging from a rice grain to an 85 mm broccoli floret prior to packaging. The prime objective is to look for any packaging materials that may have contaminated the product. As in the bulk to bulk stage optical sorting InGaAs technology is used to remove any wood plastic or cardboard that is present. The InGaAs technology has produced a step change in the performance of optical sorting on end of line with less than 0.1% of good product falsely rejected. Figure 11.12 shows a photograph of an optical sorting machine suitable for end of line sorting. Figure 11.13a and 11.13b are images of the same product stream captured by bichromatic visible and InGaAs cameras of an end of line machine. Note that the bichromatic visible image is a colour image but the reproduction in this book is shown only in greyscale. However, the main point is that the foreign material is hard to detect using visible cameras, but much easier to detect using short wave IR. Fruits The same technologies described for vegetables are applied to optical sorting of frozen fruits. Again, the technology used depends on at what stage in the line the product is to be sorted. Typical contaminants occurring in wild berries include the small stalks left on the fruit after being harvested as well as moth cocoons, leaves, sticks and other general contaminants found on a forest floor.

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 285

Cardboard Plastic Wood

Fig. 11.13 (a) An image of vegetables with foreign material as seen by a visible camera. (b) The same vegetables and foreign material as in Fig. 11.13a but as seen by an InGaAs camera.

In general the majority of these contaminants are removed mechanically prior to optical sorting. The shape sorting of the optical sorting machine provides the last cleaning process for detecting and removing any of these remaining defects.

11.6 Conclusion Optical sorting is a well-established robotics and automation technique that continues to find new applications within the food industry. This has been illustrated in this section with case studies from three different sectors of the food industry. The application of optical sorting in the processing of staple crops such as rice, pulses and maize continues to increase. In the wheat industry, recent innovations of optical sorting have led to an increasing trend of replacing

Published by Woodhead Publishing Limited, 2013

286 Robotics and automation in the food industry some of the traditional conventional mechanical cleaning equipment by optical sorters. In high-value dry commodity sectors such as nuts, recent innovations in optical sorting have enabled food producers to increase both the yield and quality standards of their final product. In the fruit and vegetable sector, step improvements of the efficiency of the processing lines have been achieved by introducing optical sorting machines to all stages of the cleaning process.

11.7 Future trends The fundamental requirements of capacity, efficiency, yield and cost that have always motivated innovation in optical sorting for the food industry will remain. But what are the influences from a broader context that will drive innovation? In this section we focus on two areas that we anticipate will motivate future changes: 1. Food safety 2. Food security

11.7.1 Food safety Food safety is the discipline of preventing foodborne illness. Food poisoning can be caused either by an infectious agent (biological) or a toxic agent (chemical). In many applications optical sorting is sufficient for the removal of items with high levels of infectious or toxic contamination, but in some applications optical sorting only reduces but does not eliminate the overall level of contamination. There is significant room for improvement in the accuracy of on-line detection and removal of foods with unacceptable contamination levels.

11.7.2 Food security The Royal Society recently published a policy document on food security (Royal Society, 2009). This Royal Society report states the need for the sustainable intensification of global agriculture in which yields are increased without adverse environmental impact and without cultivation of more land. It is clear that the needs of sustainable intensification will require major changes in both agriculture and in the bulk food industry. In turn, we expect that these changes will necessitate future innovations for optical sorting.

11.8 Sources of further information and advice www.buhlersortex.com

Published by Woodhead Publishing Limited, 2013

Automation and robotics for bulk sorting in the food industry 287

11.9 References Bee S C and Honeywood M J (2002) Colour in Food – Improving Quality, Cambridge, Woodhead Publishing. Bee S C and Honeywood M J (2007) Detecting Foreign Bodies in Food, Cambridge, Woodhead Publishing. Eisinger T (1999) Everett H. Bickley Collection, 1919–1980 #683. Available from: http:// americanhistory.si.edu/archives/d8683.htm [Accessed 23 July 2010]. Royal Society (2009) Reaping the Benefits: Science and the Sustainable Intensification of Global Agriculture, London, Royal Society. USDA (2009) United States Standards for Rice. Available from: http://archive.gipsa.usda. gov/reference-library/standards/ricestandards.pdf [Accessed 23 July 2010]. Zeng M and Williamson P J (2007) Dragons at Your Door, Boston, Harvard Business School Press.

Published by Woodhead Publishing Limited, 2013

12 Automatic control of food chilling and freezing C. James and S. J. James, Grimsby Institute of Further & Higher Education (GIFHE), UK

DOI: 10.1533/9780857095763.2.288 Abstract: It is estimated that 40% of all foods require refrigeration and rely on a well controlled cold-chain to maintain product quality and safety in order to be supplied to the consumer. This chapter looks at the key drivers and challenges for automatic controls for food chilling and freezing operations. The key operations that make up the modern food cold-chain are discussed individually in detail, and in each case the current or future applications for automation of these operations is highlighted and discussed. Key words: cold-chain, chilling, freezing, refrigerated distribution, refrigerated retail display, energy.

12.1 Introduction: key drivers and challenges for automatic control of food chilling and freezing Refrigeration stops or reduces the rate at which changes occur in food. These changes may be microbiological (i.e. growth of microorganisms), physiological (e.g. ripening, senescence and respiration), biochemical (e.g. browning reactions, lipid oxidation and pigment degradation) and/or physical (such as moisture loss). An efficient and effective cold-chain is designed to provide the best conditions for slowing, or preventing, these changes for as long as is practicable. Effective refrigeration produces safe food with a long high-quality shelf life. For example, Gill and Penney (1986) reported that a shelf life of up to 23 weeks can be achieved in lamb cuts if they are kept at −2°C. Worldwide it is estimated that 40% of all foods require refrigeration, and that 15% of the electricity consumed is used for refrigeration (Mattarolo, 1990). With the rising concern over climate change, global warming and the recent sharp

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 289

Raw material harvesting

Cold-storage

Transport

Transport

Cold-storage

Manufacture/ processing CHILLING/FREEZING

Cold-storage (Distribution centre)

Transport

Retail outlet (Supermarket/ caterer)

Consumer handling/ storage/cooking

Fig. 12.1 The sequence of events within a typical cold-chain.

increases in energy costs, there is increasing pressure to make significant reductions in carbon emissions and energy use (James and James, 2010). However, currently worldwide less than 10% of perishable foodstuffs are in fact refrigerated (Coulomb, 2008), and it is estimated that post-harvest losses account for 30% of total production (Coulomb, 2008). Since the production of food involves a significant carbon investment, which is squandered if the food is then not utilised, there is a balance to be achieved between saving energy and the need for refrigeration (James and James, 2010). To provide safe food products of high organoleptic quality, attention must be paid to every aspect of the cold-chain, from initial chilling or freezing of the raw ingredients, through storage and transport, to retail display (Fig. 12.1). The coldchain consists of two distinct types of operation. In processes such as primary and secondary chilling, or freezing, the aim is to change the average temperature of the food. In others, such as chilled or frozen storage, transport, and retail display, the prime aim is to maintain the temperature of the food. Removing the required amount of heat from a food is a difficult, time and energy consuming operation, but critical to the operation of the cold-chain. As a food moves through the coldchain it becomes increasingly difficult to control and maintain its temperature.

© Woodhead Publishing Limited, 2013

290 Robotics and automation in the food industry Table 12.1 Best estimate of the top ten food refrigeration processes in the UK ranked in terms of their potential for total energy saving Sector

Energy

Saving

‘000 t CO2/y GWh/y

%

GWh/y

1 Retail display

3100–6800

30–50

6300

2 3 4 5

2100 1200 500 20–330

5800– 12 700 4000 4800 900 309–610

30–50 20–25 20–40 20–30

2000 1200 360 180

120–220

220–420

20–30

130

50–170 130

100–320 250

20–30 20–30

100 80

80–100 60–80

140–190 110–140

~30 20–30

60 40

6 7 8 9 10

Catering – kitchen refrigeration Transport Cold storage – generic Blast chilling – (hot) processed foods: ready meals, etc. Blast freezing – (hot) processed foods: ready meals, etc. Milk cooling – raw milk on farm Dairy processing – milk, cheese, yogurt, etc. Potato storage – bulk raw potatoes Primary chilling – meat carcasses

This is because the temperatures of bulk packs of refrigerated product in large storerooms are far less sensitive to small heat inputs than single consumer packs in an open retail display cabinet or in a domestic refrigerator/freezer. Failure to understand the needs of each process may result in excessive weight loss, higher energy use, reduced shelf life, and/or deterioration in product quality. In the United Kingdom, 11% of electricity is consumed by the food industry (BERR, 2005). However, detailed estimates of what proportion of this is used for refrigeration processes are less clear and often contradictory. Efforts to determine how much energy is used in each sector of the food industry for refrigeration are often hampered by the apparent lack of measured data and limited availability of process throughput data (Swain, 2006). However, a mapping exercise (James and James, 2010) has identified and ranked the top ten food refrigeration operations in the United Kingdom in terms of potential to achieve the greatest total reduction in energy usage (Table 12.1). In this chapter, some of these key food refrigeration operations are described and the current input of control/automation and future potential discussed for each of these operations.

12.2 Automation in refrigerated food retail display In developed countries, and increasingly in developing ones, food is marketed to the consumer from a refrigerated retail display cabinet. Display cabinets can consume approximately 50% of the total electrical energy consumed in the food cold-chain (James and James, 2010). The temperature of individual consumer packs, small individual items and especially thin-sliced products responds very quickly to small amounts of added heat. All these products are commonly found

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 291 in retail display cabinets, and marketing constraints require that they have maximum visibility. Maintaining the temperature of products below set limits while they are on open display in a heated store will always be a difficult task. Average temperatures in chilled cabinets can vary considerably from cabinet to cabinet, with inlet and outlet values ranging from −6.7°C to +6.0°C, and −0.3°C to +7.8°C, respectively in one survey (Lyons and Drew, 1985). The temperature performance of an individual display cabinet does not only depend on its design. Its position within a store and the way the products are positioned within the display area significantly influences product temperatures. In non-integral (remote) cabinets (i.e. those without built-in refrigeration systems) the design and performance of the stores central refrigeration system is also critical to effective temperature control. The desired chilled display life for wrapped meat, fish, vegetables and processed foods ranges from a few days to weeks and is primarily limited by microbiological considerations. Retailers of unwrapped fish, meat and delicatessen products normally require a display life of one working day, which is often restricted by appearance changes. Frozen food can potentially be displayed for many weeks. Reducing energy consumption in a chilled multi-deck cabinet is substantially different from reducing it in a frozen well cabinet, that is, one where the food is contained within the cabinet and the access is via the top to remove products (James et al., 2009). Improvements have been made in insulation, fans and energy efficient lighting but only 10% of the heat load on a chilled multi-deck comes from these sources compared with 30% on the frozen well. Research efforts are concentrating on minimising infiltration through the open front of multi-deck chill cabinets, by the optimisation of air curtains and airflows, since this is the source of 80% of the heat load. In frozen well cabinets reducing heat radiation onto the surface of the food, which accounts for over 40% of the heat load, is a major challenge. Currently the only mechanical automation to be seen in a supermarket is limited to possibly a forklift truck to unload deliveries and the conveyor belts found at checkouts. This contrasts with the step changes in automatic recording of sales and reordering of products to keep displays full. In addition automatic remote sensing of the performance of the display cabinets themselves is now common practice. Faults can be picked up remotely and repair personnel dispatched long before the temperature of the food is compromised. It is clear that mechanical automation could have a dramatic effect on food retailing if key non-technical barriers could be overcome. The main ones are consumer trust and habits. Imagine a system where the consumer selects refrigerated products from sealed retail display cabinets and the products are automatically transferred and waiting for them at the checkout. In sealed systems energy consumption would be substantially reduced and temperature control substantially improved with consequent improvements in food safety and quality. The next stage in automatic retailing is already happening, that is, consumers ordering products from a virtual store (Cairns, 1996; Vlachopoulou & Matopoulos, 2010). In many countries ordering via the Internet and home delivery of the products

© Woodhead Publishing Limited, 2013

292 Robotics and automation in the food industry is becoming very common. Some retailers, such as Tesco, have chosen a storebased system that uses existing assets to the maximum, while other retailers have invested in highly automated distribution centres (Kämäräinen et al., 2001). At present the main benefit of automation is seen to be the reduced labour requirements; however reduced energy use, and in a food quality and safety context better temperature control are other advantages of an automated distribution model over a store-based one. Time will tell which will become the prevailing model.

12.3 Automation of refrigeration and freezing operations in food catering The second biggest use of energy in food refrigeration is in catering operations. Food catering is an expanding sector of the food chain and catering operations are using more and more refrigerated facilities to store chilled and frozen raw materials. Increasingly they are also introducing systems to chill or freeze prepared food and store work in hand. Most of these systems are small – refrigerators, freezers and blast chillers/freezers with integrated, that is, built-in, refrigeration units. Even the larger walk-in chill or freezer rooms are manually loaded and unloaded. At present, operations in the catering area are generally ‘low tech’, even in the major multinational fast food chains (Rodgers, 2008). Although there would be advantages in terms of energy, food safety and quality if such systems were sealed with automatic loading and unloading, it is difficult to envisage this happening in the foreseeable future. In large scale catering operations supplying cook-chill, cook-freeze or sous-vide products continuous tunnel or spiral coolers/freezers (described later) may be used. Some integrated cook–chill systems are available with robotic loading and unloading, or automatic/robotic transfer of the product between the cooking and cooling operations. The other catering operation where automation is common practice is in food vending machines. Many thousands of chilled drinks in cans and bottles, and prepacked chilled foods for example, sandwiches, pies, pasties, salads, etc., are automatically dispensed every minute from refrigerated vending machines throughout the world. Some food vending machines now incorporate real time data collection of sales for automated replenishment (Poon et al., 2010). Robotics and control are also being incorporated into the machines themselves, for example automated ice cream portioning (Friedrich and Lim, 2001). Soft ice cream is generally processed by machine but dispensed manually. Automated robotic delivery and management of serving has been shown to reduce material and operating costs.

12.4 Automation in refrigerated food transport systems The third biggest use of energy in food refrigeration is in transportation. Over a million refrigerated road vehicles, 400 000 refrigerated containers, and many

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 293 thousands of other forms of refrigerated transport systems, are used to distribute refrigerated foods throughout the world (Gac, 2002). Although air-freighting was initially used for high-value perishable products, such as strawberries, asparagus and live lobsters (Sharp, 1988; Stera, 1999), foods do not necessarily have to fall into this category to make air transportation viable since it has been shown that ‘the intrinsic value of an item has little to do with whether or not it can benefit from air shipment, the deciding factor is not price but mark-up and profit’ (ASHRAE, 2006). This sector now accounts for 14% by volume of total global air freight (Bridger, 2008), and is continuing to grow rapidly. All these transportation systems are expected to maintain the temperature of the food within close limits to ensure its optimum safety and high-quality shelf life. Developments in automatic temperature controlled transportation systems for chilled products have led to the rapid expansion of the chilled-food market in particular. It is particularly important that the food is at the correct temperature before loading since the refrigeration systems used in most transport containers are not designed to extract heat from the load but to maintain the temperature of the load. If products have been cooled to the correct temperature before loading and do not generate heat then they only have to be isolated from external heat ingress. Surrounding them with a blanket of cooled air achieves this purpose. Care has to be taken during loading to stop any product touching the inner surfaces of the vehicle because this would allow heat ingress by conduction during transport. Modern refrigerated containers are equipped with temperature sensors and controllers that automatically control the refrigeration unit. In the large containers used for long distance transportation food temperatures can be kept within ±0.5°C of the set point. With this degree of temperature control transportation times of 8–14 weeks (for vacuum packed meats stored at −1.5°C) can be carried out and still retain a sufficient chilled storage life for retail display (Gill and Penney, 1986). Products such as fruits and vegetables that produce heat by respiration, or products that have to be cooled during transit, also require circulation of air through the product. This can be achieved by directing the supply air through ducts to channels at floor level or in the floor itself. Most ISO (International Organization for Standardization) containers are either ‘refrigerated’ or ‘insulated’. The refrigerated containers have refrigeration units built into their structure. The units operate electrically, either from an external power supply on board the ship or dock or from a generator on a road vehicle. Insulated containers either utilise the plug type refrigeration units already described or may be connected directly to an air-handling system in a ship’s hold or at the docks. Close temperature control is most easily achieved in containers that are placed in insulated holds and connected to the ship’s refrigeration system. When the containers are fully loaded and the cooled air is forced uniformly through the spaces between cartons, the maximum difference between delivery and return air can be less than 0.8°C (Heap, 1986). The entire product in a container can be maintained to within ±1.0°C of the set point.

© Woodhead Publishing Limited, 2013

294 Robotics and automation in the food industry Electronic dataloggers are frequently used for monitoring, certification or verification that shipment requirements have been maintained at all times, particularly where long periods of shipment or storage have been involved (Studman, 2001). Radio frequency identification (RFID) is enabling far tighter control of cold-chain distribution than was previously possible (Pramatari et al., 2010). General packet radio service (GPRS) equipped systems are available that enable real time monitoring and control of storage and product temperatures during transportation in any geographical location with public mobile network coverage. Controlled atmosphere (CA) storage is often used during transportation of many fruits and vegetables to extend storage life and delay ripening, thus enabling perishable products to be made available to markets they could not otherwise reach (Bishop, 1996; Markarian et al., 2003; Thompson, 2010). Such conditions require a control system that can dynamically control both temperatures and gas levels, as these change due to physiological changes in the products during storage and transportation. These changes are affected by produce age, temperature, oxygen and carbon dioxide concentrations and other factors that make it difficult to predict or model the process (Markarian et al., 2003). Computer-based controls provide the capability of monitoring and adjusting many operations independently and simultaneously. Proportional, integral and derivative (PID) controllers are the most popular feedback controllers used, while more advanced control techniques include adaptive, fuzzy logic, knowledge-based and artificial intelligent controllers (Markarian et al., 2003). As in retailing and catering, there is currently little mechanical automation of food transportation operations. Most transport containers for refrigerated foods are loaded by hand or using forklift trucks. In a few cases retractable conveyors are used to directly transport packaged refrigerated food products into the transport container or vehicle. Automation is extensively used in container depots throughout the world to load and unload ships and distribution vehicles. In addition the routing of refrigerated products throughout the world and for local delivery is almost totally automated. Computer programs are used to optimise routes, construct the best loads and ensure that there is adequate shelf life after transportation. One concept that is currently being considered is automatic processing of refrigerated food during transportation. Many foods are processed pre- or posttransportation to add value to the raw material. Since seaborne transportation of refrigerated food can take many weeks, if not months, could autonomous systems process the raw materials within the transportation period? Studies are currently being considered at the Food Refrigeration and Process Engineering Research Centre (FRPERC) to identify suitable products and processing options.

12.5 Automation in food chilling and freezing systems During chilling and freezing heat can only be removed by four basic mechanisms: radiation, conduction, convection or evaporation. To achieve substantial rates of

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 295 heat loss by radiation large temperature differences are required between the surface of the product and that of the enclosure. This occurs in the initial stages of chilling or freezing of cooked or warm products. Physical contact between the product and the source of refrigeration is required to extract heat by conduction. Plate conduction coolers are used for quick cooling of some packaged products and highly perishable products such as fish blocks. For the majority of foods the heat lost through evaporation of water from the surface is a minor component of the total heat loss, though it is the major component in vacuum cooling. Evaporation from a foods’ surface reduces yield and is not desirable in many refrigeration operations, but can be useful in the initial cooling of unpackaged cooked food products. Convection is by far the most important heat-transfer mechanism employed in the majority of food refrigeration systems. Most foods are cooled by the convection of heat into air or another refrigerated medium. For the majority of chilled and frozen foods, air systems are used, primarily because of their flexibility and ease of use. However, other systems can offer much faster and more controlled chilling or freezing. It is not unusual for food products (or ingredients found in food products) to be chilled or frozen a number of times before they reach the consumer. For example, during industrial processing frozen raw material is often thawed, or tempered, before being turned into meat-based products, for example, pies, convenience meals, burgers, etc., or consumer portions, fillets, steaks, etc. These consumer-sized portions are often refrozen before storage, distribution and sale.

12.5.1 Chilling and freezing systems There are a large number of different chilling and freezing systems for food based on moving air, direct contact, immersion, ice, cryogenic, vacuum and pressure shift. The degree of control and automation used is a function of both the system used and the food being refrigerated. In general the higher the throughput, and the smaller the individual item being refrigerated, the more likely the process is to be automatically controlled. Air Air systems range from the most basic in which a fan draws air through a refrigerated coil and blows the cooled air around an insulated room, to purpose-built conveyorised blast chilling tunnels or spirals. Relatively low rates of heat transfer are attained from product surfaces in air-cooled systems. The big advantages of air systems are their cost and versatility, especially when there is a requirement to cool a variety of irregularly shaped products. Chilling/freezing tunnels (Fig. 12.2) are often included in a continuous processing line, with either packaged or unpackaged raw materials or processed food being conveyed through the refrigerated space. To increase the rates of heat transfer and therefore reduce the chilling/freezing time impingement systems have been designed to break up the boundary layer around the food (Fig. 12.3). Chilling/freezing is often the longest operation in a

© Woodhead Publishing Limited, 2013

296 Robotics and automation in the food industry

Fig. 12.2 Continuous chilling/freezing tunnel.

Ceiling-mounted evaporators Air 3 m/s

Air 20 m/s

Belt

Fig. 12.3 Schematic of impingement chilling/freezing system.

food processing line and if floor space is limited spiral chillers/freezers are often used to increase cycle times (Fig. 12.4). There are many manufacturers of tunnel and spiral systems including Jackstone Freezing Systems ltd, Frigoscandia, ADZ Zephyr, Lomax Technical Services Ltd and Eurotek, etc. The majority of continuous air chiller/freezers have a fixed single retention time. However, a number of automated in-line chilling/freezing systems are now available that can cope with chilling/freezing multiple different products at the same time in the same equipment. In general these are tray systems in which products are loaded on trays, which in turn are loaded into predetermined levels within

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 297

Fig. 12.4 Continuous spiral chiller/freezer.

the chiller/freezer. Each level of trays can be treated independently. Theoretically, the system can handle as many different products, each with their own retention times, as there are levels available. This allows the system to chill/freeze a wide range of products, volumes, sizes and weights with different entry and exit temperatures, all simultaneously. Such systems are available from companies such as Intec USA LLC, and Odenberg. Direct contact Contact refrigeration methods are based on heat transfer by contact between products and metal surfaces, which in turn are cooled by either primary or secondary refrigerants. Contact cooling offers several advantages over air cooling, such as much better heat transfer and significant energy savings. Contact cooling systems include plate coolers, jacketed heat exchangers, belt coolers and falling film systems. Automatic horizontal plate freezers are suitable for fish, meat, vegetables, ice cream and ready meals packed in cartons, polybags or in trays for bulk freezing. In a typical system the product travels on a conveyor belt to the front part of the freezer to be frozen, a sensor registers the packages and, once a row has been filled, it is pushed forward onto the freezing plate. When this plate is full, it is raised to make room for a new empty plate. Once all the plates have been filled they are lowered to return to the initial position. At this point, as one row of fresh product packages is loaded onto the freezer, another row of frozen packages is unloaded off the opposite side. Some products, such as surimi and blocks used for fish fingers, require blocks of uniform density, exact dimensions and precise tolerances. In this case, the product has to be frozen and compressed at the same

© Woodhead Publishing Limited, 2013

298 Robotics and automation in the food industry time, thus also avoiding the formation of air pockets that can cause imperfections in the subsequent processing. Many models of automatic plate freezers are available with capacities ranging from 400 to 5000 kg/h. Manufacturers of plate freezers include Sinrofreeze Equipment Co. Ltd, Jackstone Freezing Systems Ltd, Dybvad Stal Industri (DSI), NH3 Refrigeration Ltd and Milmeq. Some systems combine air and contact cooling, an example being the Torry IQF Continuous Steel Band Freezer (Hall, 1997). The Torry Freezer consists of a solid stainless steel belt on which the product is conveyed through an insulated enclosure. Refrigerated air is circulated both over and under the belt at high velocity, thus freezing the product rapidly. It was developed for fish fillets, but can also be used for a wide range of fish products, fruit and processed meat. It has a capacity ranging from 150 kg/h for a single belt to 1000 kg/h for a double belt freezer. The freezing time is typically from 6 min for extruded scampi to 40 min for 30 mm thick fish fillets. Immersion/spray Immersion/spray systems involve dipping product into a cold liquid, or spraying a cold liquid onto the food. This produces high rates of heat transfer due to the intimate contact between product and cooling medium. Both offer several inherent advantages over air cooling in terms of reduced dehydration and coil frosting problems (Robertson et al., 1976). Clearly if the food is unwrapped the liquid has to be ‘food safe’. Automatic immersion cooling systems with either ice water mixtures or brines are often used for the cooling of packs of sauce, soup or sous-vide products. Similar systems have also been designed for freezing (Fig. 12.5). Cooling using ice or cryogenic substances are essentially immersion/spray processes. The freezing point of the cooling medium used dictates its use for chilling or

Fig. 12.5 Automatic immersion freezing system for whole fish.

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 299 freezing. Obviously any immersion/spray freezing process must employ a medium at a temperature substantially below 0°C. This necessitates the use of non-toxic salt, sugar or alcohol solutions in water, or the use of cryogens or other refrigerants. Direct spraying of liquid nitrogen onto a food product while it is conveyed through an insulated tunnel is one of the most commonly used methods of applying cryogens. Cooling is brought about by boiling off of the refrigerant. As well as using the latent heat absorbed by the boiling liquid, sensible heat is absorbed by the resulting cold gas. Due to very low operating temperatures and high surface heat-transfer coefficients between product and medium, cooling rates of cryogenic systems are often substantially higher than other refrigeration systems. Cryogenic freezing systems are available from Air Products Ltd, BOC, Praxair and Airgas among others. Vacuum cooling Food products having a large surface area to volume ratio and an ability to readily release internal water are amenable to vacuum cooling. The products are placed in a vacuum chamber (typically operating at between 530 and 670 Nm−2) and the resultant evaporative cooling removes heat from the food. Evaporative cooling is quite significant, the amount of heat released through the evaporation of 1 g of water is equivalent to that released in cooling 548 g of water by 1°C. Suitable products, such as lettuce, can be vacuum cooled in less than 1 h. In general terms a 5°C reduction in product temperature is achieved for every 1% of water that is evaporated. Since vacuum cooling requires the removal of water from the product, pre-wetting is commonly applied to prevent the removal of water from the tissue of the product. Windham Harvest Specialists produce vacuum coolers for spinach, lettuce, etc in sizes ranging from a 2- to a 12-pallet system. Vacuum coolers for raw products are batch loaded systems, however similar systems for the cooling of pie fillings, soups and sauces often operate with totally automatic filling, cooling and unloading cycles. The Food Machinery Company Ltd has developed vacuum coolers specifically for cooked meats. High-pressure freezing High-pressure freezing and in particular ‘pressure shift’ freezing is attracting considerable scientific interest (LeBail et al., 2002). The food is cooled under high pressure to sub-zero temperatures but does not undergo a phase change and freeze until the pressure is released. Rapid nucleation produces small even ice crystals. However, studies on pork and beef have failed to show any real commercial quality advantages. The majority of current high-pressure freezing plants are totally automatic, with a series of discrete chambers being loaded and unloaded in sequence.

12.6 Automation in food cold storage systems After a raw material or finished product is chilled or frozen it is usually stored in a refrigerated facility. Storage systems range from domestic refrigerators or

© Woodhead Publishing Limited, 2013

300 Robotics and automation in the food industry freezers through to very large cold-storage complexes. A typical UK cold-store will have 75 000 m3 of storage space and be fitted with 10–14 m long mobile racks. However, larger cold-stores operate throughout the world that are almost three times as large (200 000 m3) and have 32–38 m long automated racks (http://www.grimsby.ac.uk/documents/defra/sectrep-frozenstorage.pdf). Food can spend many months if not years in a frozen cold-store typically operating at −20°C to −30°C, or between a few days and a year in a chilled cold-store operating at 0–5°C. The common feature of all cold-stores is that they are designed to maintain, not change, the temperature of a previously cooled food. As discussed in the transport section, CA storage is often used for the cold storage of many fruits and vegetables to extend storage life and delay ripening (Bishop, 1996; Markarian et al., 2003; Thompson, 2010). The long times used for such storage and the relative importance and value of the perishables stored require reliable control system that can dynamically control both temperatures and gas levels. Increasingly medium and large cold-stores use completely automatic loading and unloading systems that eliminate the need for any manual entry into the store except during breakdown or for maintenance. The need for access by forklift trucks can require up to 60% of the floor area for gangways (Refrigers.com, 2011). In a typical cold storage warehouse where fork lifts are used, the lift’s maximum reach is around 6 m, allowing about 4 vertical pallet positions for a maximum 6 m rack height (McMahon, 2011). Automatic stacker cranes were first used in cold-stores in the 1960s but automation only began to become widespread from the 1980s and there are now many installations throughout the world. With automated stores, the store height can be increased considerably, up to 40 m (Duiven and Binard, 2002). One crane can service some 4000 pallet positions at the rate of 50 pallets per hour (Refrigers.com, 2011). The footprint reduction is also an important factor in energy savings since much of the heat gain occurs in the roof (McMahon, 2011). Mobile racking, where the lines of racking are on transverse rails, can be closed together when access is not needed, but rolled apart to provide an aisle for a forklift truck. This system is best for a limited range of products moving in rotation, since the racking will not have to be moved very often. A typical small installation might have seven mobile racks; each 25 m long by 4 pallets high, and require an extra 3 m width for one access aisle, plus an end access of 4 m (Refrigers.com, 2011). This results in a store of 504 pallet capacity and a floor area of 270m2. The tight stacking when the racks are closed impedes the airflow around the pallets, so this system is not suitable where some cooling of the product may be required. There are developments in the use of robots in refrigerated storage areas. In Australia robots are being used in an abattoir to palletise cartons of frozen meat (RA Group, 2011). The robots used are capable of palletising twenty 30 kg cartons boxes of frozen meat per minute. The robots are located in the cold-store area where temperatures approach 0°C to cater for both domestic and export

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 301 requirements, the robots are programmed to palletise a range of different sized cartons into different palletising patterns.

12.7 Advances in research and future trends The food industry is under increasing pressures to reduce unit production costs whilst maintaining, if not improving, the safety and quality of the products they produce. Optimising the refrigeration of food can increase throughput, maximise yield and reduce energy consumption. Worldwide it is estimated that 40% of all food requires refrigeration (Mattarolo, 1990). On the best available data, the energy saving potential in the top five refrigeration operations (retail, catering, transport, storage and primary chilling), in terms of the potential to reduce energy consumed, lies between 4300 and 8500 GWh/y in the United Kingdom (James et al., 2009). New/alternative refrigeration systems/cycles, such as Trigeneration, Air Cycle, Sorption-Adsorption Systems, Thermoelectric, Stirling Cycle, Thermoacoustic and Magnetic refrigeration, have the potential to save energy in the future if applied to food refrigeration (Tassou et al., 2010). However, none appear to be likely to produce a step change reduction in refrigeration energy consumption within the food industry within the next decade. Thus improved automated control and handling of current systems is needed to improve energy consumption. In food chilling and freezing processes automatic systems are widely used. However, most refrigerated food processing areas are currently a compromise between the conflicting requirements of the food being processed and the people doing the processing. In the majority of cases the eating quality and microbial safety of the food is optimal if the food is processed at a low temperature in the absence of light. People do not operate well in the dark and at low temperatures. Research is required to develop low cost robotics and automation systems that can reliably and hygienically process food in dark, low temperature environments.

12.8 Sources of further information and advice FRPERC at the Grimsby Institute is one of the few international sources of information and advice on both food automation and food refrigeration (www.grimsby. ac.uk/What-We-Offer/FRPERC/Contact-FRPERC/). The International Institute of Refrigeration 177, boulevard Malesherbes, 75017 PARIS, France (www.iifiir.org) is a valuable source of information on all aspects of food refrigeration. While ASHRAE (The American Society of Heating, Refrigerating and Air-Conditioning Engineers) provides a very similar role in the USA (www.ashrae.org/).

© Woodhead Publishing Limited, 2013

302 Robotics and automation in the food industry In the UK the Food Manufacturing Engineering group (FMEG) and the Centre for Food Robotics and Automation (CenFRA) have been specifically set up to address the automation needs of the food industry (www.fmeg.org.uk/).

12.9 References ASHRAE (2006). ASHRAE Handbook – Refrigeration. Atlanta: American Society of Heating, Refrigerating and Air-Conditioning Engineers. Bridger R (2008). Food Air Freight, the Global Infrastructure Expansion. UK: Food Ethics Council. Cairns S (1996). ‘Delivering alternatives: Success and failures of home delivery services for food shopping’, Transport Policy, 3, 155–176. Duiven J E and Binard P (2002). ‘Refrigerated storage: New developments’, Bulletin of the IIR, no. 2002–2. Friedrich W and Lim P (2001). ‘Robotic food applications example: ice cream portioning’, Proceedings of the 2001 Australian Conference on Robotics and Automation, Sydney, 14–15 November 2001, 56–60. Gac A (2002). ‘Refrigerated transport: What’s new?’, International Journal of Refrigeration, 25, 501–503. Gill C O and Penney N (1986). ‘Packaging of chilled red meats for shipment to remote markets’, Recent Advances and Developments in the Refrigeration of Meat by Chilling, Meeting of IIR Commission C2, Bristol (UK), Section 10, 521–525. Hall G M (1997). Fish Processing Technology, 2nd edition, Chapman & Hall, London. Heap R D (1986). ‘Container transport of chilled meat’, Recent Advances and Developments in the Refrigeration of Meat by Chilling, Meeting of IIR Commission C2, Bristol (UK), 505–510. James S J, Swain M J, Brown T, Evans J A, Tassou S A, Ge Y T, Eames I, Missenden J, Maidment G and Baglee D (2009). ‘Improving the energy efficiency of food refrigeration operations’, Proceedings of the Institute of Refrigeration, 105, 1–8. James S J and James C (2010). ‘The food cold-chain and climate change,’ Food Research International, 43, 1944–1956. Kämäräinen V, Småros J, Jaakola T and Holmström J (2001). ‘Cost-effectiveness in the e-grocery business’, International Journal of Retail & Distribution Management, 29, 41–48. LeBail A, Chevalier D, Mussa D M and Ghoul M (2002). ‘High pressure freezing and thawing of foods: A review’, International Journal of Refrigeration, 25, 504–513. Lyons H and Drew K (1985). ‘A question of degree’, Food, December, 15–17. Markarian N R, Vigneault C, Gariepy Y and Rennie T J (2003). ‘Computerized monitoring and control for a research controlled-atmosphere storage facility’, Computers and Electronics in Agriculture, 39, 23–37. Mattarolo L (1990). ‘Refrigeration and food processing to ensure the nutrition of the growing world population’, Progress in the Science and Technology of Refrigeration in Food Engineering, Proceedings of meetings of commissions B2, C2, D1, D2-D3, 24–28 September 1990, Dresden (Germany), Institut International du Froid, Paris (France), 43–54. ISBN 2-903-633-533. McMahon J (2011). ‘Automated storage and retrieval system cuts energy in cold storage warehouses’, Available at: http://www.automation.com/content/minimizing-e nergy-usage-and-improving-efficiency-in-cold-storage-warehouses-with-automate d-storage-and-retrieval-systems [Accessed 6 April 2011]. Poon T C, Choy K L, Cheng C K and Lao S I (2010). ‘A real-time replenishment system for vending machine industry’, 8th IEEE International Conference on Industrial Informatics (INDIN), 13–16 July 2010, Osaka, Japan, 209–213.

© Woodhead Publishing Limited, 2013

Automatic control of food chilling and freezing 303 Pramatari K, Karagiannaki A and Bardaki C (2010). ‘Radio frequency identification (RFID) as a catalyst for improvements in food supply chain operations’, In Delivering Performance in Food Supply Chains, edited by Mena C. and Stevens G. Woodhead Publishing Ltd, Cambridge, 432–455. RA Group (2011). ‘Robotic palletising in an abattoir’, Available at: http://www.ragroup. com.au/robotics/palletising.htm [Accessed 6 April 2011]. Refrigers.com (2011). ‘Automated cold stores’, Available at: http://www.refrigers.com/ content/view/31276/28/ [Accessed 6 April 2011]. Robertson G H, Cipolletti J C, Farkas D F and Secor G E (1976). ‘Methodology for direct contact freezing of vegetables in aqueous freezing media’, Journal of Food Science, 41, 845–851. Rodgers S (2008). ‘Technological innovation supporting different food production philosophies in the food service sectors’, International Journal of Contemporary Hospitality Management, 20, 19–34. Sharp A K (1988). ‘Air freight of perishable product’, Refrigeration for Food and People, Meeting of IIR Commissions C2, D1, D2/3, E1, Brisbane (Australia), 219–224. Stera A C (1999). ‘Long distance refrigerated transport into the third millennium’, 20th International Congress of Refrigeration, IIF/IIR Sydney, Australia, paper 736. Studman C J (2001). ‘Computers and electronics in postharvest technology – a review’, Computers and Electronics in Agriculture, 30, 109–124. Tassou S A, Lewis J, Ge Y T, Hadawey A and Chaer I (2010). ‘A review of emerging technologies for food refrigeration applications’, Applied Thermal Engineering, 30, 263–276. Thompson A K (2010). Controlled Atmosphere Storage of Fruits & Vegetables, CABI International, Wallingford, UK, ISBN: 978 1 84593 6464. Vlachopoulou M and Matopoulos A (2010). ‘Adoption of e-business solutions in food supply chains’, in Delivering Performance in Food Supply Chains, edited by Mena C. and Stevens G. Woodhead Publishing Ltd, Cambridge, 417–431.

© Woodhead Publishing Limited, 2013

13 Robotics and automation in meat processing G. Purnell, Grimsby Institute of Further & Higher Education (GIFHE), UK DOI: 10.1533/9780857095763.2.304 Abstract: Tasks in the meat processing sector are physically challenging, repetitive and prone to worker scarcity. Despite the potential for automation, the inherent biological variation of meat and the commercial characteristics of the supply chain have limited the widespread implementation of automated systems. This chapter describes potential benefits and challenges, and gives an overview of some of the robotic and automation equipment available and in development for beef, pork and lamb processing. Key words: meat processing automation, robotics, primal cutting, boning, trimming.

13.1 Introduction The benefits of automation are well understood and frequently adopted by many manufacturing industries. The food sector generally has been slow to capitalise on the opportunities, particularly in the primary production operations before packing. The specific issues associated with automation for the meat sector are discussed in the following sections. 13.1.1 Scope of chapter Butchery tasks are unpleasant, physically arduous and carry a high risk of worker injury. This suggests them as prime targets for the benefits of robotisation; however, the skilled nature of the butchery task, combined with the biological variation of the raw material, poses substantial challenges. This chapter considers the applications of robotics and automation in primary meat production processes in the abattoir and cutting plant for beef, sheep/lamb and pork meat.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 305 Automation in poultry and fish production are dealt with in other chapters, as are operations such as packing and palletising that occur after retail portion cutting.

13.1.2 Drivers for automation in meat production Robotics and automation have been relatively slow to permeate the meat production industry, and the majority of tasks are still performed manually. Whilst individual machines can be developed for skilled and complex tasks, their function always remains specific to the task for which they were developed. Replacing a skilled slaughterman or butcher is difficult. For the more (seemingly) mundane tasks, such as putting lamb chops into packs, laying up sliced beef, handling ‘bundles’ of wafer ham, etc., the human is difficult to replace at a viable cost. Research projects have tackled the technical aspects of these problems for several decades, but commercially viable systems are only just beginning to emerge. There are a wide variety of commercial and product quality reasons leading many companies to investigate robotics and automation applications on meat production lines. Ultimately all drivers to adopt automation have the same aim – increased profitability. If no profit or long-term benefit is foreseeable then no changes will be implemented. The use of automation in meat processing in place of human operatives has many potential benefits, which may be tangible, intangible, social or economic. Many generic drivers are quoted to support the introduction of automation including: • Production quality: It is widely accepted that meat cuts best in the range 2–5°C, just above the initial freezing point. As the temperatures reduce, the cut quality improves, but cutting forces increase (Brown et al., 2005) to an extent where human strength could be insufficient to maintain production rates. Automation can be used to exert higher forces, maintaining or improving on cutting quality and production rates. • Product consistency: Boredom, stress and tiredness are not an issue with automated systems, typically performing a task more consistently than a human. This consistency can additionally permit efficiencies in other aspects of the business thus further aiding profitability. ‘Getting things right’ reduces waste and increases overall yield. • Added functionality: One key benefit of robotics is in performing tasks that the human cannot. Automation can make subtle adjustments beyond the skill of an operative, or be endowed with ‘superhuman’ sensory, recall, reasoning or other capabilities such as infrared detection, increased strength, X-ray vision, huge memory, etc. Machines can be designed to operate under conditions where humans could not perform effectively. This can allow processing in environments beneficial to quality, for example, sustained low temperatures, aseptic atmospheres, etc. • Worker safety: Injuries cause lost production and absence from work, not to mention costly compensation claims. The meat processing industry has a poor

© Woodhead Publishing Limited, 2013

306 Robotics and automation in the food industry safety record when compared to all types of manufacturing industry surveyed by the UK Health and Safety Executive (HSE). For 2005–2008, the meat and poultry sectors had a mean annual injury rate of 1313 per 100,000 employees compared to a mean of only 913 for all manufacturing industries (HSE, 2008). Injury occurs to both experienced and trained staff, illustrating that it is the nature of the work rather than inexperience causing the danger. Cuts made with high force towards the body, bad knife design and cold fingers contribute to the poor safety record (North, 1991). • Food safety: Foreign bodies and microorganisms can be transferred to foods from operatives. Replacement of potentially contaminating human labour by machine can reduce this risk. The costs of preserving hygiene with the large numbers of staff present in a normal meat plant increase the overall production cost. A number of studies of specific systems (Holder et al., 1997; Clausen, 2002) suggest that automating and removing staff from the production process can improve the microbial condition of processed meat. • Legislation: The minimum legal continuous working temperature for a standing, active labourer in the UK is 10°C (UK Factories Act, 1961). EEC directive 95/23/CE states that during cutting meat temperatures should not exceed 7°C, and the processing rooms should be at a maximum of 12°C. Automation and robotics can work closer to the optimum temperatures for meat processing than can be legally achieved with human operatives. • Difficulties in recruiting staff: There is a shortage of skilled labour for many of the tasks in the meat industry. The work is typically repetitive, physically intensive and takes place in an unpleasant environment. Many employers have substantial difficulties in recruiting and retaining useful staff. The continual recruitment and training introduces unwanted additional costs.

13.1.3

Barriers to introduction of robotics and automation into meat processing A traditionally conservative, cash-poor meat industry with low margins has some fundamental financial, attitudinal and commercial challenges in implementing automation systems. Developments made during the last decade have removed or reduced many of the technological barriers to automation of meat production tasks. The predominant limiting factors are now related to business and commercial factors. Despite the advances, automation technology is still a long way from the generic robot-type system capable of replacing people in most food operative situations as envisaged by Khodabandehloo and Clarke (1993). Commercial and organisational challenges Automated processing has been successfully implemented in the automotive industry where regular components and a high value product, coupled with relatively low production rates, make vehicle production an ideal process for

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 307 robotisation. Despite the product and process differences, some business experiences and observations can be transferred into the meat sector. A longer-term, less risk averse, company culture is required, and employees at all levels must be prepared to accept change. Where automation projects have failed is often in the lack of ‘buy-in’ throughout the company and lack of awareness of the skills and organisational changes required to support the implementation. The same organisational risks apply to the food sector, with additional challenges of high product variability and a constricting market structure. The low margin on most meat products reduces the finances available for investment and a marketplace dominated by major multiple retailers exacerbates the situation. The majority of labour in the food sector is unskilled, and thus sums saved by manpower substitution are low. Supply, demand and processing specifications are flexible, seasonal and regional. Many meat processing plants currently lack the in-house skills to specify and support automated systems. The skills required stretch beyond the basic engineering function of current mechanised lines into more complex aspects of automation system specification, installation, support, maintenance and reconfiguration of the system to deal with changing production requirements. Management, and production staff working alongside the automated systems, need to understand the strengths and weaknesses of the equipment and adjust practices accordingly. The entire organisation, from cleaners to directors, has to embrace a positive ‘mindset’ to automation of traditionally manual operations. Inappropriate attitudes at any of many levels can cause automation projects to fail. Technical challenges From an automation viewpoint, the complexity of meat production tasks should not be underestimated. Humans possess sophisticated, integrated sensory abilities with inbuilt reasoning and manipulation capabilities. The majority of tasks within meat production have evolved to utilise these inherent abilities. The human is excellent at evaluating situations and acting accordingly, while a typical machine system has a predestined function, and correction of only a limited number of possible perturbations can be incorporated into the design. An automated system to replicate even a small subset of human abilities can require very sophisticated systems integration. Despite the advances in meat automation progress made in recent years, the greatest technical problem is still that of coping with the natural biological variation in the product. Variable products require variable production strategies and thus flexible processing methods. This has implications for sensing systems and system elements in contact with the meat such as fixtures, grippers and cutting tools. Many meat products are relatively delicate and can be damaged by inappropriate handling. These factors tend to exclude direct technology transfer from other industries. The secondary technical challenge is in equipment longevity and suitability for food production environments. Hygienic and robust systems that can resist highpressure wash down, hot or cold, and condensation can be designed and built, but

© Woodhead Publishing Limited, 2013

308 Robotics and automation in the food industry at additional cost and complexity. This further increases costs for implementation of automation for food production. A number of studies of specific automated meat production systems (Holder et al., 1997; Clausen, 2002) acknowledge the automation production benefits but express concern over cleanability of the often complex equipment.

13.1.4 Current status of meat processing automation Pigs, cattle and sheep are all quadruped mammals and the basic operations for conversion of animals to meat are similar. There are some differences however. Cattle and sheep have their entire skin removed whereas pigs are de-haired. The details of the evisceration process are different for ruminant cattle and sheep and for omnivorous pigs. Pork and beef carcasses are split whereas smaller lamb carcasses are typically not split. All species have different cutting patterns to produce different meat products and portions; these cutting patterns can vary between countries and regions, and seasonally. Pork meat production is the most widely automated. Many of the key developments have been made by the Danish Meat Research Institute (DMRI). Whilst DMRI have a stated goal of producing a virtually fully automated pork process, some operations such as shackling, sticking, gambrelling, veterinary inspection, final trimming and removal/separation of specific organs are not included in the plan (Clausen, 2002). This ambitious target can be attempted due to the cooperative and nationally integrated structure of the Danish pork industry, research establishments and equipment producers. Whilst the size and weight of the beef carcass suggests automated processes would be of benefit to reduce the physically arduous nature of the tasks, this carcass type has received relatively little automation research and development (R&D) effort compared to lamb and pork. The key challenge for automation is the large variation seen in cattle. Slaughter animals may be from a wide variety of breeds, ages, type (bull, steer, heifer, cow, etc.) and range in weight from 200 to 1000 kg. The variations seen in other carcass types are substantially less. Mechanised processing aids, guided by human staff, have been in existence for many years, but the ‘Fututech’ Australian R&D programme (White, 1994) sought to develop the world’s first truly automated beef processing line. The system was developed through to a commercial prototype stage and designed for a minimum processing rate of 60 carcasses per hour. The system included a large number of automated or semi-automated modules that performed the majority of the slaughter tasks. These modules included rectum clearing and bagging, aitch bone cutting, head removal, brisket cutting, evisceration and tail cutting. Lamb and sheep farming and meat production form a major part in the economies of New Zealand and Australia. Not surprisingly, the majority of automation for these carcass types has originated in these regions. The Meat Industry Research Institute of New Zealand (MIRINZ) have developed a series of machines for sheep processing that use minimal sensing or compliance to adapt the motion or work piece. However, by rearranging the various tasks on the slaughter line

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 309 and by redeploying some labour to act as the ‘sensing’ or adaptation element, relatively simple machines for sheep and lamb meat processing have been successfully developed and commercialised. In the early 1980s researchers at MIRINZ developed an improved manual dressing system, later called the ‘Inverted Dressing’ system because the carcass is hung from the front feet for a proportion of the dressing process whereas on a traditional sheep production line carcasses were hung from the hind feet. This simple change reduced manning levels by 10–20% and achieved a throughput of 3200 carcasses per shift (Annan, 1982). By 1990 a typical sheep dressing line making use of all available technology developed by MIRINZ over the previous ten years required only 26 butchers. This is almost half of the manning that had been required for the traditional manual sheep production line ten years earlier. The majority of slaughter, butchery and meat processing operations are currently performed manually with simple tooling. Automation and robotics has much to offer the slaughter and meat processing industries, and whilst not widespread some automated systems are available and have been implemented. The next section describes meat processing automation for pork, beef and sheep/lamb operations.

13.2 Automation of carcass production processes before primary chilling All slaughterhouses follow a similar sequence of operations to transform the live animal into meat for consumption. After slaughter, inedible and many non-muscle parts (skin, hair, intestines, etc.) are removed to produce a carcass which is then chilled to reduce spoilage rates. After chilling, the carcass is typically subdivided into smaller sections (‘primals’) for ease of handling. These primals are then processed (boned, trimmed, cut, etc.) to produce retail joints and portions. This section deals with automation of carcass preparation processes taking place before chilling.

13.2.1 Lairage Ideally all animals rest in the lairage after arrival at the abattoir. This allows them to recover from the stresses of transport and acts as a pre-slaughter process buffer. Maintaining low stress levels in the live animal is an important welfare issue, but also has a beneficial effect on meat quality. An animal experiencing stress will have physiological changes including changes in heart rate, blood pressure, body temperature and respiration. Several stress hormones are released into the blood stream that speed the breakdown of glycogen stored in the liver and muscles, creating by-products of lactic acid and water in the meat carcass. In turn these can contribute to undesirable effects on the final quality of meat such as pale, soft, exudative (PSE) meat and dark, firm, dry (DFD) meat. The enforced herding required to move the animals around the lairage increases stress further. An automated pig lairage where these movements are

© Woodhead Publishing Limited, 2013

310 Robotics and automation in the food industry performed gently and without human presence was developed in Demark in the 1990s (Madsen et al., 2006). This equipment uses automated walls to gently herd pigs towards the slaughter raceway and is installed in several Scandinavian pork plants (Pigsite, 2009).

13.2.2 Stunning, sticking and killing The slaughter raceway leads animals from the lairage to the slaughter area. The activities of slaughter process are influenced by cultural, animal husbandry and occupational health and safety considerations. Errors have far reaching effects on animal welfare, meat quality and all downstream processes. Two methods of killing are commonly in use: stun-and-bleed, or gas kill. Stunning is carried out with an electrical shock across the head, or concussive pistol head blow, to halt brain function, and then a cut is made to the artery in the neck (sticking) to drain the blood. There are automated systems to convey animals to the stun station, most consisting of V-shaped conveyors to carry the animal to the stun operator. The current and voltage of the stunning shock is controlled, and varies for different animal sizes and species. An alternative to stun-and-bleed is gas killing, whereby animals are immersed in a CO2 environment that renders the animal unconscious before bleeding. This automated method is gaining popularity, particularly in pork production. The automated CO2 stunning units operate like enclosed Ferris wheels, with multiple compartments rotating cyclically. Small batches of around six pigs are herded into each compartment. The compartment then descends into a deep well area filled with CO2, emerging on the opposite side to pig entry where the compartment tilts, and the animals slide down a chute to the shackling line below. Residence time is typically around three minutes in 82% CO2 (Butina, 2010). Whilst gas stunning can produce higher quality meat (Channon et al., 2002), there are some concerns for animal welfare (Grandin, 2008). Beef stunning and sticking processes are ergonomically difficult to perform manually because of the size of the animal. Food Science Australia, a joint venture organisation of Australia’s Commonwealth Scientific and Industrial Research Organisation (CSIRO) and the State of Victoria, has investigated automatic systems for these tasks. This system is based largely on the Fututech module, in which a machine vision system was used to determine correct stun and sticking locations. The Fututech slaughter module separated one animal from a group of cattle using a moving floor conveyor that transferred the animal to a moving conveyor between the animal’s legs as the floor dropped away (White, 1994). Two bails captured the neck and applied an electrical current to stun the animal. The electrical pathway was then altered to effect a spinal inactivation. A pneumatically powered knife with oscillating blades was used to enter the thoracic cavity and sever the aorta. Horns were also removed at this stage using hydraulic cutters. A prototype automated sheep stunning machine was developed in the mid 1980s in New Zealand (Authier, 1990). This machine was quickly commercialised

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 311 and is commercially available. For a variety of reasons including ritual slaughter, automated sheep sticking systems have not been successfully developed to date.

13.2.3 Shackling Once stunned, animals are manually shackled, usually with a chain loop around one hind leg, and hoisted to hang head down. A human operative then makes the ‘sticking’ cut to the artery in the throat to drain the blood. These shackling and cutting operations are complex and difficult to automate due to the complexity of the operations, the unstructured environment, the implications on downstream processes if performed incorrectly, and the need to maintain animal welfare if stunning fails.

13.2.4 Removal of hair or hide Pork carcass production differs from beef and sheep carcass production in that typically hairs are removed from the skin, whereas beef and sheep plants remove the entire hide/fleece from the carcass. The removed hide has value as a base product for the leather industry. Whilst it is usual to remove only the hairs from pork and leave the skin on the carcass, some pork plants also perform de-hiding for pig leather, although this is not a common practice. De-hairing pork Once drained of blood, pork carcasses pass through a sequence of mechanised operations, typically consisting of first a hot water or steam scald to loosen hairs, and then through a de-hairing machine where rotating metal-tipped rubber fingers scrape and brush most of the hairs from the carcass surface. This is followed by a singeing operation, whereby the carcass passes through gas flames to burn off remaining fine hairs. Finally, carcasses pass through a second ‘polishing’ station, where burnt hair stubs are removed by rotating rubber flail fingers. These operations avoid the need to adapt to carcass geometries by using techniques that conform to the product shape. Fingers on flexible rubber mounts, gas flames and water jets can all act on the carcass without detailed knowledge of surface position. This approach allows simple mechanisation to be used for these tasks. De-hiding beef The first task of beef de-hiding is to cut the hide along the belly from the crotch to the neck. This is a demanding task requiring a consistent cut typically 2 m or more in length, along the centre line of the carcass severing only the skin. Industrial Research Ltd (IRL), based in New Zealand has developed automation for this task (Templer et al., 2002). The profile of the belly is detected with an infrared laser distance sensor, and this information is processed to form a smooth trajectory for the cutting tool. The purpose-designed tool consists of a guidance spike mounted tangentially to a rotating circular knife. The spike protects the

© Woodhead Publishing Limited, 2013

312 Robotics and automation in the food industry underlying meat from cutting damage and serves as an anvil to improve the cutting efficiency. The tool is moved by robot to place the spike between the skin and meat and then follow the previously determined path to sever the hide along the belly. The system has been proven in a slaughterhouse in Nebraska, successfully cutting many thousands of carcasses. Whilst the initial development work used a purpose-built robot, later commercialisation work used an off-the-shelf food grade KUKA robot. Before this hide opening cut is made on feedlot cattle, there are often large ‘dags’ or deposits on the skin that must be removed. In 2000, Food Science Australia staff developed a hand-held dedagging tool. A later MLA project (MLA, 2012) sought to automate this process using a robot. The project was not successful due to problems restraining the carcass while the robot was operating. After the skin-opening cut is made, the hide is removed or ‘pulled’. Mechanical pulling arms supply the majority of the effort, but a human butcher is required to make specific preparatory cuts, attach the pulling mechanism and make assisting cuts during the pulling operation. The Fututech system used bed-dressing for hide removal where the carcass was resting on its back (White, 1994). After appropriate manual hide preparation the carcass was suspended from four hooks, one in each hock, while remaining in the supine position. The hide was removed automatically using a 3-stage process that involved (1) pulling the hide downwards, (2) separating the hide from the back fat using a blunt knife and (3) pulling the hide over the head and off the carcass. De-fleecing sheep The sheep de-fleecing, or pelting, process is complex and traditionally used 30% of the labour force on a sheep dressing chain (Longdill, 1984). Early attempts to automate this process were reasonably successful although the machinery was complex (Robertson, 1980). Researchers at MIRINZ developed a rotary pelting machine that automated the majority of the pelting process. The machine was physically large and operated on a rotary turret principle to achieve the required throughput. Commercial versions of this machine were installed in a number of sheep processing plants in New Zealand during the 1980s; however, none are operating now, as they were superseded by the technologies described below. Sheep pelting starts with the ‘Y-cut’, whereby initial incisions are made on the forelegs and chest. A robotic Y-cutting system was developed at IRL in New Zealand (Taylor, 1993) in the 1990s and this is now operating in several meat plants in New Zealand. Brisket clearing is another pelting subprocess most slaughter staff consider physically difficult that has been automated in New Zealand researchers. This is another example of a human operator being assisted by a relatively simple but powerful machine. The operator performs all the sensing and delicate positioning operations, which are not physically difficult, while the machine can reliably and repeatably perform the difficult and physically demanding tasks without tiring or becoming injured. A similar approach is used for the sheep-fleece shoulder puller

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 313 with a combination of skilled labour for sensing and positioning, and a machine performing the heavy work. The last in the series of pelting machines developed by MIRINZ is the final puller. This machine is deceptively simple in its operation, although its design and set-up is the key to its successful operation. This machine was released commercially in 1985 and has been installed extensively on lamb chains around the world.

13.2.5 Evisceration and dressing Once the hairs or fleece/hide have been removed, all carcasses are eviscerated, whereby the internal organs are removed. The viscera generally separate into edible (heart, liver, kidneys, etc.) and inedible (intestines and bowel). There are substantial hygiene implications of error when removing the inedible organs as the contents contain faeces and pathogenic organisms that can contaminate the carcass meat if spilt. Although all processes follow the general sequence of rectum loosening, belly opening, and viscera removal, there are substantial detail differences for example, trimming, washing, additional cutting, other organ removal, other processing, etc., between species and plants. Pork evisceration Work at DMRI in the 1990s (Fig. 13.1) and a later collaboration with SFK has developed automation for pork evisceration (Madsen and Nielsen, 2002). The equipment makes a few simple anatomical measurements that guide the process. These gross measurements allow for coarse positioning of the evisceration automation, and conformation of the flexible carcass or adaptive tooling is also used to reduce complexity and hence increase reliability of a relatively complex operation. A prepared carcass is automatically clamped open and an arm moves the viscera to expose the sternum, where a second set of arms loosens the leaf lard and severs the attachment of the diaphragm to the chest cavity wall. A back cutter is then moved into the carcass to penetrate the diaphragm adjacent to the spine and sever the connective tissue between the organs and spine in the hind section of the carcass. A tenderloin tool then moves in opposition to the leaf fat arms to separate organs from the thoracic cavity. Activating these automation motions concurrently resolves forces in the system and removes the need for forceful carcass clamping and fixturing. The released organs are then pulled forward out of the carcass with a horizontal movement of the tenderloin tool, the clamps are released and the carcass is moved out of the supports. The automated system ensures tools are washed before the next carcass arrives. This automated evisceration system performs all these operations in 10 s giving a line speed of 360 carcasses per hour. DMRI are currently working on equipment for the subsequent separation and sorting of the organs. Microbial analysis has shown that carcasses automatically eviscerated possess fewer pathogens (E. coli) and aerobes than conventionally eviscerated carcasses (Clausen, 2002).

© Woodhead Publishing Limited, 2013

314 Robotics and automation in the food industry

Fig. 13.1

Developmental pork evisceration system.

Beef evisceration Once the cattle hide has been detached, the abdominal cavity is opened and the organs removed. Part of the opening process involves sawing the sternum bone to gain full access to the chest cavity. IRL in New Zealand have been working towards automating this beef task (Templer et al., 2000). Through projects running over a number of years, the team have demonstrated first static, then line-synchronised, brisket sawing. Using the same robot and guidance system as the hide opening system, a reciprocating bone saw similar to, but more powerful than, a manually manipulated brisket saw, is moved down the centreline of the sternum. However, when implemented on a production line the equipment did not perform satisfactorily, as a large number of the carcasses had been damaged in the previous de-hiding process. This resulted in a twisted carcass at the sternum saw station. The automation was unable to cope with straightening the carcass and completing the cut in the 9 s cycle time available. The Fututech system used an automated evisceration system comprising a paddle that was pushed against the spine and pulled down the carcass to peel the viscera from the abdominal cavity and push it into the viscera tray for sorting (White, 1994). This mechanised approach relied heavily on compliance of the carcass and organs to a fixed trajectory.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 315 Sheep evisceration Researchers at MIRINZ developed a mechanical sheep evisceration system in the late 1980s (Authier, 1990). The system comprised a brisket and belly cutter, an eviscerator and offal handling system. The system was trialled in an Australian sheep processing plant and offered to several companies for commercialisation (Authier, 1994). Later trials using a variant of the pelting Y-cut robot were aimed at opening the brisket and the belly in a more conventional chain configuration. The technology has yet to be commercialised.

13.2.6 Splitting Pork and beef carcasses are generally split into right and left sides to ease handling and increase rates of chilling. Lamb and sheep carcasses are typically left unsplit. Automatic carcass splitting equipment has been available for many years. Systems are sold by such suppliers as Stork, SFK, Danfotech, Durand, Automeat, etc. These machines have a range of cutting actions and complexities. The basic systems use a simple downwards motion of a circular saw through the space where the carcass should be. A higher level of complexity uses a series of rollers to locally position the spine onto the cutting device. ‘Back finning’ is sometimes carried out as part of splitting for pork carcasses. This process reduces damage to the eye-muscle during the splitting operation by separating it from the dorsal spine ‘fins’ before splitting the carcass. An automated system using a relatively complex arrangement of rotary knives, plain blades and active rollers has been developed for this task in the Danish pork industry. Automation for beef splitting was among the first examples of mechanisation in the slaughterhouse, and many equipment manufacturers now include beef splitting machines in their product range. Whilst this equipment removes the arduous manual process, many users of the equipment are still dissatisfied with its performance in terms of accuracy of splitting down the centre of the spinal column and the hygiene aspects associated with deposition of bone dust and other detritus on edible surfaces of the carcass. The Fututech system included a module that automatically split a beef carcass into two sides using a guided bandsaw (White, 1994). Later work by Food Science Australia funded by Meat & Livestock Australia (MLA, 2012) used a robot to guide a band saw for carcass splitting. This equipment had a vertebrae sensing system based on ultrasound and this experienced difficulties on some carcasses due to voids caused by the hide puller disrupting the consistent passage of the ultrasonic wave necessary for ultrasound sensing. Most splitting equipment producers claim an increased accuracy of automatic carcass opening and splitting over human-based splitting operations. However, the experience of some users is that there is still deviation from the precise centre line of the carcass. This can cause problems for carcass inspection and subsequent automated systems using the spine as a reference or datum position.

© Woodhead Publishing Limited, 2013

316 Robotics and automation in the food industry 13.2.7 Automated inspection and grading Farmers and animal growers are typically paid based on weight and conformation (muscularity) of carcasses. Impartial, accurate and reliable automated grading automation has been the subject of much development activity for all species. Automated carcass weighing systems are common on most slaughter lines. Automatic grading and classification systems typically compare image(s) of each carcass against standard reference carcass images for the various grades. This is impartial and removes variation due to individual graders. The captured image can be stored and used for traceability, production management or process quality audit. For pork, DMRI have developed the Danish Carcass Classification Centre, and SFK produce automated grading system called AUTO-FOM (Madsen and Nielsen, 2002). Machine vision systems that are non-invasive are in development, but some studies show them to be less accurate in predicting saleable yield than existing technology (McClure et al., 2003). Whilst laboratory development systems show the potential for rapid, economic, hygienic, consistent and objective assessment systems, there are still limitations in the industrial environment (Brosnan and Sun, 2002).

13.2.8 Automated chill rooms Certain wavelengths of visible light can reduce shelf life and encourage rancidity of stored chilled meat (Field, 2004). Automation to move carcasses in darkened chill rooms could improve product quality through reducing a contamination route from the human operative to the meat and reducing the spoilage organism growth rate. This type of automated carcass loading and unloading system has been commonplace in the New Zealand sheep meat industry for the last 20–30 years.

13.3 Automation of carcass separation processes after primary chilling All slaughterhouses follow a similar sequence of operations to transform the live animal into meat for consumption. After slaughter, inedible and many non-muscle parts (skin, hair, intestines, etc.) are removed to produce a carcass which is then chilled to reduce spoilage rates. After chilling, the carcass is typically subdivided into smaller sections (‘primals’) for ease of handling. These primals are then processed (boned, trimmed, cut, etc.) to produce retail joints and portions. This section deals with automation of carcass preparation processes taking place after chilling.

13.3.1 Primal cutting After chilling, carcasses are commonly cut into smaller ‘primal’ sections that are then further subdivided into retail joints, boned out, or processed into a wide variety of end products. Automated and robotic systems to produce primals are of

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 317 major importance for meat processing because relative price differences between primals requires accurate and consistent control of cut paths to optimise overall carcass value. Pork primalisation Early work on robotic systems for pork primalisation was performed in Western Australia (Clarke, 1985). The system comprised a computer-controlled pork carcass break-up machine that automatically broke down a full carcass into eight pieces in less than 30 s. Simple automated cutting systems that separate a pork half carcass into fore, middle and hind sections were developed in Europe by DMRI and others in the early 1990s. The tenderloins, head and forefeet are manually removed as preparatory operations, then carcasses hanging on a standard gambrel are pulled across a conveyer belt and the hind feet cut off. This releases them from the gambrel onto the conveyor. At a second station each carcass side is moved against a datum surface and the length between the pubic bone and the foreleg is measured. This measurement is used to position circular saws further down the line to anatomically derived cut positions for that carcass side. A second machine is available for the longitudinal cut to separate the belly from the loin. A robotic solution that performs all cuts in a single system was developed in the early 2000s as part of an EU funded project (Purnell, 2004). The Advanced Robotic Technology for Efficient Pork Production (ARTEPP) system has been patented and is arguably the most advanced robotic meat production system available to date. Because of the need for accurate cut placement, compliance of the carcass is not used and each cutting path is specifically adapted to the individual carcass being processed. Significant interaction between various expert organisations in cutting-blade design, machine vision, robots, systems integration and meat production were required for the project to be successful. This project is described in more detail as a case study below. Case study: Development of the ARTEPP pork primalisation robot In initial R&D studies a purpose-built Cartesian food grade robot wielding a pneumatic cutting tool was used to make the cuts. Whilst successful cutting was demonstrated, several factors limited the industrial exploitation potential of the system. The gantry-based Cartesian robot was very large, did not withstand the rigours of the food production environment, and spare parts and engineering support were not readily available. The pneumatic cutting tool was prone to stalling at the high cutting rates possible with robotisation, thus negating some of the potential benefits. This illustrates some of the fundamental differences between human and machine performance of the same task. The physical strength of the slaughterman regulates the cutting process; as the cutting forces build up, the human slows, allowing the tool to make the cut. With a higher strength robot making the cut, the separation made by the pneumatic tool could not keep pace with the rate the robot was moving the tool along the path. Using a more powerful 3-phase electric saw and developing a special blade for high-speed cutting of meat and bone reduced these difficulties.

© Woodhead Publishing Limited, 2013

318 Robotics and automation in the food industry A second-generation ARTEPP system used off-the-shelf 6-axis KUKA KR125 anthropomorphic robot to avoid development time and costs within the project, and to produce a system with readily available spare parts and engineering support. A modular approach was taken with the carcass transport, orientation and support subsystem operating independently of the sensing, cut-path derivation and robotic cutter subsystems. This allowed standard subsystems to be used in different installations, with customisation only required in a few subsystems, thus reducing costs. In one implementation of the ARTEPP system, incoming pork sides are orientated to align the carcass split plane to the overhead rail. At the orientation station, use is made of a previous processing line feature in that the hook through the Achilles tendon always faces the split plane. An inductive sensor detects the hook and the side is rotated if required to place split plane facing the robot and vision measurement system. The side then indexes on to the cutting station where adaptive fixturing grasps the side and supports the carcass to resist cutting forces. Once clamped, structured light techniques are used to extract 3D locations of anatomically pertinent carcass features. Cut paths are defined anatomically relative to these features for every carcass separately. After cutting the clamps are released and the side ejected from the system. In another configuration of the system, where fewer cuts per carcass are required but at higher line speeds, an overhead rail and inclined conveyor carry each side past the vision sensing and cutting stations at a fixed speed. The image processing is performed as the side travels to the cutting station and the robot is synchronised to the line speed allowing carcasses to be cut whilst moving. Comparing the robotic system with manual primal cutting showed manual cut placement was within 20 mm of the anatomically desired location and 89% of manually cut carcasses had cut accuracy of better than ±5 mm. The robot system performed to better than ±5 mm for 97% of cuts (Fig. 13.2). The system can tirelessly produce consistent, anatomically accurate cuts. However, the most important commercial feature of the automated primal cutting system is its ability to finely adjust cuts in response to seasonal and market price Frequency

Automated cutting

Manual cutting 20 mm 5 mm 5 mm 20 mm Lower value primal Higher value primal Nominal cut position

Fig. 13.2 ARTEPP automation versus human cutting performance.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 319

Fig. 13.3 Production ARTEPP system.

fluctuations. Shifting cuts to favour high value primals (moving the ‘automation’ peak to the left in Fig. 13.2) can result in significant value improvements for each carcass. Other cost benefits connected with not having to find, train and retain human staff for the task are a bonus. The latest system (Fig. 13.3) has been used online for a full year in a Norwegian plant doing the work of three staff with a 3% yield improvement. This produces a payback in less than 18 months for equipment costs. Beef primalisation Mechanical boning aids that exert pulling forces while a human butcher makes key separation cuts have been used for many years. Whilst not at the forefront of automation technologies, these human augmentation systems have enabled higher throughputs with less physical effort for the same number of staff than using traditional individual cutting tables (Field, 2004). French researchers at the Institut national de la recherche agronomique (INRA) developed a prototype robotic system for subdividing beef forequarters (Damez and Sale, 1994). The system is relatively slow because major sensing and trajectory planning problems had to be solved. The prototype was successful in producing beef primals, but has yet to be developed into a commercial system. An ambitious beef sectioning system was proposed by the Texas Beef Group in a 1993 patent (O’Brien and Malloy, 1993). A chilled eviscerated carcass would be mounted horizontally on an automatic guided vehicle and appraised using X-rays, 3D machine vision and ultrasonic sensing. The results of the inspection would be used to generate cutting paths to enable the carcass to be cut into optimal primal sections. A robot would be used to effect this separation with high-pressure water, abrasive and air jets. Flesh would be cut with the water jet while the air jets would

© Woodhead Publishing Limited, 2013

320 Robotics and automation in the food industry keep the severed meat clear of the cutting area. The abrasive jet would be invoked when making cuts through bone. This is a particularly high-tech proposal in a patent and no record can be found of a prototype or commercial system. Sheep primalisation automation Up until the 1970s, most of the sheep and lamb traded internationally was in the form of frozen whole carcasses. After that time, carcasses began to be progressively broken down into a series of cuts, initially frozen and later chilled and vacuum packed. The main items of technology in a lamb boning room in the early days were band saws and packaging machines. Research at New Zealand’s AgResearch Crown Research Institute in the late 1990s was directed at the automation of cutting lamb carcasses into primals. The drivers were to produce clean, square cuts and hygienic handling of primals, to improve product yield and shelf life from subsequent processing. A prototype machine was produced during 2000 and further enhanced the following year by automatically locating bones within carcasses.

13.3.2 Boning Automation and dedicated machinery for boning out of specific pork sections are commercially available or under development in many parts of the world. Much of this work has been led by DMRI and their commercial partners and includes boning equipment for fore-ends and hindlegs, and combined boning and trimming equipment for belly and loins (Madsen and Nielsen, 2002). Robotic technology has also been used to separate pork flank ribs from the pork belly (Anon, 2000). The system uses a machine vision system to assess the size and shape of an incoming belly. The 3D data are used to calculate the cutting path. A Fanuc M710 robot equipped with a curved double-edged ‘Denver knife’ executes the path, pulling the shaped knife through the belly in the prespecified trajectory. The robotic cell includes automatic tool changeover and can select from eight different knives. When not in use knives are sterilised as part of the production process. This system can process 1400 bellies an hour, equivalent to a six-man crew. Final trimming and manpower requirements are reduced and the yield is optimised over both the belly and flank rib set. Automated systems for deboning specific beef meat sections are under development or in production, as with the other meat types. A beef rib deboning system has been designed and manufactured by Food Science Australia. This machine automatically strips the meat from a beef ribset in 21 s. Longdell (1996) describes other beef deboning machines for heads, loins and forequarters. All systems improve carcass yield, but the levels of sensing and adaptive automation are low, with most systems separating meat from bone with combinations of mechanical force, compliance in the equipment, different rigidities of meat and bone in the beef section, and bespoke shaped blades (Trow and Ng, 1994). A human butcher is required to operate the equipment and assist cutting in a similar manner to the primalisation pulling arms mentioned above.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 321

Fig. 13.4 (a) Cutting sub-system for beef foreleg deboning research equipment. (b) Beef foreleg deboning research system showing machine vision and lighting.

A vision guided, force-feedback controlled beef deboning research system has been constructed at the University of Bristol (Purnell et al., 1993). The laboratory based system (Fig. 13.4) demonstrated the technical feasibility of sensory guided robotic deboning, but further R&D would be required to bring the concept to a commercial reality. The technique made an initial 2D visual assessment of the beef joint, and sought to match that current meat section to a database of previous experience. If a match was found, the previous cut paths were replayed for the current meat section; if no match was found then force feedback from the boning blade was

© Woodhead Publishing Limited, 2013

322 Robotics and automation in the food industry used to guide the robot along the bone and in doing so create another experience example to augment the database. This process showed promise for the 2D deboning of beef forelimb taken as the example process. However, these initial concepts would need to be extended substantially to produce a fully automated beef boning line for commercial use. Research on automated sheep boning systems began at MIRINZ in the early 1980s as part of the mechanical boning project (Roberts, 1984; Wickham, 1988). The automated sheep boning process consisted of the following processes: 1. The load station lifted the carcass off the rail, removed the gambrel and loaded the carcass onto the carcass support. 2. The pedestal rotated the carcass support about the horizontal axis to present it to the boning head. 3. The linear drive cleared the pelvis by grasping and pulling the rear legs on the upward stroke. 4. On the downward stroke, a combination of rotating knives, flexible disks, ploughs and a moving wire separated the soft meat sides from the skeletal frame. 5. The skeletal frame was ejected at the pedestal during rotation of the carcass support. A programmable logic controller and range of sensors controlled the sequencing and the compliance was used in places to adapt cutting paths to carcass profile and deform the carcass to the blade trajectory. The production rate was estimated at 190 carcasses per hour with a payback period of less than 1 year. However, this machine was never commercialised. The frame boner laid the groundwork for a very successful second-generation boning machine (Wickham, 1990). A carriage was manually loaded with a loin section of sheep carcass, and then transported through a set of fixed knives followed by a set of semi-rigid plastic ploughs. The frame for the knives and ploughs could move vertically to partially accommodate different loin sizes. Further commercial machines to come out of the MRINZ mechanical boning programme were the chine and feather bone removal machine, rib frenching machine, shoulder fleecing machine and the shoulder boning machine (Ng, 1992, 1994; Wickham, 1992). A trunk boning machine and a leg boning machine operating on similar principles are commercially available (Macpro, 2010). In the trunk boning machine the trunk is conveyed away from the operator after manual loading where two blades clear the meat from the vertebrae approximately 50 mm either side of the centreline. The fleecing blades sweep around the ribs to separate the meat, while a second set of knives simultaneously clears tissue from the neck. In the leg boning machine the leg is placed vertically between shaped metal boning chucks. The two chucks move towards each other, boning the leg using a scraping and cutting action until the two chucks meet. The bone is finally ejected through the lower chuck.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 323 Scott Automation, in association with meat processor PPCS, has developed a robotic system for boning lamb legs (Templer, 2004). A KUKA robot is been fitted with a boning knife incorporating force feedback allowing the robot to guide the knife along the bones of the lamb leg.

13.3.3 Slicing and portioning The development of slicers illustrates the importance of sensing for benefits of automation, particularly in slicing for multi-slice packs where control of individual slice weight is desired. Initially meat slicing was carried out by hand, as this was the only method possible, then mechanised fixed slice thickness machines with much higher throughput rates came to the fore. Further reductions in giveaway were gained with the first generation of automated slicing machines that changed thicknesses by scaling a slicing pattern from input meat section weight. This gave advantages over fixed-increment mechanised slicing. However, because of the inherent variations in weight–length ratio of nearly all meat sections some giveaway losses were still apparent. The latest 3D scanning slicing machines make the next step with greater sensor data to improve performance further. A full 3D representation of input meat section is gained using a laser stripe or similar technique, which allows for better utilisation and less giveaway. However, the slice angle is often constant, limited by the mechanical arrangements of the equipment to maintain production rates. There is a trade-off between machine complexity, line speeds and yield attainable.

13.3.4 Trimming People are becoming increasingly health conscious and consequently there is a growing demand for lower fat products. In traditional manual trimming each individual operative makes the decision on how much fat to remove and then makes the cut using a standard butcher’s knife. This often results in straight cuts to trim fat from a curved surface and a high degree of variability between individual trimmers. Water-jet based trimming machines have been on the market for some time but high capital costs requires relatively high throughputs, which consequently suit only higher volume production. Many meat producers are too small to afford and benefit from this trimming technology. Additionally, there are specific problems with meat section geometry for some sections. Water-jet cutting is most commonly used on ‘planar’ food sections, such as beef steaks, pork chops, chicken breast fillets, pork bellies, etc., that are laid flat on the input feed conveyor and remain stable in this position through sensing and cutting. For some sections, such as lamb or pork chops, the narrow ‘tail’ of the chop can twist under its own weight to lie flat on the conveyor, thus preventing effective trimming. Research at FRPERC (Purnell and Brown, 2004) developed specific equipment for trimming lamb chops. Machine vision provided fat thickness profile along the length of each chop. The system then conformed the meat section to

© Woodhead Publishing Limited, 2013

324 Robotics and automation in the food industry Conforming wall to change shape of chop Required cut path at fixed distance from fat–lean interface

Clamp force holding chop against wall

Simple path followed by cutter

Fig. 13.5 Uniform fat trimming concept.

Fig. 13.6 Trimmed lamb chops.

place the fat–lean interface at the desired fat thickness from a fixed cutting path. The cutting path was thus simplified and could be made with basic tools and an uncomplicated trajectory (Fig. 13.5). After the cut is complete, the meat section is released and returns to its natural shape but with a uniform covering of fat over the fat–lean interface (Fig. 13.6). Industrial plant trials with the demonstrator system have achieved improved accuracy and product appearance over manual trimming.

13.4 Future trends Until relatively recently, technical issues associated with adapting sufficiently accurately to the inherent biological variations in meat limited the potential for

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 325 automation in the meat processing sector. Many of these technical issues have now been resolved and demonstrated, and commercial business issues will be the major influence on implementation and uptake of meat processing automation in the near future. Yield control, legislation, difficulties in staff availability will increase commercial pressures and encourage more meat processor organisations to automate, simply to maintain throughput. The economic breakeven point for implementing automated slaughter lines will be affected by an increasing cost of not automating, coupled with a decreasing cost to automate. As automation levels rise, the staff skill levels will rise accordingly. As the pressures imposed by the global marketplace, regulatory agencies, distribution channels, media and customers increase, slaughterhouse staff with profiles closer to surgeons, and skilled automation engineers by their education, training and working habits will begin to emerge. Due to development costs, new automation meat systems are expected to arrive in a piecemeal manner, rather than as major projects addressing the entire slaughter line. However, as a result more automation subsystems will become available reducing one current barrier of technology cost. These pockets of automation will have significant impact in small areas in their specific roles, but widespread automation will not occur immediately. Pork slaughter automation is a possible exception with fully automated lines expected in next decade.

13.5 Conclusion The development of automated meat processing systems has received substantial effort and investment, but uptake has been limited by technical and business issues. With improved technology and reducing production costs for automation, it has become increasingly possible to overcome these limitations. The potential advantages and rewards to the meat industry have resulted in a considerable number of process-specific applications and continue to drive R&D of more sophisticated systems. Despite the differing operations for beef, pork and sheep meat production, some general trends are common across a number of projects. Initially many meat automation research projects developed bespoke robots for their particular task (Maddock et al., 1989; Wadie et al., 1995; Taylor and Templer, 1997; Templer et al., 2002). In these projects as the developments neared commercialisation, the teams changed direction to using standard industrial robots, protected against the rigours of the food production environment. In conjunction with, and in some cases as a result of, these developments, most robot manufacturers now supply off-the-shelf food-grade robots. This in turn provides food sector appropriate subsystems for equipment integrators, easing development, operation and maintenance of robotic meat processing systems (Ranger et al., 2004). Food companies that have been successful in introducing automation tend to have good working relationships between all grades of staff and have longer-term financial viewpoints.

© Woodhead Publishing Limited, 2013

326 Robotics and automation in the food industry A modular approach has been proved worthwhile at both process and individual task levels. DMRI seek to automate all pork production tasks through developing a series of modular components, each performing a different task in the slaughter process. This has allowed a number of different projects and partnerships to be established, leading to more flexibility in implementation for both the automation user and supplier. The ARTEPP primal cutting system uses modular subsystems to accommodate variations in plant specific processes. The series of lamb boning machines developed by MIRINZ use similar mechanised approaches. Some automation systems have been successful because they perform tasks currently not possible for a human operative. A human butcher could not perform the multi-armed cutting and handling operations achieved by evisceration automation. Even the strongest, most skilled butcher cannot match the consistency and high-force cut accuracy achieved with automated primal cutting. Automation of these types of tasks, unperformable by a human, are often the first to exhibit an acceptable cost–benefit ratio. Currently it is mostly uneconomic to replace a slaughterhouse operative with automation unless the automation yields additional benefits. Some successful projects have demonstrated an improvement over manual labour in terms of speed, consistency, accuracy and control.

13.6 Sources of further information and advice ABB robotics: http://www.abb.com/robotics Danish Meat Research Institute (DMRI): http://www.dti.dk/services/danishmeat-research-institute/31729 Food Refrigeration and Process Engineering Research Centre (FRPERC): http://www.frperc.com, http://www.grimsby.ac.uk/Industry/FRPERC.php KUKA robotics: http://www.kuka-robotics.com Meat and Livestock Australia: http://www.mla.com.au Meat Industry Research Institute of New Zealand (MIRINZ): http://www. agresearch.co.nz/mirinz SFK systems A/S: http://www.sfk.com

13.7 References Annan, D. 1982. MIRINZ manual dressing system. Proceedings of the 22nd Meat Industry Research Conference, Hamilton, New Zealand, 37–39. Anon. 2000. Slaughterhouse slashes costs and pork bellies. Advanced Manufacturing, 2 (1), 43–44. Publ. Clifford Elliot Ltd, Burlington, Ontario, Canada. ISSN 1481–8354. Authier, J.F. 1990. Mechanical dressing developments. Proceedings of the 26th Meat Industry Research Conference, Hamilton, New Zealand, 225–231. Authier, J.F. 1994. Lamb evisceration by machine. Proceedings of the 28th Meat Industry Research Conference, Auckland, New Zealand, 187–191. Brosnan, T. and Sun, D-W. 2002. Inspection and grading of agricultural and food products by computer vision systems – a review. Computers and Electronics in Agriculture, 36, 193–213.

© Woodhead Publishing Limited, 2013

Robotics and automation in meat processing 327 Brown, T., James, S.J. and Purnell, G. 2005. Cutting forces in foods – experimental measurements. Journal of Food Engineering, 70(2), 165–170. doi:10.1016/j. jfoodeng.2004.09.022 Butina. 2010. http://www.butina.eu (Accessed August 2010). Channon, H.A, Payne, A.M. and Warner, R.D. 2002. Comparison of CO2 stunning with manual electrical stunning (50 Hz) of pigs on carcass and meat quality. Meat Science, 60(1), 63–68. Clarke, P.T. 1985. Automatic break up of pork carcasses. Agri-Mation 1, Proceedings of the Agri-Mation 1 Conference & Exposition. ASAE Publication. 01-85, 183–189. Clausen, V. 2002. Automation in the pork industry. In Proceedings of The Nordic Veterinary Congress, Helsinki, Finland, 27–29 November 2002. Damez, J.L. and Sale, P. 1994. Studies on automation of cutting of the forequarter of cattle carcasses (In French). Viandes et Produits Carnes, 15(4), 103–107. Field, M. 2004. Evolution in practice. New Food, (Issue 4), 41–43. Grandin, T. 2008. Effect of genetics on handling and CO2 stunning of pigs. Meat Focus International, July, 124–126. (with May 2008 updates). http://www.grandin.com/ humane/meatfocus7-92.html (Accessed August 2010). HSE. 2008. http://www.hse.gov.uk/food/slaughter.htm#a1 (Accessed September 2012). Khodabandehloo, K. and Clarke, P.T. 1993. Robotics for Meat Fish and Poultry Processing. Blackie Academic Press. ISBN 0 7514 0087 4. Longdell, G.R. 1996. Recent developments in sheep and beef processing in Australasia. Meat Science, 43, nS165–nS174. Longdill, G.R. 1984. Advances in Mechanical Dressing Technology. Proceedings of the 23rd Meat Industry Research Conference, Hamilton, New Zealand, 127–130. Macpro. 2010. http://www.macpro.co.nz/mutton.html (Accessed August 2010). Maddock, N.A., Purnell, G. and Khodabandehloo, K. 1989. Research in Application of Robotics to Meat Cutting. Proceedings of 20th International Symposium on Industrial Robots (ISIR). Tokyo, Japan, 4–6 October, 957–963. Madsen, K.B. and Nielsen, J.U. 2002. Automated meat processing. In Meat Processing: Improving Quality, Kerry, J., Kerry, J. and Ledward, D., eds. Woodhead Publishing Ltd, Cambridge, England. ISBN 1-85573-583-0. Madsen, N., Nielsen, J.U. and Mønsted, K. 2006. Automation – the meat factory of the future. Proceedings of 52nd International Congress of Meat Science and Technology (ICoMST), Dublin, Eire. 13–18 August. ISBN-10: 90–8686-010-9 / ISBN-13: 978-908686-010-4. McClure, E.K., Scanga, J.A., Belk, K.E. and Smith, G.C. 2003. Evaluation of the E+V video image analysis system as a predictor of pork carcass yield. Journal of Animal Science, 81, 1193–1201. MLA. 2012. http://www.redmeatinnovation.com.au/innovation-areas/new-technologies/ beef-slaughter/products. (Accessed September 2012). Ng, W.Y. 1992. Advantages on machine boning of lamb. Proceedings of the 27th Meat Industry Research Conference, Hamilton, New Zealand, 322–326. Ng, W.Y. 1994. Lamb shoulder fleecer and rack frenching machines. Proceedings of the 28th Meat Industry Research Conference, Auckland, New Zealand, 201–208. North, S. 1991. Public health and safety aspects. DTI Seminar: Robotics for the meat processing industry. University of Bristol, Bristol, UK, 27 September. O’Brien, W.H. and Malloy, J. 1993. Method and apparatus for automatically segmenting animal carcasses. US Patent 5205779, 27 April 1993. Pigsite. 2009. Automated Production and Logistics. http://www.thepigsite.com/articles/13/biodigestion-biofuels/2917/environmental-issues-control-slaughter-production (Accessed August 2010). Purnell, G. 2004. Robotic Automation for Pork Primal Cutting. Food Technology International. ISBN 1-85938-535-4 / ISSN-1460–3179, 54–58.

© Woodhead Publishing Limited, 2013

328 Robotics and automation in the food industry Purnell, G. and Brown, T. 2004. Equipment for controlled fat trimming of lamb chops. Computers and Electronics in Agriculture, 45(1–3), 109–124, doi:10.1016/j. compag.2004.06.004 Purnell, G., Maddock, N. and Khodabandehloo, K. 1993. Robotic deboning: A fundamental approach to engineering a system. Proceedings of Artificial Intelligence in Food and Agriculture (AIFA 93) Conference, Nîmes, France, 27–29 October. Ranger, P., Ottley, G. and Smith, N. 2004. The pitfalls of purchasing. Food Processing, October, 20. Roberts, C.A. 1984. Mechanical boning and trimming of mutton. Proceedings of the 23rd Meat Industry Research Conference, Hamilton, New Zealand, 90–93. Robertson, A.A. 1980. Mechanical pelting: Introduction. Proceedings of the 21st Meat Industry Research Conference, Hamilton, New Zealand, 24. Taylor, M.G. 1993. Automated Y-cutting of sheep carcasses. Meat’93. The Australian Meat Industry Research Conference, Gold Coast, Australia. Taylor, M.G. and Templar, R.G. 1997. A washable robot for meat processing. Computers and Electronics in Agriculture, 16, 113–123. Templer, R. 2004. Cutting and Boning | Robotics and New Technology, in Encyclopedia of Meat Sciences, Ed. Jensen W.K., Elsevier Ltd., Oxford, UK, 381–388. Templer, R., Nicolle, T., Nanu, A., Osborn, A. and Blenkinsopp, K. 2000. New automation techniques for variable products. In Proceedings of Meat Automation Congress (MAC). Malaga, Spain, June 2000. Templer, R., Osborn, A., Nanu, A., Blenkinsopp, K. and Freidrich, W. 2002. Innovative robotic applications for beef processing. In Proceedings of Australasian Conference on Robotics and Automation (ARAA), Auckland, New Zealand, 27–29 November, 43–47. Trow, D. and Ng, W.Y. 1994. Beef and pork loin boning. Proceedings of the 28th Meat Industry Research Conference, Auckland, New Zealand, 187–191. Wadie, I.H.C., Maddock, N., Purnell, G.L., Khodabandehloo, K., Crooks, A., Shacklock, A. and West, D. 1995. Robots for the meat industry. Industrial Robot, 22(5), 22–26. White, R.M. 1994. Fututech. Proceedings of the 28th Meat Industry Research Conference, Auckland, New Zealand, 165–168. Wickham, G. 1988. Progress with boning and cutting technology. Proceedings of the 25th Meat Industry Research Conference, Hamilton, New Zealand, 160–164. Wickham, G. 1990. Developments in mechanical boning and cutting. Proceedings of the 26th Meat Industry Research Conference, Hamilton, New Zealand, 232–237. Wickham, G. 1992. An overview of boning machinery developments. Proceedings of the 27th Meat Industry Research Conference, Hamilton, New Zealand, 317–321.

© Woodhead Publishing Limited, 2013

14 Robotics and automation in the poultry industry: current technology and future trends G. McMurray, Georgia Tech Research Institute, USA

DOI: 10.1533/9780857095763.2.329 Abstract: This chapter discusses the current and future state of robotics and automation to enable a more efficient poultry processing facility. The chapter begins with a discussion of the unique challenges to automating poultry processing as compared to automobile or electronics manufacturing. The chapter then discusses the current production process with a focus on available solutions as well as prototype systems developed in academic labs. Finally, the chapter ends with a look into the future to see how robotics and automation can help maximize yield and improve food safety while minimizing labor costs and the use of natural resources such as water and energy. Key words: robotics, automation, poultry industry.

14.1 Introduction The modern poultry processing facility is an impressive combination of high technology solutions that have been arranged with the sole purpose of economically delivering a safe and affordable product to the customer. In the United States alone, the chicken industry, which is the largest component of the poultry industry, processes over 8.8 billion birds a year. As this number has climbed over the years, the industry has increasingly turned to robotics and automation to allow it to process birds in a faster and more efficient manner.

14.1.1

Unique characteristics of poultry processing and associated challenges for automation In the modern chicken processing plant, the slaughter and evisceration line is typically running between 140 and 180 birds/min (3 birds/s). At these speeds, it is

© Woodhead Publishing Limited, 2013

330 Robotics and automation in the food industry critical to have some automation because people do not perform these tasks well at these speeds. Typically, without causing injury to the worker, the upper end in speed for people tends to be in the range of 2–3 s per task. While these tasks were originally designed for the workers, the increase in production rates has made it difficult for the workers to keep up. Compare this to other industries, such as the automotive and electronic industries where robotic systems have been very successful. In those industries, the speed of the tasks is much slower – anywhere from 10 to 30 s per task. Also, they are typically doing tasks that the typical worker does not do very well – be it lifting heavy objects (hoods of cars, engine blocks, etc.) or very accurately positioning components. In addition, the cost of an error is relatively large. If a tray of computer wafers is dropped or incorrectly positioned, then the entire tray might be lost at a cost of several millions of dollars. In addition, the robot is typically handling product that is very dry, rigid, and its physical attributes are very well defined. In the poultry industry, the product is very different from that of the automotive and electronic industries. Birds vary rather dramatically in size, and the product is not rigid (especially when dealing with cut-up product like boneless breast fillets, and even bone-in-product is not as rigid as manufactured components due to compliance in the meat). This makes the use of traditional clamping style end-effectors difficult. In addition, the use of hard stops to precisely position product for handling does not work well, due to variations in product size and the compliance of the material. Further, the product can be slippery (boneless product or whole bird). This makes the development of end-effectors (hands) for robots very difficult. Washdown is also a major driver in the design of any electro-mechanical system for the poultry industry. Unlike other industries, all equipment in a poultry processing plant must undergo a high-pressure washdown with the application of chemicals to the surface. Washdown typically will consist of using high-pressure water (at approximately 1000 psi (6895 kPa) at 120°F (49°C)) and the use of chemical cleaners/foams that have high acid and high base pH levels. The reason for the high-pressure washdown is one of the largest drivers for the entire industry: food safety. The extent that this affects all equipment cannot be overstated. The USDA and the National Sanitation Foundation (2002, USDA, 2001) have both issued very detailed guidelines on machine design for food processing. These requirements include the materials used, the types of fasteners that can be used, and how mechanical elements are to be joined. All of these factors have a tremendous impact on food safety, complicate the design of any electro-mechanical system and establish a set of ambitious requirements (AMI, 2003). Differences in poultry processing around the world Poultry processing is generally the same around the world, but there are some differences. Some of these are related to regulatory issues in the various countries, while others are due to local specialty products or to the types of birds raised in that area. For the most part, the initial processing of the bird (live hang, kill, and evisceration) are the same around the world. Most of the differences arise in the

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

331

choice of chiller (water or air) and the further processing area (deboning and creation of the final product). One significant difference between European and American production is the chiller. In the United States and most of South America, the water chiller is the method of choice to reduce the slaughtered bird’s body temperature down to 40°F (4.5°C), immediately after slaughter. The chiller provides an excellent thermal transfer coefficient such that the process can be done relatively quickly and with a small footprint. In Europe, an air chiller system is used to cool the birds. This method reduces the chance of cross-contamination between birds from immersion in a water bath. However, care must be taken to prevent birds from dripping on other birds throughout the air chiller. This reduction in cross-contamination is achieved at the expense of chilling time and floor space. In further processing, the method of deboning the bird can vary based upon the preferences of the customer. Leg deboning in America is very different from leg deboning in Japan, for example. This unique market requirement dictates the design of the equipment and, in some cases, whether or not the cut can be automated.

14.2 Robotics and automation in live hanging and first processing of poultry In the modern processing plant, the initial tasks of getting the live birds onto the processing line, killing the birds and eviscerating the birds (first processing) are key tasks. These tasks are fairly standard in the industry. While the live hang process still remains a huge challenge, first processing is very automated. This section will specifically address one method to automate the live hang process and by doing so, it will illuminate the complexities of the task. 14.2.1 Live hang Birds are brought to the processing plant live and the goal of the first step in the modern processing plant is to hang birds onto the shackle line by their feet. The typical plant is running the live hang shackle line at 180 birds/min in the United States and over 240 birds/min in Europe. The task is laborious and very unpleasant for the workers as well as a potential safety hazard due to the dust, feathers and feces that are in the air. It is also dangerous because the birds want to peck and scratch at anything that grabs them. In an attempt to keep the birds calm, the live hang process typically occurs in an enclosed room that uses only red lights for the workers to see. Birds do not see well in red light so they tend to be more relaxed. In a processing plant, this is often one of the higher paid jobs because of the harsh working conditions. Automating this task has been a priority for a number of companies and inventors for a good half a century. The economic justification for automating this task is clear: the salary for the worker, the extremely high turnover (in some plants, it is over 200% for this task), the safety of the worker and the safety of the birds. From an examination of the requirements for this task, it is clear that this is an extremely difficult task to automate.

© Woodhead Publishing Limited, 2013

332 Robotics and automation in the food industry

Fig. 14.1

Live transfer.

The main difficulty in automating this task is obvious: the system must handle live birds. Live birds do not sit still on the conveyor and the designer must also take into account animal welfare. These are two very important points to consider in the design. Potential solutions for this application are numerous and plentiful, but none to date have solved the problem. An exhaustive description of all the solutions attempted to solve this problem is outside the scope of this chapter, but a reference is provided. At Georgia Tech, we have been looking into this problem for many years. Dr. Lee, the leader of this effort, and his team (consisting of researchers from both Georgia Tech and the University of Georgia) have developed a concept based on the following distinct steps: singulation of the birds, orientation, leg capture, and inversion (Lee, 2001; Webster and Lee, 2002). This process would ideally be done in the dark or in a red-lit area to put the bird in as relaxed state as possible. Figure 14.1 is a picture of the prototype system being tested. For the singulation step, a variety of approaches were considered. The team settled on a design that utilizes a set of counter-rotating drums with compliant fingers similar to those found in the defeathering process. These fingers are rubber and are pointed outward from a rotating cylindrical surface. The rotating fingers allow for only a single bird to pass at a single time without damaging the birds. The reaction of the bird to the fingers has also been the subject of much research (Lee et al., 1999, 2001). At this point, the key is to use a light that relaxes the bird but does not distract them. As the bird exits the singulation system, it automatically aligns itself with the motion of the conveyor in one of two manners – either facing forward on the

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

333

conveyor or facing backwards. Utilizing a vision system (either an IR camera if the room is completely dark or a regular CCD if the room is lit by a red light) to recognize the orientation of the bird, the system rotates the bird while in the grasp of the compliant grippers so that they are all facing forward (Lee et al., 2000; Li and Lee, 2005). (The IR camera is the special requirement for the system to work. IR cameras do not need any additional lighting or special consideration other than avoiding hot spots in the thermal image.) The key variables to prepare the bird for the leg-capture phase are the difference between the speed of the body of the bird in the grasper relative to the speed of the conveyor (called the translational velocity), the posture of the bird, and the lighting conditions. The importance of the translational velocity was a key discovery for the team. This also led the team to begin to analyze the kinematics of the legs of the bird in order to determine an optimal translational velocity for each individual bird. The translational velocity of the bird is an important component of the bird’s natural reaction to extend their legs downward in order to maintain contact with the downward sloping conveyor or retract their legs into their body. The desired response from the bird is to have it extend its legs so that they can be placed into the shackle. If the translational velocity is too high, the bird’s natural reaction will be to pull its legs up and into its body. If the velocity is too low, then the bird will stumble and possibly fall, thus making the loading into the shackle impossible. To customize this for each bird, a vision system was developed to identify the physical parameters of the bird as well its posture as it approaches the compliant grasper. A neural network is then used to optimize the rotational velocity of the compliant grasper to achieve the desired translational velocity for each bird. The results of preliminary testing of this system validate the described approach, but much more testing is needed (Webster and Lee, 2002). The next step in the process is the actual leg-capture phase. As the bird exits the compliant grasper, its legs should be extended and it can be placed directly into a moving shackle that is placed horizontally on a conveyor. Once in the shackle, the conveyor drops away and the shackle is free to rotate into the vertical position and the bird is successfully hung on the shackle (Shumway, 2002). It should be noted that this system is still very much in the research and development phase. Initial testing of the system has shown the validity of the core concepts, but a complete system has not yet been developed and tested.

14.2.2 First processing First processing is typically defined as slaughter, evisceration and chilling. These are standard processes that are almost identical in every poultry plant. This has been driven by a number of different factors, and has resulted in a certain amount of standardization of the equipment to perform these tasks. This equipment is categorized as fixed automation – mechanized machinery to perform fixed and repetitive operations to produce a high volume of similar parts (AMI, 2010).

© Woodhead Publishing Limited, 2013

334 Robotics and automation in the food industry As discussed earlier, one of the biggest problems for processors is dealing with the size variation of the birds. However, instead of adjusting the equipment to each bird, the equipment is adjusted based on the estimated weight for an entire flock. Thus, as a flock moves through the process, the machines are manually adjusted for the weight range. While this might seem rudimentary, the system actually works quite well. The majority of tasks performed during this part of the process do not generally require the accuracy that robotics or sensor-based automation offers. While the equipment for this process is relatively expensive, the actual cost to process a bird is very low. Given the economics of the situation, it is hard to justify additional automation that is an incremental change to the system. The main producers of equipment for first processing include Marel/Stork, Meyn, and Baader.

14.3 Robotics and automation in second processing of poultry Once the bird exits the chiller, it moves into second processing. This is defined as the deboning and portioning of the product, as well as packaging of the product into tray packs or bags/pouches. While first processing has almost been universally standardized, second processing is more defined by the product types and product mixes that a particular processor runs. Thus, each plant is unique. Second processing is an area that has been the subject of considerable development of robotic and automated systems (Wyvill, 2005). These are typically tasks that the worker does not naturally do well because of processing speeds and accuracy requirements. While most of these tasks are justified on labor replacement, some of these tasks, such as breast deboning, can affect the yield of the product. In this specific case, the losses due to manual deboning can result in yield losses totaling several millions of dollars a year per plant. This is a strong incentive to automate this process.

14.3.1 Rehang after chiller As discussed earlier, water chilling is not found in Europe since they use air chilling and the birds remain on the shackle during the entire process. In the United States and many other places, they still rely on immersion of the birds in a cold water bath to reduce their body temperature to under 40°F (4.5°) at exit. The birds are placed in the chiller by removing them from the shackle and dropping them into the chiller where they are typically, but not exclusively, moved through the process via an auger. It should be noted that at this point in time, the bird has not been split into the front half and legs. It is typically referred to as a WOG (without giblets). It is important to note that once it is removed from the shackle, the pose (position and orientation) of the bird is lost. Upon exit from chiller, the birds exit via a slide onto a table or circular conveyor system. At this point, the pose of the bird must be recovered so that it can be manually rehung onto the shackle line. This

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

335

Fig. 14.2 Robotic shackle loader.

typically requires somewhere between 4 and 8 people per line to meet the production requirements. The task also requires that the workers hang the birds in a particular orientation on the shackle line to enable further processing. At the Georgia Tech Research Institute, we have been active in developing an automated solution to this task. We have attempted to develop systems to automatically rehang the birds in two distinct manners: a fully robotic solution and a sensor-based fixed-automation solution. The fully robotic solution was intended to be a stand-alone solution to this problem (Fig. 14.2). The system was designed to identify the position and orientation (the orientation of the legs as well as whether the bird was breast up or breast down) of the bird on a moving conveyor, with software to track the moving bird, an end-effector to grasp it, and a system to track the moving shackles. The system used a KUKA 15SL, a stainless steel washdown robot, with their conveyorTech software for tracking the bird and the shackle. The end-effector was a particularly challenging part of the project. The natural variations of the bird, as well as the requirement to grasp birds that are either breast up or breast down, makes the design of a general purpose end-effector very difficult. Also, the legs must be kept a specified distance apart so that they can be placed into the shackle. The final design selected is sufficient, but clearly additional work is required prior to commercialization.

© Woodhead Publishing Limited, 2013

336 Robotics and automation in the food industry The system as developed did demonstrate several key concepts. First, the team was able to develop image processing software to accomplish the task. The system could identify the orientation of the legs on the conveyor and it could determine if the bird was breast up or breast down with a high level of accuracy. In addition, the project demonstrated how to handle a WOG without damaging the bird and a method to keep the legs in a position that allows for the hanging process. However, the system was not economically feasible. The cost of the robot and vision system was too high and the cycle time of the robot too slow. From our initial tests, the cycle time was between 7 and 9 s. A second design effort was undertaken to develop a solution using only fixed automation to accomplish the same task. The birds are first singulated using a commercial system and fed individually into the mechanism. The system has been divided into a two step process. The first step is to put the randomly oriented bird into a specially designed box – breast up with the legs protruding from the box. The second step is to transfer the bird from the box onto the shackle. The first step is shown in Fig. 14.3 and consists of the following procedures: 1. WOG slides onto a platform. Due to the physical properties of the bird, the slide orients the birds either legs first or legs last. 2. An imaging system takes a picture of the bird to determine if the bird is legs first or legs last and whether it is breast up or breast down. 3. The bird is rotated, based on the orientation as determined in step 2, such that it is pushed legs first into the flipping mechanism on the right side of the picture.

Fig. 14.3

Non-robotic rehang system, first step.

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

Fig. 14.4

337

Non-robotic rehang system, second step.

4. The bird is oriented such that it is breast up. 5. The bird is dropped into a specially designed box. The second step is shown in Fig. 14.4 and consists of the following processes: 1. The box from each loading station is integrated into a single line of boxes on a conveyor. 2. The box is transferred to the shackle loading system. 3. The bird is placed on the shackle line. 4. Empty boxes are removed from the shackle loading system back onto the conveyor. This system is more attractive in terms of cost–effectiveness because it relies on relatively simple fixed-automation concepts. The only complicated part of the system is the image processing, but that problem was solved during the development of the fully automated solution discussed above. The problem with this system is the footprint of the system. The conveyors to move the boxes through the process and manage a buffer result in a footprint slightly larger than the current process. Given the congested nature of most plants, this could be a significant problem. Further work is needed to minimize the footprint while maintaining the adaptability of the system.

© Woodhead Publishing Limited, 2013

338 Robotics and automation in the food industry 14.3.2 Cone loading Following the rehang after the chiller, the birds are typically automatically sorted on the shackle line based on their weight to meet the particular needs of each processing plant. Once the birds are sorted, each bird is split in two by running it over a fixed blade. This creates a front half (wings and breast meat) and the bottom half (legs and thighs). Each half is processed in different manners. The front half is typically placed onto a cone for either manual or automatic deboning (see Fig. 14.5). This task involves taking a stream of randomly located front halves and placing them on a cone. The placement of the bird on the cone is important. The goal of this is to place it such that a pin that is located on the cone (called the keel pin) penetrates a specific spot just above the rib cage. If this is done properly, the keel pin will hold the bird onto the cone and prevent it from rotating or moving during processing. For automated deboning processes, the placement of the bird is very important. Therefore, most companies have some mechanism that is used to pull it into the proper position on the cone. Most of these devices are cam operated and are simply fixed pieces of automation. There is not sensing or modification of the device based on the product. This can result in breaking or damaging the skeleton if the bird is not properly positioned on the cone. At GTRI, a team designed an end-effector to work on the front half of the bird (Fig. 14.6). This task is complicated by several factors (Socha et al., 2004). First, the cavity of the bird must not be compromised for several reasons. First, the cavity must not reduce in diameter such that the bird could not be placed on the cone or prevent the keel pin from penetrating the desired location. Second, the grasping must not damage the wing in any manner. A vision algorithm was also developed to identify the orientation and position of the bird on the conveyor. For this work,

Fig. 14.5

Cone loading gripper.

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

Fig. 14.6

339

Robot placing front-half on cone.

GTRI researchers used stereo and time-of-flight (TOF) cameras to determine the 3D shape of the product on the conveyor. From the shape data, it was possible to identify the long axis of the bird as well as whether it is laying breast up or breast down. Because of the moments of inertia, the bird will not typically rest on its side when it is fresh. From this data, the algorithm also estimated the target point for the keel pin on the bird to facilitate the placement of the bird on the cone. As can be seen in Fig. 14.6, the team was successful in developing an endeffector and imaging system to accomplish this task. The current system uses a commercial ABB robot to move the end-effector. A logical next step in this development would be to design and develop a commercial system to accomplish this task.

14.3.3 Breast deboning Deboning operations are one of the largest users of on-line labor in today’s poultry plants. Efforts have been made over the years to automate this function, but to date they have achieved only limited success. The main difficulty in this task is its unstructured nature due to the natural variability in the sizes of birds and their deformable bodies (Fig. 14.7). To increase product safety and quality, the industry is looking to robotics to help solve these problems. This research has focused on developing a new method of automating the deboning of bird front halves. If this task can be automated, the technology would naturally be extended to other cuts and trimming operations in poultry and red meat. The value in accomplishing this work would be not only to reduce labor costs but also to increase the yield of breast meat and reduce/eliminate bone chips. It is estimated that an increase in yield of a single percentage point could represent

© Woodhead Publishing Limited, 2013

340 Robotics and automation in the food industry

Fig. 14.7 Poultry deboning line.

several millions of dollars of additional revenue for each and every plant. Current attempts at automation of the shoulder cut impose several percentage points of yield loss in return for lower labor costs. Currently, automated solutions are offered by the major poultry processing companies including Marel/Stork, Meyn, and Baader. These systems typically rely on a variety of mechanical adjustments to perform the shoulder cut. These adjustments are to account for the size variation of the birds based on lot sizes, as discussed earlier. However, the size variation of the bird at this step is particularly difficult to adjust for without resorting to robotics. The yield of these systems is approaching that of the average worker, but only if the machine is kept in constant adjustment based on the size of birds being run at that time. Actual practice has shown that this is difficult to do at best. In the manual process, while generally providing a higher yield of breast meat, the quality of the product varies dramatically based on the skill of the worker, and the labor costs are significantly higher. It is the goal of this work to develop a system that eliminates labor and consistently provides a yield similar to the best manual worker. The starting point for the work is that the bird is placed onto a cone as shown in Fig. 14.8. The overall vision for this project requires the development of various technology components that will be unified into a single operational system. This includes a system to identify the initial cutting point, a system to specify the nominal cutting trajectory based on the size of that specific bird, a model to predict the location of the joint and shoulder tendons given the position/orientation of the wing tip, a mathematical model of the cutting process that allows the control system to interpret force/torque data and make intelligent motion commands to avoid cutting through the bone, and a robotic platform capable of executing these commands in real-time. The prototype cell is shown in Fig. 14.9.

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

341

Fig. 14.8 Front half of bird on a cone.

For our work, we have divided the cutting process into three distinct cutting steps. These steps are found in every manual cut, but they can vary between individual plants. The first step is a cut from the clavicle to the beginning of the shoulder joint (Zhou et al., 2007a). This cut is important to separate the breast meat from the carcass in order to maximize yield. The second cut is to cut all of the tendons and ligaments that attach the wing to the shoulder. This cut can be a source of bone chips if not performed correctly. Finally, the last cut is from the shoulder joint down to and along the scapula bone. This cut is important for yield and it is also a possible source of bone fragments. The team has identified the second cut, the cut through the shoulder joint, as the most technically challenging. We initially focused on the third cut and we were able to show the ability to cut along the scapula bone with a simple force-control algorithm. The success of the first cut is directly affected by the ability to accurately identify the initial cut position which is also very important to the second cut. However, the second cut is particularly challenging because of the requirement to cut meat and tendon, but not bone. In addition, the position of the wing plays a rather significant role in the success of the second cut, but it plays no significant role in the other cuts (Zhou et al., 2007b).

© Woodhead Publishing Limited, 2013

342 Robotics and automation in the food industry

Fig. 14.9 Intelligent cutting work cell.

The importance of the position of the wings for the second cut cannot be underestimated. The width of the gap in the shoulder joint is determined by the position of the wings as well as the tension in the tendons and ligaments. Much like cutting a string with a pair of scissors, the string must be in tension in order to be cut. For the shoulder cut, the tendons and ligaments can lose tension while the shoulder is being cut and the gap in the shoulder joint can completely close if the wing is not properly held. For our work, we are investigating both active and passive wing manipulation systems. The advantage of the active system is that the tension and gap in the shoulder joint can be controlled during the cutting process to insure a quality cut (Claffee, 2006). However, this does result in a more costly system due to the extra manipulation systems. The passive system is much simpler to implement but it does not allow the system to maintain the ideal tension in the tendons and ligaments during the cut. The trade-offs between these two concepts are still being investigated. To address the identification of the cutting point and predict the location of the bones and tendons in the shoulder joint, the GTRI conducted extensive modeling based on measurement of birds of all sizes (Daley and Grullon, 2010). This analysis allowed the team to develop an analytical model of the bird’s anatomy as a function of the three key points on the bird – one point for each wing where the wing meets the body of the bird and the tip of the keel, Fig. 14.10. These three points can

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

343

Fig. 14.10 Key exterior points on bird for cutting.

be identified using a stereo vision system to get the 3D location of the points or a mechanical system can be used to identify and measure the points as well. Based on the location of the three points, the model of the bird anatomy is used to identify the initial cut point for the robot and it generates a nominal cutting trajectory for that individual bird. This individualized cutting trajectory not only takes into account the size of each bird, but also its unique location on the cone. Using the nominal cutting trajectory, the system moves the cutting blade such that it is located above the cutting point. As the cone line moves the bird underneath the waiting blade, the cutting robot inserts the blade into the carcass. At this point, the system begins to monitor the cutting force to ensure that the system cuts meat and tendon, but not bone. As might be expected, an increase in cutting force normal to the blade would indicate that the blade is contacting a more rigid material – meat, tendon, or bone. However, there are many factors that affect the force being applied to the blade during the cutting process. These include: properties of the material being cut, velocity of the blade, angle of the blade relative to the

© Woodhead Publishing Limited, 2013

344 Robotics and automation in the food industry material (slice angle), and the sharpness of the blade (Zhou et al., 2006a, 2006b). Based on our modeling of cutting of biomaterials (Zhou and McMurray, 2007), GTRI has developed unique criteria for determining what the blade is cutting even though all of the above properties can vary significantly. This technique is able to identify when the blade has made contact with the bone such that the blade never penetrates more than 1 mm into the bone. An active research project at GTRI is focused on extending the bone detection algorithm such that it is insensitive to the blade sharpness. Blade sharpness has been a particularly difficult parameter to adjust for given the complex interactions of the above parameters and the difficulty in defining sharpness in practical terms. This work is focused on adapting the force cutting and bone detection algorithms to the ever changing blade sharpness, thus ensuring that the system can identify meat, tendons and bones from the very first cut with a sharp blade to the dullest blade used after several hours of cutting without resharpening (Zhou et al., 2006b). Of course, the goal is to cut through the shoulder joint without encountering the humerus or coracoid bones, but due to the very small gap in the shoulder joint this is inevitable in a practical system. Once the bone has been detected by the above technique, the system engages a force-control algorithm that attempts to maintain a constant force on the unknown shape of the bone. This is required because the tendons and ligaments that span the shoulder joint are attached to these bones at some point. By maintaining contact with the bone, the blade is able to move around the bone to cut all of the tendons and ligaments without generating a bone chip. The system that has been described is still an active research project at GTRI. There are still a number of technical challenges that lie ahead of the team before the system will be ready for commercialization. However, the core technology has been developed and the more basic research topics have been addressed. At the present time, the team is also investigating the more practical side of automating the cutting process. This includes defining the best way to hold/manipulate the wings of the bird during the cut, and the optimal angle of the bird to facilitate the cut. This second point is important in developing a robust cutting system. Given that this system might be cutting 300 000 shoulder joints a day, aligning the joint with the blade to minimize the side forces and provide a straight path to the major tendons only makes sense from a practical point of view.

14.3.4 Tray packing After the product has been deboned, the parts move down a conveyor where a final manual trimming operation is performed (if necessary). In many plants, the product is now ready to be placed into a tray pack where it will be sold to the customer. The placement of product into trays is another very labor intensive process in the modern poultry plant. This task typically can require up to ten people per line. In recent years, several commercial systems have been developed to address the tray packing problem. Robotic companies such as ABB and Fanuc have recently

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

345

come out with solutions to automating this task. The systems typically consist of several key components: a vision system to identify the position and orientation of the product on the conveyor, an end-effector to grasp the product on the conveyor, and a robotic arm to move the end-effector. One of the recent advances in the area of robotics that has been the basis of many of the commercial solutions has been the parallel robot. These robots have a much smaller work space and can handle much lighter payloads (typically under 1 kg) than the typical industrial arm, but they are significantly faster. The speed of these systems is what makes them well suited for this and many other food applications. These systems are gaining wide acceptance in the baking industry for these reasons. Typical applications include transferring cookies or croissants from one conveyor into the final package. Because of the high volume in which these products are produced, the system must be very fast and inexpensive, such that the price per part is very low. At this time, only the parallel robots can meet these criteria. However, for the poultry industry, there is another criterion that the robotic system must meet: food safety. The machine must be able to go through a high-pressure washdown every day and the system must be tested to show that it will not allow pathogens to grow on it. As was discussed in the introductory section, the ability to clean the modern processing plant and the equipment in the plant has been an obstacle to the introduction of robotic technology in the past. Today, however, solutions to this are beginning to appear on the market and the research labs. The current products being offered on the market today are significantly more rugged than traditional industrial robotic arms, but they still cannot undergo the standard high-pressure washdown procedures. This means that the machines must be hand washed. At GTRI, researchers have developed washdown technology to allow robotic devices to undergo a high-pressure washdown cycle that includes the caustic chemicals (Zhou et al., 2007c). The system is shown in Fig. 14.11. All of the materials used in the design of the robot are USDA/FDA approved, and all of the components have been tested for chemical compatibility with the cleaning agents used in a typical plant. The prototype system that is shown in Fig. 14.12 was placed in a meat processing plant for field testing. The system was only run off-line, but actual product was used in the testing. The system was run at a maximum production rate of 70 parts/min and a vision system was used to identify the product and its position/ orientation on the belt. If required, the throughput of the system can be improved through the addition of a second end-effector by extending the linear axis in the other direction. The results of the test showed that the design was sufficient to allow a robot to undergo a normal high-pressure washdown procedure without impacting the performance of the machine. In addition, surface swabs from the machine were also tested to determine the effectiveness of the design in insuring food safety. The testing revealed no signs of pathogens or any food safety issues. A key result of this project were a lessons learned for developing a washdown robot. These include:

© Woodhead Publishing Limited, 2013

346 Robotics and automation in the food industry

Fig. 14.11 Tray packing robot for washdown environments.

Fig. 14.12 Robot for washdown environment.

• Seal considerations – Redundant sealing is recommended for all critical sealing points where caustic cleaners have the potential of causing severe damage to the machine. FDA approved lip seals can go a long way to prevent ingress into critical areas. • Bearing selection – FDA approved bearings that do not require lubrication are preferred due to the lack of maintenance and durability; however, care must

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

•

•

•

•

•

347

be taken to select a robot configuration to limit loads and speeds (pressure– velocity ratings) to within the manufacturer’s ratings. Horizontal surfaces – Even if the base material is resistant to all food processing cleaners such as some stainless steel alloys, the lack of draining can still result in pools of liquid on the surface which are undesirable. Laser-cutting comments – Many of the machine elements were fabricated by using a laser-cutting system to cut stainless steel sheets. This method of fabrication must be handled with care because splatter from the laser-cutter supports can create contamination on surfaces of the parts, which, if not sanded or blasted with media, can result in severe rusting. Waterjet cutting removes the risks of contamination present in laser-cutting fabrication methods. Drain holes – Unless electronic equipment undergoes the costly procedure of hermetic sealing, drain holes are compulsory to prevent moisture from building up in equipment in washdown areas. Location of primary electronics equipment – Most machines can be operated by controllers located several meters from the actual manipulator. This can be used to the advantage of equipment providers if less severe areas are available for placing critical electronic equipment. Although pneumatic valves are one item that must be located close to the robot, low-cost commercial solutions exist to address this issue if they are properly guarded in a washdown capable enclosure. Underwater cables – Underwater-rated cables and connectors have integrated O-rings in the back shells to help to ensure signal integrity.

From a regulatory point of view, the following lessons learned resulted: • Designing for cleanability from the beginning – Placing cleaning considerations as early in the design process as possible prevents excessive redesign later in the design process where changes to the design are more costly. Referring to the AMI equipment design guidelines will also help to ensure equipment will meet the requirements of inspection staff at the plant where the equipment is to be installed (AMI, 2003). Keep in mind that inspection of equipment can differ from plant to plant meaning more robust hygienic designs will have a larger potential market. • FDA/USDA-approved materials do not guarantee acceptance – Although various materials and platings are sufficient for FDA/USDA approval, appearances and acceptance requirements which are plant specific may preclude the use of some materials such as plated aluminum surfaces. • ANSI/RIA standards – All safety standards must be followed to ensure the safety of personnel.

14.4 Robotics and automation in bulk packing and shipping of poultry meat Once the product has been shrink-wrapped or boxed in some form that protects the raw product from the environment, all handling tasks become very similar

© Woodhead Publishing Limited, 2013

348 Robotics and automation in the food industry 160.00

Index, 1990 = 100

140.00 120.00

Labor costs

100.00 80.00

Robot prices, not quality adjusted

60.00 40.00

Robot prices, quality adjusted

20.00 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 Data from International Federation of Robotics

Fig. 14.13

Estimated price index for industrial robots in US.

to those found in non-food facilities. Therefore, this is the area that has in recent years taken advantage of the commercial robotic systems. While there is still a temperature requirement for the product to inhibit the growth of pathogens in the product, it can, otherwise, be handled exactly like other industrial products. The modern industrial arm has become a very cost-effective alternative to manual labor (Fig. 14.13). In addition, the mean-time-to-failure for the modern robotic arm is approaching 100 000 h. This translates into a lower initial cost of a robotic system while dramatically decreasing the operational cost of the system through less down time and repair costs. For the modern poultry plant, one common final form for the product is the tray pack. This is a tray that has been individually shrink-wrapped and is meant to be sold individually to the consumer, typically at the grocery store. The task of loading the individual trays into a shipping carton has been a task that has been difficult for the industry to automate to date. The reason that this task has not been automated can be traced back to the requirements for this task. First, the typical plant is running between 30 and 60 trays per minute. As discussed earlier, this is a difficult speed for traditional robotic arms. Because of the diverse packing patterns used by the plants for their diverse product set, it is generally not possible to create a packing pattern and then load multiple trays simultaneously, as is common practice in many industries. Another difficulty in automating the tray packing operation is the vertical distance between the trays and the bottom of the boxes. While the vertical depth of many of the shipping boxes is less than 12 inches (30 cm), the boxes are usually formed with the flaps standing straight up. This means that the vertical distance between the box conveyor and the tray conveyor must be the height of the box plus the height of the flaps. This can easily result in a travel distance of over 24 inches (60 cm) for the robot to pick up the tray pack and place it in the bottom of the box without dropping it. This vertical stroke is greater than the vertical motion range of most parallel robots (the ABB Flexpicker, shown in Fig. 14.14, has a 12 inch (30 cm) vertical stroke limit).

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

349

Fig. 14.14 ABB Flexpicker.

To address this market need, GTRI worked with a commercial company, CAMotion, to develop a case packer to meet the unique requirements outlined above (Dickerson et al., 2005). The system developed consists of 3 degrees of freedom (XYZ motions) that can grasp and place a tray at the rate of 60 trays a minute. The specially designed end-effector was designed to account for the variations in the surface of the tray as well for frozen or non-frozen trays. A picture of the system is shown in Fig. 14.15. The next step in the process is typically the pallet loading process. This task is generally fairly easy to cost justify for a poultry company. Pallet loading takes place after the product has been placed into a shipping box and the boxes are loading onto a pallet for bulk shipping. What makes this task easy to automate is that the boxes are fairly rigid and well-defined products at this point and, as such, easy to manipulate with an industrial robot. The speed of the task is not as demanding as those found in other parts of the plant and the packing patterns on the pallet are very well defined. In many of the newer plants around the world, this task is almost always automated. The next task that can be automated is the wrapping of the pallets to keep the boxes on the pallet from shifting during transport. There are a variety of semi-automated solutions as well as fully automated systems on the market today. The semi-automated solution typically requires the user to either set or build the pallet on a rotating platform and then input the size of the pallet into the controller.

© Woodhead Publishing Limited, 2013

350 Robotics and automation in the food industry

Fig. 14.15 Case loading robot from CAMotion.

After this is complete, the user manually starts the shrink wrap on the pallet and then turns on the system. The shrink wrap material is automatically dispensed as the pallet rotates on the table. The system automatically moves in the vertical direction to provide complete coverage of the pallet. Internal sensors are used to maintain the proper tension in the shrink wrap material during the entire process. The automated solutions are more convenient to use. One new system on the market is actually a small automated mobile robot from Italdibipack Group and the product is called the Robot Leonardo2000. The user can move the robot to the pallet anywhere in the plant and simply turn on the automated shrink wrap process. The robot has a sensor on the front that maintains contact with the pallet while it drives around the pallet applying the shrink wrap material while maintaining the proper tension in the material. The operator still has to input the height of the pallet, but the rest of the process is automated. Once the pallet has been created, the pallet can be handled just like any other product in the supply chain. If desired, automated guided vehicles (AGV) can be used to transport the pallets to and from the freezer or refrigerated areas. These systems are mobile flatbeds or forklifts that can follow predefined paths to move product from one location to another inside of a company. Companies such as JBT Corporation and Egemin Automation can provide full service products to meet the demands of the poultry processing companies. In addition, automated storage retrieval systems can be used to handle the warehousing needs. Two companies in this area are Consafe Logistics and Dematic, but there are many other companies in this space. These systems automatically transfer product from the AGV to

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

351

a mechanized tote to be stored in a warehouse and retrieved from the warehouse: all without requiring any human involvement.

14.5 Future trends As the poultry moves into the next decade, there is a sense that the industry is ready for a change in processing technology that has brought them so much, but now seems so fragile. As food safety requirements are increased by government agencies and the cost of feed and oil increase as well, this presents an opportunity to rethink the process and optimize the process for a new set of variables: animal welfare, sustainability, food safety, yield, and labor. While it is impossible to look at our crystal ball and accurately predict what the process will be, it is possible to identify technologies that will play a major role in reshaping the processing landscape in the future as the industry tries to meet these new set of goals. First, I see the plant of the future to be a data-driven process. Data and analytics will be used to better estimate the size of the birds at the grow-out house to optimize their delivery to the processing facility so that their weights match the current orders of the plant. Live hang will eventually be replaced by a more acceptable slaughter process that will require a new method of loading the birds onto the shackle. Given that the birds will be killed or stunned in bulk, the loading problem will manageable since they will not be moving. Once the birds enter the plant, a 3D model will be constructed and the weight and yield of each bird will be automatically calculated. From this data, the exact settings of the entire process will be determined and communicated throughout the process. This will allow something that the automotive industry and other manufacturing industries have been searching for decades for: processing down to lot sizes of one. Each process will be optimized for that bird so that nothing is wasted and the most value of every bird is obtained by the processor. It will also minimize the use of precious natural resources such as water and energy in the process. As the bird moves through the process, data will be collected during the process to verify food safety (temperature of the product, digital image of product as well as a time stamp). Product traceability from the grow-out house to the tray pack leaving the plant will be guaranteed as well. This means that the water chiller in the US will have to change. Will it go away and be replaced by the air chiller or will cryogenic freezing of the bird replace the water chiller? I do not know, but the drive to reduce water and food safety tracking will eventually cause it to go away. Throughout the entire process, no hands will touch the product unless there is an exception in the process. I dare not predict the order in which the changes in the process will occur in the future as there are many possibilities that have yet to be explored to answer that question. However, of one thing I can be sure: the creativity of the people working in this industry will be released like it has never been

© Woodhead Publishing Limited, 2013

352 Robotics and automation in the food industry before as academia and the industry begin to team up together with a clean sheet of paper and ask the question: ‘what would the process look like if...?.’

14.6 References 2002. Requirements for the Design of Meat and Poultry Processing Equipment. NSF/ ANSI/3-A 14159-1 – 2002. AMI 2003. Fact Sheet: Sanitary Equipment Design. American Meat Institute. AMI 2010. Available from: http://www.toolingu.com/definition-300100-12475-fixed-au tomation.html [Accessed]. Claffee, M. R. 2006. The effects of wing manipulation on automated cutting of biological materials. M.S., Georgia Institute of Technology. Daley, W.D., Grullon, S. and stewart, J. M. 2010. Patent Application Number 12/608,939. Dickerson, S., S. Coleman, G. V. McMurray, W. Holcombe and J. Prince 2005. The commercialization of a low-cost, high-speed pick-and-place casepacker. XVII European Symposium on the Quality of Poultry Meat. Golden Tulip Parkhotel Doorwerth, Doorwerth, The Netherlands. Lee, K.-M. 2001. Design criteria for developing an automated live-bird transfer system. IEEE Transactions on Robotics and Automation, 17(4), 483–490. Lee, K.-M., B. Webster, J. Joni, X. Yin, R. Carey, M. Lacy and R. Gogate 1999. On the development of a compliant grasping mechanism for on-line handling of live objects. Part II: Design and experimental investigation. ASME International Conference on Advanced Intelligent Mechatronics (AIM’99). Atlanta, GA: IEE. Lee, K.-M., J. Joni and X. Yin. 2000. Imaging and motion prediction for an automated live-bird transfer process. In: ASME Dynamic Systems and Control Division-2000, 5–10 November, Orlando, FL, IEE, 181–188. Lee, K.-M., J. Joni and X. Yin. 2001. Compliant grasping force modeling for handling of live objects. In: May 21–26 2001 IEEE International Conference Robotics and Automation, Seoul, Korea. Li, Q. and K.-M. Lee 2005. Effects of color characterization on computational efficiency of feature detection with live-objective handling applications. In: 2005 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM 2005), 24–28 July, Monterey, CA, 225–230. Shumway, C. 2002. Dynamic modeling and analysis of body inversions. M.S., Georgia Institute of Technology. Socha, K.G., G. V. McMurray, W. D. Holcombe and H. Lipkin. 2004. Design of a compliant end effector for grasping non-rigid materials. In: International Conference on Intelligent Manipulation and Grasping, 2004 Genoa, Italy. 253–258. USDA 2001. USDA Guidelines for the Sanitary Design and Fabrication of Dairy Processing Equipment. Webster, A. B. and K. M. Lee 2002. Toward automation of the transfer of broilers to the processing line. WATT Poultry USA, September, 28–42. Wyvill, J. C. 2005. Dynamic Process Control in Poultry Processing and the Technology Tools Helping to Make It Possible. Athens, GA: University of Georgia. Zhou, D., M. R. Claffee, K.-M. Lee and G. V. McMurray. 2006a. Cutting, ‘by pressing and slicing,’ applied to robotic cutting bio-materials. Part 1. Modeling of stress distribution. In: 2006 IEEE International Conference on Robotics and Automation (ICRA 2006), May 2006, Orlando, FL, 2901. Zhou, D., M. R. Claffee, K.-M. Lee and G. V. McMurray. 2006b. Cutting, ‘by pressing and slicing,’ applied to the robotic cut of bio-materials. Part 2. Force during slicing

© Woodhead Publishing Limited, 2013

Robotics and automation in the poultry industry

353

and pressing cuts. In: 2006 IEEE International Conference on Robotics and Automation (ICRA 2006), 15–19 May 2006, Orlando, FL, 2256–2261. Zhou, D., J. Holmes, W. Holcombe and G. McMurray. 2007a. Automation of the bird shoulder joint deboning. In: 2007 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, 4–7 September 2007a, Zurich, Switzerland. Zhou, D., J. Holmes, W. Holcombe, K.-M. Lee and G. McMurray. 2007b. Automation of bird front half deboning procedure: design and analysis. In: FToMM 2007 World Congress in Mechanism and Machine Science, 17–21 June 2007b, Besancon, France. Zhou, D., J. Holmes, W. Holcombe, S. Thomas and G. McMurray. 2007c. Design of a fresh meat packing robot for working in washdown environment. In: 2007 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, 4–7 September 2007c, Zurich, Switzerland. Zhou, D. and G. McMurray 2007. Formulation of the relationship between tension loading rate and internal stress for uni-axial bio-materials. In: IASTED 2007 Modeling and Simulation Conference, 30 May–1 June 2007, Montreal, Canada.

© Woodhead Publishing Limited, 2013

15 Robotics and automation in seafood processing J. O. Buljo and T. B. Gjerstad, SINTEF Raufoss Manufacturing AS, Norway

DOI: 10.1533/9780857095763.2.354 Abstract: Automation and the use of robots are enabling technologies in the seafood industry when the goal is reduced production costs and increased product quality. The seafood processing industry has a relatively small robotic involvement compared to some other industry sectors, and needs to increase this involvement in spite of the challenges regarding robotic handling of fresh food products and hygienic requirements. Some technological possibilities are briefly described in this chapter; among other technologies, some new gripper solutions are mentioned. The application and adoption of automation and robotics in different unit operations in fish processing are described and some future trends are briefly discussed. Key words: robotics, automation, seafood processing, fish, gripping.

15.1 Introduction As an introduction, we will provide some observations regarding driving forces for increased automation of fish processing.

15.1.1 Key drivers for enhanced automation Seafood means food of animals from the sea, especially fish or sea animals with shells. In this chapter the focus will be on some aspects of automation primarily in processing of fish products. For marketing and sales of seafood products, the main issues are consumers’ demands and preferences. Many food products nowadays are ready-made meal solutions instead of traditional meal ingredients, prepared as fresh/chilled, organic and functional foods. To meet new market trends, innovation is often required regarding products, food processing, packaging, distribution

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

355

channels and eating-out-of-home opportunities. Price and quality are two essential factors to be competitive in the market. Automation and the use of robots are enabling technologies when the goal is reduced production cost and increased product quality. In most cases automation in the food industry enables improved production efficiency and reduced production costs, improves working conditions in food processing factories, enhances hygiene standards, improves yield margins, increases profitability, and eases conformation to legislation pertaining to food processing. Automation and robotics are therefore seen by many enterprises, including those in the seafood sector, as a necessity to secure their future survival. Seafood consumption is increasing worldwide. The fishing industry harvested about 141 million tons of seafood globally in 2005. That was eight times as much as in 1950, with each person on average eating four times as much seafood as before. We cite from the document ‘State of the world. Innovation for a Sustainable Economy’ (Halveil and Nierenberg, 2008): For people living in wealthy nations, seafood is an increasingly popular health food option; with its high levels of fatty acids and trace minerals, nutritionists recognize seafood as essential to the development and maintenance of good neurological function, not to mention a reduced risk of cancer, heart disease, and other debilitating conditions. In poorer nations in Asia, Africa, and Latin America, people are also eating more fish if they can afford it.

Modern people want to eat seafood products that are easy to prepare, of good quality and not too expensive. This justifies enhanced automation use within the seafood industry.

15.1.2

Current status and challenges of adopting robotics and automation in seafood processing Robots are often seen in most of the manufacturing industries, such as e.g. automotive and machinery industries. According to International Federation of Robotics (IFR) more than a million robots had been installed worldwide by 2008 (stock of robots) and more than one hundred thousand units were purchased in 2009. The development is shown in Fig. 15.1. The food industry has a relatively small robotics and automation involvement compared to the automotive industry and there have been great barriers to the purchase of industrial robots. The situation is the same in the seafood industry sector. However, the number of new robot installations in the food industry has increased each year. In 2003 the total number of robot sales was 1003, and in 2008 the number had increased to 3961. Despite the economic crisis in 2008, when robot shipments decreased in most of the manufacturing industries, the number of robot shipments in food and drink industries followed an upward trend. About two-thirds of the worldwide sales to this industry were made in Europe. According to Robotstatistikk (Litzenberger, 2009), these investments were made

© Woodhead Publishing Limited, 2013

356 Robotics and automation in the food industry 140000 No. of robot shipments

Total number Food and drink 120000 100000 80000 60000 40000 20000 0 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 Year

Fig. 15.1

Shipments of robots during the year 1999–2008.

to optimize production processes. The robots purchased were predominantly for material handling tasks, such as secondary (case packaging) and tertiary (palletizing) packaging. In the fish processing industry, machines are currently available for most of the unit operations. The machine and equipment supplier industry is in continual contact with the fish processing companies, and wishes to improve its machine solutions in order to increase its competitive strength. As a result of this, the machines are getting ever better (quicker, lower energy consumption, higher production output, longer uptime, better product quality, etc.). Nevertheless, we still find that several machines must be fed manually or they may require regular assistance from an operator. To a small extent, the machine suppliers have developed systems which automatically feed the machines. Some automated feeding systems exist, but they are not an everyday sight in the fish processing industry. There can be several reasons why the fish processing industry does not act more quickly in regard to the utilization of new automated technical solutions. One reason is the low profit margin and the small number of employees with high technical skills in fish processing enterprises. The low involvement of robotics in food processing is also due to the fact that food products, and seafood products particularly, are highly variable both in shape, size and structure, which poses a major problem for the development of sensor systems and manipulators for handling such products (Litzenberger, 2009). In particular, there has been a lack of suitable grippers that can handle non-rigid objects such as fish fillets or whole fishes. In recent years, some gripping solutions showing promise for such use have been developed, but the fish processing industry has only to a limited degree been interested, and willing, to put such solutions into use. New gripping solutions will be further described later in this chapter. In addition, though there is a huge assortment of industrial robots in the market today, not all of them are well suited for the fish processing industry. Requirements regarding hygienic design of machines and equipment, and the

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

357

fact that few materials are suitable to withstand the environment where seafood is processed, lead to the development of good solutions in automated processing being costly. In the fish processing industry, different production areas have different hygiene-graded levels. In high-risk areas, strict hygiene requirements are set for the equipment, facilities and operators. Robots and gripping tools, like other machinery, should be designed using hygienic design principles, being easily washable using both alkaline and acidic detergents. One solution is that the robot could possibly be covered in a plastic shroud to simplify the cleaning (Kuka, 2011) and this is used to some extent. The humidity in the production area is a greater challenge for robots than the temperature, and is a complementary factor for the reduced use of robots in high-risk areas. Traditional industrial robots are not adapted to the hygiene requirements of clean rooms or humid working environments, and only specialized clean-room robots can be used. These robots are more costly than traditional industrial robots due to their special design, the material used, and a smaller market. In fact, a robot satisfying the clean-room standard is approximately 10–20% more expensive than standard industry robots. The development of good solutions in automated processing is costly for the reasons mentioned above, and the market for a special processing system is in many cases small. The economic risk the system supplier has to meet often seems to be high and many good ideas are therefore not realized. In spite of the challenges mentioned above, we can recognize that the degree of automation in seafood processing is increasing (Litzenberger, 2009). Present-day fish processing is mainly based on the principles of continuous manufacturing where the fish product moves down a manufacturing line to be processed at specific single operations, both mechanical and manual. Due to a reduction in the cost of elements required for automation, it is now technically feasible to automate tasks in food processing and handling to an increasing extent. We suggest that technology for integration of machine vision and robotics has reached the level of maturity that is required in order to solve grading, inspection and processing tasks, especially in the pelagicfish processing industry. Due to the similarity in size and shape, processing of pelagic fish is often easier than processing fish in general. Also the automation tasks are easier with pelagic fish for the same reasons. However, when following a modern line of seafood processing, we find both automated machines (operations) and manually operated stations to possess the flexibility needed to be able to respond to an ever-changing market situation. The seafood industry is today highly dependent on manual labour. Increased automation is needed and will be useful, but faces challenges with regard to the seasonality and variability of the raw material and, in the case of at-sea mechanization, the need for motion compensation.

15.1.3

Benefits of replacing manual handling with automated handling in seafood processing The most pronounced benefit of implementing gripper solutions in the fish and meat processing industry is the replacement of operators. There are several

© Woodhead Publishing Limited, 2013

358 Robotics and automation in the food industry arguments that support the use of automated handling in preference to manual handling. In order to make the use of end-effectors beneficial, we presuppose that the design is carried out according to object requirement and hygienic design principles. In an automated handling operation, the object is grasped only once, while in a manual operation objects are handled several times before being placed in a fixed position (e.g. between conveyor belts and cases, and cases and packaging machines). Automated handling will therefore reduce the physical stress on the object and also to a greater extent maintain the object quality. An end-effector designed by hygienic design principles brings no or minimal negative influence to the object’s condition, neither physically nor bacteriological. In that case the end-effector is more hygienic than operator handling, avoiding bacteriological contamination from the operator. To be able to automate the operation, the product flow needs to fit the cycle time of the automated operation. An uneven product flow lowers the performance of the automated solution. Operators, with their high degree of flexibility and dexterity, are able to hide this bad product flow, for example, using unplanned buffers, resulting in a lower performance of the automated solution than should have been possible. There are also other benefits that are often referred to as potential benefits from robotic systems: • Reduced requirements for floor space, for example, use of robots installed in the ceiling and less need of intermediate bearings like cases and boxes. • Improved efficiency. • Maintained and improved quality of handling operations. • The ability to work in cold or hostile environments. • Increased yield and reduced wastage. • Increased consistency – the handling operation is carried out in the same way for each cycle. • Increased flexibility for some operations – the end-effector may be able to handle a wide range of objects. The primary packaging of food is difficult because the speed is higher and the product is probably at its most vulnerable with regard to quality and safety (Wallin, 1997). The reason for the complexity is that the personnel who are currently used to pick and place products are able to carry out several additional tasks: • • • •

inspect for colour, shape, texture, size, type, etc., stop the line if there are downstream problems, adapt to new products, make decisions based on previous events.

As shown in Fig. 15.2, the throughput, and the ability to satisfy hygiene requirements and meet legislation are decreased with manual handling and increased by robotic handling. On the other hand, the number of labour intensive processes, final cost of products, and the environmental effect of manual handling, are

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

359

Hygienic requirements Throughput Decrease

Increase

Labour intensive process

Robotic handling

Manual handling

Meeting legislation

Final cost of products Environmental effects on Increase

Fig. 15.2

labour

Decrease

Requirements for robotic handling versus manual handling of food products. (Source: Erzincanli and Sharp, 1997a.)

decreased with robotic handling. This supports the arguments of implementing robotic solutions in fish processing industries when appropriate.

15.2 Technologies for robotics and automation in the seafood industry Applicable technology is a requirement of increased automation. Some important areas of technology are discussed briefly in the following text.

15.2.1 Gripping tools Robotic systems require end-effectors to carry out the required handling tasks. An end-effector for handling products is an interface between the robot or programmable arm and the material or product to be handled. The success or failure of an application depends on how well the end-effectors are designed, developed and implemented (Erzincanli and Sharp, 1997a, 1997b). The handling of fish objects using gripping tools is a more challenging task than selecting an appropriate robot. Most grippers are developed to handle rigid, three-dimensional objects, but there are also end-effectors developed to handle different categories of objects, from rigid to non-rigid objects. Compared to objects in mechanical industries, where the quality tolerances/limits are within millimetres, fish vary in shape, size and texture. There are great variations in object size and geometry within the same species of fish and within given weight classes. In addition, the raw material quality changes between seasons: the texture is often softer in the spring than the summer/autumn. A problem that occurs when handling non-rigid materials is that such objects change their geometry under the influence of force. This non-rigidity makes automatic handling of these materials very difficult. Conventional robot grippers

© Woodhead Publishing Limited, 2013

360 Robotics and automation in the food industry (such as two-jaw grippers used for handling rigid material) cannot be used because these methods can result in insufficient grasping (Zoller et al., 1999, pp. 3–29) whereby the objects are loosened or damaged. These product challenges are described in more detail later in this chapter. There are different requirements for equipment used in a hygienic high-risk area than in a non-high-risk area. Robots handling wrapped products are normally not working in this high-risk area. These objects have the same shape and often a flat surface suitable for grippers based on principles developed for traditional mechanical and automotive industries. Different handling operations, in combination with different object shapes and product conditions, set different requirement specifications for the gripping tools even though the raw material is from only one species, such as cod or salmon. Different gripper tool principles are used depending on the object characteristics and the handling task. More gripper tools are needed to increase the use of robots in fish processing industries. Standardized industrial robots and gripping tools are often used in automatic palletizing operations of food products packed in cardboard boxes. Defining object characteristics and behaviour Before the gripper is selected and objects are gripped, the characteristics that define the object type and the way the gripper is supposed to pick it up must be described. Thereafter, how the object is presented to the gripper must be established. An object’s geometry defines the options for applying force to it. Physical object characteristics determine the way forces are applied. The latter is of particular interest for handling non-rigid objects where the surface cannot be damaged (Wolf et al., 2005). Object shape and size are the most important geometric characteristics when selecting/developing an appropriate end-effector for handling non-rigid food objects such as fish. The variations between the species, and also between objects from the same species (e.g. portions of fish), are great. The size of the fish in different weight classes may vary a great deal, where some fish are thin and long and others might be short and muscular. These differences result in great variation of shape and size. In comparison with metal manufacturing industries, the limits of tolerance of metal objects are far stricter. The most important physical characteristic is texture. Texture comprises groups of physical properties that derive from the structure of the food. Functions of mass, distance and time are objective measurements used in analysing these properties. There is a change in texture for most of the food during storage. In most of the food industries, the processing is directed to weakening of the structure in order to make the objects easier to masticate. Fish is classified as non-rigid raw material, which means that the geometry changes under influence of forces. In the fish processing industry, traditional salmon processing is carried out after a minimum of three days, when the rigor mortis phase is passed. Wild-caught fish such as cod and pollock are also ripened when processed, since the fish in most cases is not landed the same day they are caught. During the ripening period, the fish meat becomes softer.

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

361

Physical characteristics determine the way forces are applied to the objects. Fish fillets and portions of fish have a wet, sticky and uneven surface. Changes to these characteristics, together with variation in texture properties from object to object, require careful calculation of forces applied to the objects from both the gripper and the handling procedure. Neither the robot nor most of the grippers have weight constraints that preclude the handling of fish objects due to weight restrictions. However, the object weight is important for the grasping operation where no physical closure may hold the object and where the objects are under influence of acceleration and retardation forces. The force that might be added to the object through acceleration and retardation depends on the object weight. Object behaviour is defined when the object is at rest and when it is in motion. For rigid objects, the behaviour becomes most interesting before the pick operation starts, while changes during motion are also important for non-rigid objects such as fish. An object which may change position must be kept steady to prevent it from changing position while it is about to be picked up. The non-rigid fish are not in this relation technically defined, and the object characteristics fluctuating within specific tolerances lead to immediate challenges (Wolf et al., 2005). Gripper developments The gripping tools mostly found in the food processing industry handling raw materials and non-wrapped products are mechanical grippers, vacuum grippers, and non-contact grippers. Mechanical grippers are often based on the positive closure principle. Different designs have been reported. A jaw gripper was developed to handle boneless portions of poultry into correct position in packaging trays (Brett et al., 1991; Karakerezis et al., 1994; Chua et al., 2003; Harwes, 2004; Seliger et al., 2000). This gripper was further developed to handle dough (Stone et al., 1996). Similar grippers have been commercially developed by AEW Delford (Fig. 15.3) (Harwes and Nierenberg, 2008), where the objects are grasped by jaws. Several versions are adapted to handle different objects dependent on size, shape and non-rigidity. The object size and gripping type define how far the operating elements of the gripper need to be opened. The distance which the gripper fingers/jaws cover to apply force on the object, the so-called action radius of the fingers, also influences gripper flexibility. In this case, some space is required between the objects to be able to open the gripper before grasping and to avoid damaging the objects nearby. Vacuum grippers have been reported that were developed for handling chocolate bananas. These objects have a relatively flat surface, which is a prerequisite for the use of vacuum gripper. A big advantage is that the objects are grasped from the top surface and only one flat surface is needed. On the other hand, improper positioning causes reduced holding power, thereby loosening the grip. However, meat juice, air leakage and variation in object size, shape, surface condition, and texture, make maintaining the vacuum and satisfying hygiene requirements when handling fish difficult (Wögerer et al., 2005). A non-contact gripper is reported by Erzincanli and Sharp (1997a), Rawal et al. (2008) and Ozcelik et al. (2003) that takes advantage of the radial outflow

© Woodhead Publishing Limited, 2013

362 Robotics and automation in the food industry (a)

(b)

Fig. 15.3 Pick and place gripper device developed by AEW Delford, Publication number US2010133862 (A1) (Patent pending). (Source: Harwes and Nierenberg, 2008.)

phenomenon when generating a high-speed fluid flow between the nozzle(s) and product surface. Thereby a vacuum is created which levitates the object. This gripper is said to be able to lift objects having compliant structures and viscous or slippery surfaces. It also has a simple, rugged, compact design and no moving parts. The drawbacks of this gripper are that this end-effector is not able to handle objects that have a relatively rough surface or a surface covered with loose materials, or have an irregular surface structure (because the non-contact end-effector requires a flat surface to handle a product), or are air permeable (the air will pass through the object instead of flowing radially outwards and no vacuum is created) (Erzincanli and Sharp, 1997b; Davis et al., 2008; Rawal et al., 2008). At SINTEF four different grippers have been developed based on the positive closure principle (surface hooking) and frictional engagement. Based on the requirements of surface hooking, a needle gripper (crossing needles), a compact needle gripper and a compact freezing gripper have been developed (Figs. 15.4–15.6). The

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

363

(a)

(b)

Fig. 15.4

Compact needle gripper, prototype design (developed by SINTEF).

Fig. 15.5 Needle gripper with crossing needles, prototype design (developed by SINTEF).

needle gripper penetrates the objects. However, the needle gripper seems to satisfy most of the requirements set by industry and researchers, giving the best mechanical functionality. Use of needles is common in the meat processing industry especially, and also more and more common in the fish processing industry, for example, used to inject brine and marinades into the products. These are hollow needles and may have less strength than needles used in end-effectors/gripping tools. Tests carried out where this compact needle gripper were handling pieces of salmon, cod or pollock showed promising results. With an acceleration of 6.0–8.0 m/s2, more than 80% of the pieces were grasped and handled without harming the objects. The needle gripper with crossing (Gjerstad et al., 2006) needles gives an even better hold between the gripper and the non-rigid object.

© Woodhead Publishing Limited, 2013

364 Robotics and automation in the food industry

Fig. 15.6 Three compact freeze gripper units (developed by SINTEF) fasten to a bracket for fixing to robot arm.

Freeze gripper development has several unsolved problems of a technical and operational character (e.g. finding the right temperature for optimal grasp for different species, mapping the heat transmission between object, freezing surface and cooling agent), and the timeframe until an industrialized version of this gripper is available seems to be much longer than for the commissioning of the needle gripper. However, this concept is of special interest regarding achieving the best hygienic design in combination with use of the freezing principle (creating a thin layer of ice in the contact area between gripper and object). It is the principle that demonstrates the gentlest handling of the objects with virtually no influence on object condition. The fourth gripper, a mechanical finger gripper, was developed to handle leg of ham, but is also suitable for handling whole fish. Two rows of movable fingers placed on each side of the object gave a soft grip with no or small changes to the object surface. More development is needed to adjust the fingers and to reduce the total weight of the gripper so as to reduce the cycle time of the handling procedure.

15.2.2 Control systems Modern automated processing plants require suitable control systems. During the last decades, there has been a considerable development of control systems, and these have become relatively cheaper and acquired a far higher capacity for data processing. New low-cost computer technology is one of several reasons for this

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

365

development. Introduction of new control systems must include some adaptation to the particular sector. Such adaptation has also happened in the fish processing industry. Control systems must cover different levels of control at the processing plant. Often the machines on the shop floor have their own control, which does not necessarily communicate with the controls of other machines, chilling equipment, etc. ERP (Enterprise Resource Planning) systems integrate internal and external management information across an entire organization embracing finance, supply chain, manufacturing, human resources, sales, service, customer relationship management, etc. Another aspect is systems for planning and execution of maintenance. There is often a need for these different systems to exchange data and put to use the existing information in the company. Access to reliable real-time information gives new possibilities to take good decisions that answer the visions and goals of the company. Even though there has been a gradual development in this direction of coordinated and integrated control systems, there is still a long way to go before optimal solutions will come into use in this part of industry. Production control systems should be based on the principles of flexible and reconfigurable systems (open-architecture control). If the food factory of the future is close to being a fully automated processing plant, a considerable effort regarding further development of the control systems is still required.

15.2.3 Machine vision and other sensor-related technology Machine vision is a sector in engineering and is related to computer science, optics, mechanical engineering and industrial automation. Vision as a sensor system is based on the use of camera technology. A vision system is a powerful tool for the mapping of the production plant’s surroundings and details, because of the large quantity of information which can be read out of a picture in a relatively short period of time. A vision system includes software for picture analysis, where key information is sorted from the pictures and analysed. There is also a camera in the system, a signalling device initiating the photographing, and often also a lighting arrangement. After the picture analysis, the resulting information is sent to the control system to act as a basis for actions related to the processing. Vision systems are gaining ever wider ground, also in the food processing industry. Machine vision has been used for sorting fish according to species, grading herring roe, analysing whole fish and fillets with respect to freshness, estimating brown trout cutlet fat contents by automated colour image analysis, determining the fat and connective tissue amounts in salmon fillets, and sorting of whole Atlantic salmon into ‘superior/ordinary’ and ‘ordinary’ quality grades (Misimi, 2007). The development of machine vision algorithms and related technology in recent years has also resulted in the wider utilization of these technologies in food processing applications. Automation presupposes the use of controlling systems, which again presupposes sensors that can give signals from the processes to the control systems to monitor and control. Compared to many other branches of industry, the use of

© Woodhead Publishing Limited, 2013

366 Robotics and automation in the food industry sensors in the fish processing industry is relatively moderate. This is of course related to the fish processing industry not being at the front of the technological development, and having a lower degree of automation than for instance the automotive industry. New and more functional sensors are continually being developed in all categories. The main categories of sensors are internal sensors and external sensors. External sensors are classified as non-contact and contact sensors. Vision systems are part of the non-contact group. The sensitivity, response time, work area, precision, solution, operating safety and price–effect relation of the sensors are continually improving. There is a powerful trend in the direction of miniaturizing. All these improvements can also be utilized in the fish processing industry in the future. Nevertheless, in many situations the sensors will have to be adapted to the specific conditions related to this part of industry (temperature, moisture, chemical cleaning agents, hygiene requirements, etc.). The positive development of better sensors for different purposes allows for development of more automated fish processing plants than is typical in the industry today.

15.3 Application of robotics and automation in fish slaughtering, filleting, portioning and associated unit operations Fish processing comprise a number of unit operations. The state regarding automation has been outlined for some important basic unit operations.

15.3.1 Slaughtering operations Wild-caught fish of a certain size (not typical pelagic fish, such as herring and mackerel) normally have gill arches cut shortly after having been hauled aboard the boat. This is especially important if the method of fishing is such that the fish are still alive when brought aboard, as the gill-arch cut quickly kills the fish. Another common method of killing the fish is gutting and de-heading immediately when hauled aboard. Chilling or freezing of wild-caught fish is very important to secure fish quality. Pelagic fish, such as herring and mackerel, are normally pumped or hauled into the hold, and are not individually killed. These die due to the lack of oxygen in the boat’s hold. It is very important to cool down and freeze pelagic fish very quickly as they normally have not been gutted. The intestines contain nutrients, and these must not be allowed to spread in the haul, as they provide fertile conditions for rapid bacterial growth, which will impede the quality of the caught fish. This process must be conducted with care. Farmed fish, such as salmon, trout, cod, etc., will, in compliance with the ethical rules in many countries, be anaesthetized before slaughtering. Previously this

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

Fig. 15.7

367

Robotized line for gill arch cut delivered by SeaSide.

anaesthetising had been conducted through the addition of carbon dioxide to the water tank (chilled water) holding the fish. This method of anaesthesia (i.e. sedation) is now being phased out, and is being replaced by stunning, either by physical impact or electrical discharge. These methods are to a large extent automated, the fish being fed into the machine where the process is conducted automatically, before moving on to the station where the gill-arch cut is performed, if these processes do not occur at the same station. Installations for automated gill-arch cutting have been developed. For instance, SINTEF Fisheries and Aquaculture has in collaboration with the equipment supplier Seaside AS developed an installation for automated gill-arch cutting on salmon, trout and similar fish (Fig. 15.7). This type of installation is intended to be adaptable to different species of fish. Blood drainage usually takes place in water tanks with chilled water, or by the fish being placed head down individually in small boxes on an overhead conveyor. For this last method the fish is manually fed onto the conveyor, but this operation can be automated. Farmed fish, such as salmon and trout, are normally transported from net cage to slaughter house by fish carriers (vessels) by sea. From a net cage at the slaughter house, the fish are pumped into treatment stations for anaesthetising and slaughter. These operations are to a large extent mechanized or automated. The process of handling the net cage requires manual effort. Alternatively, the farmed fish are slaughtered in the immediate range of the farming facilities and then transported to a plant for further processing.

15.3.2 Sorting operations At a fish processing plant, there is normally a need to perform some sorting based on one or several of the following factors: fish type, size, shape, weight, external and internal characteristics, freshness and other quality-related factors. Sorting can happen at different stages of the processing, according to need. Presently,

© Woodhead Publishing Limited, 2013

368 Robotics and automation in the food industry simple mechanical appliances (V-belts, rotating rollers, etc.) and weight systems, combined with visual methods performed by operators, are used to a large extent. At plants processing pelagic fish (herring, mackerel, etc.), there are often large volumes that are to be processed (90 000 to about 300 000 individual fish per hour is normal). It is not possible to perform a quality-related inspection of such quantities of fish with adequate precision with only a handful of operators. In addition, manual methods have their obvious limitations and drawbacks. Therefore, there is a need for automated methods for sorting fish, be it pelagic fish or other species. In recent years, new methods applicable in the fish processing industry, based on machine vision (see Section 15.2.3 on machine vision) have been developed. These new methods will most probably come into use in the fish processing industry, even though they are presently applied only to a limited extent. Also, new sensors, for instance electronic noses (Natale et al., 2001; Haugen et al., 2007), can determine the freshness of the fish. A combination of different methods (mechanical, machine vision, advanced sensors) will be able to cover the need for quick and efficient sorting of fish raw material. This is an area where further development of methods and industrial solutions are in prospect. Electronic capture of data, in combination with sorting operations, will give useful data regarding optimal usage of the fish being processed. This is possible, and to some extent applied, by the help of signals from sensors integrated in the process equipment.

15.3.3

Preparation of fish for filleting and for some other alternative processing methods After the fish have been slaughtered, they are readied for filleting. In order to maintain the quality of the raw material, the fish are kept chilled. Different methods are used for transport and storage in this regard. Conveyors are often used for transport operations. For short-time storage, plastic boxes lined with ice are much used and also larger cases lined with ice slurry. Some processing lines also include removal of scales, with suitable machines for this operation. The Icelandic supplier BASIS International delivers automatic high-speed de-scalers for different sorts of fish. Before filleting, the intestines of the fish must be removed from the belly. This is called gutting. Machines performing the gutting process have been available in the market for quite some time. Examples of gutting machines are Kronborg 500/510 (designed for cod, saithe, haddock, etc.; 28–75 cm fish length; 33–52 fish/ min) and Baader 142 (salmon, sea trout, coho; 1.3–3.5 kg, 2.0–7.0 kg, 5.0–10.0 kg; 16 fish/min), both delivered by Baader Food Processing Machinery (Germany). Peruza (Latvia) delivers de-heading and gutting lines for small fish. There are also other firms supplying solutions for these operations. The machines are mostly fed manually, even though there are some machines available with automated feed. The intestines of the fish are transported from the gutting machine, and are treated as a resource (by-product). After the intestines have been removed, the belly of the fish needs cleansing, and this normally happens in a later station on the process line. In part, machines are used for this operation, and in part, it is performed

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

369

manually. The fish now are ready for filleting and are still kept chilled to ensure good quality. Large volumes of gutted fish (sometimes also the head cuttings) are transported from slaughter facilities to processing plants over huge distances by the use of refrigerated transport units. During transport the fish are kept in boxes with ice in sufficient quantities to keep them well refrigerated. A traditional Norwegian and Icelandic way to treat gutted and headless cod, is hanging it on wooden racks until well dried (stockfish). This takes place in northern Norway (mainly in Lofoten) and in Iceland, and is performed with considerable manual labour. This sector is automated only to a very small degree. Fish exceeding a certain size are split (except for the tail) in special splitting machines, which seen from the outside have a great resemblance to filleting machines. Another method of processing is salting. Lines for preparation and salting of white fish use machines, but handling and different operations are still performed manually. Salted white fish (mainly cod) are dried in special drying tunnels, and are turned into clipfish. Clipfish are mainly used as a basis for bacalao. Such processing lines also require manual operation to a large extent, even though a certain degree of automation has been implemented on some plants. The machine supplier Nordic Supply Systems is among the leading actors in the European market for splitting machines for saltfish/clipfish production. Fish caught by trawlers out on the fishing banks are gutted and frozen in blocks days before the trawler can deliver the cargo on shore. The processing aboard the trawlers is mainly automated, but manual feeding of machines, as well as assistance in some other operations is still required. When the fish are delivered at the processing plant ashore, they must be thawed (defrosted) in special defrosting appliances before further processing. The thawing process is mainly automated.

15.3.4 De-heading and filleting De-heading is an operation that must be performed prior to filleting. There are de-heading machines that execute this operation. These machines are, similar to the filleting machines, normally limited to operate on a specific species in a certain size spectrum. Normally these machines are fed manually, even though there exists to some extent equipment for feeding. Baader is a well-known supplier of de-heading machines for use with various fish. The Baader 409 is a machine designed for de-heading cod, haddock and saithe; 40–80 cm (round fish length), capacity: 60–80 fish/min, supporting two filleting machines. Marel (Iceland) is a supplier of a de-header called the CT 2620 designed for de-heading and tail-cutting salmon. The fish heads are a by-product of the fish processing lines, as are the intestines, and can be further processed in separate lines. Nevertheless, many of the by-products are presently used for production of ensilage, which again is further processed into dry substance (fish meal) and fish oil. These are ingredients in animal and fish fodder. The processing takes places in automated processing plants. De-headed fish are manually fed into filleting machines where the muscles of the fish are separated from backbone (spine), rib bones, fins and some of the skin.

© Woodhead Publishing Limited, 2013

370 Robotics and automation in the food industry

Fig. 15.8

Fish processing line from Baader, including filleting and trimming.

The fillets will still have their skin intact at this point. In a fish processing plant, the fillets will normally be moved onwards via conveyor belts on to a skinning machine, which removes the skin from the fillets. When the machines are performing as intended, the de-skinned fillets are transported from the skinning machines onto the conveyor belt without manual assistance. The machine supplier Trio Food Processing Machinery delivers skinning machines that remove the skin from the fillet in a very gentle manner where the fish muscle is not exposed to stress. This is done by freezing the fillets onto a rotating freeze drum (placed inside the skinning machine) and a drawknife shears the fillet off the drum, leaving the skin. Baader Food Processing Machinery is a major supplier of machines for de-heading, gutting, filleting and other processing of fish (Fig. 15.8). Baader has in the last few years developed combination machines which perform subsequent operations in one machine, thereby eliminating the need for separate conveyor belts between operations. The transport function is a built-in part of the combination machine. The main rule, though, is that such machines must be fed manually by an operator.

15.3.5 Trimming operations and de-boning Fillets of fish are normally trimmed, whether fillets of white fish, salmon fishes or other fish. Trimming operations include removal of unwanted elements (bones, fin or skin remnants, nematodes, belly fat areas, membrane, etc.), as well as correction of the shape of the fillet towards the sought-after standard. Some of these operations have been automated, but not all. Some machines provide the main part of the total operations. Nevertheless, there is often something that must be done manually. There are presently machines available that cut the fillet into the desired form. Also, there are machines that remove pin bones, but there are often bones left after the mechanical treatment, and these have to be removed manually. So far there have not been developed any machines for the

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

371

removal of nematodes, but a machine detecting these in the fish meat has been developed, based on the use of NIRS (Near Infrared Spectroscopy) (Heia et al., 2007; Stormo et al., 2007). At traditional processing lines for white fish, the fillets are trimmed manually by operators using filleting knives. Pin bones are removed via a V-shaped cut. During the same procedure, the operator divides the fillets into loins, belly flaps, centre cuts and tails. Other partitioning of the fillets is also possible. The Icelandic equipment supplier Marel delivers so-called flow lines, whereby these operations are performed manually by operators in the line. The control system, with its advanced software and weight sensors, gathers real-time information of all material flow within and through the flow line. Such lines are also available from other suppliers. Some years ago the Danish supplier Carnitech developed a prototype solution for removal of pin bones in white fish, such as cod and haddock. The limited use of this de-boner in the industry can possibly be explained by the fact that it also removes some fish meat when in use. The Norwegian supplier Trio Food Processing Machinery has for years been searching for new technology for the removal of pin bones both in salmon and white fish. Several types of solutions (machines and handheld equipment) are available from this company. Trio has also developed a patented method of pin-boning salmon fillets immediately after slaughter (pre-rigor). In order to cut down on operator work regarding trimming operations, Baader has recently developed a new combination machine that both removes the area in the fillet containing pin bones and also divides the fillet into different pieces. The machine is called the Baader 988 W, and is designed to operate on farmed cod and pollock, with a capacity of up to 80 fillets/min. The Baader 988 S is an advanced unique trimming machine for salmon and sea trout (Fig. 15.9). Each fillet is evaluated and the shape is trimmed according to stored trim parameters.

15.3.6 Portioning operations The fish fillets represent different qualities of meat when regarding the different parts of a fillet. A loin of cod fillet is considered to be the most valuable piece of the fillet, and is accordingly prized. The area called belly flap will usually fetch the lowest price. The most common partitioning of cod fillets is in loins, centre cuts, belly flaps and tails. For other fish, partitioning into other pieces can be considered. Regarding smoked salmon fillets, it is customary to smoke the fillets in one piece. After the smoking process, they can be divided into smaller pieces, but they can also be sold as whole smoked fillets. Salmon fillets and fillets of other fishes are also cut into portion-sized pieces and packed in plastic in packing machines. Marel and some other suppliers deliver machines for portioning fillets or round fish (Fig. 15.10). Packing machines are delivered by several suppliers (e.g. Multivac, Sealed Air, Cryovac, CFS and others). The portioning operation is conducted by feeding the fillets to the portioning machines on conveyor belts, and the cutting operation happens automatically. The portions are then conveyed on belts into sorting, further processing, freezing and/ or packing, according to need.

© Woodhead Publishing Limited, 2013

372 Robotics and automation in the food industry

Fig. 15.9

Baader 988 S (salmon trimming machine) is a relatively new machine from Baader Food Processing Machinery; capacity up to 40 fillets/min.

Fig. 15.10 Trimming and portioning operations at a flow line from Marel.

In-line weights in combination with mechanical solutions exist for sorting smaller pieces from the portioning operation, pieces that are not suited for further processing together with the clearly defined portion-sized pieces. Sorting systems called graders are normally based on in-line weights as process sensors. Modern weight systems of this sort have achieved high accuracy. They are also used when the end-packing is to contain a defined number of portion pieces that together should keep to a certain weight level. Systems such as these can be fully automated, but there are also systems based on a limited manual assistance.

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

373

Slicing fish fillets, for example salmon fillets, is also part of the industrial picture. Several suppliers deliver slicing machines. Some of these machines require manual feeding and assistance, while others are more or less automated. Some slicers are fed by use of conveyor belts, for example, the Trio FDS 105 HS from Trio Food Processing Machinery.

15.4 Automation in other unit operations in fish processing The current status and examples of automated solutions for some of the other unit operations in fish processing are briefly presented in the following text.

15.4.1 Pick-and-place operations There is presently available a huge spectrum of machines for processing and cutting fish and meat. The machine suppliers have gone to great lengths in optimizing the processes that are to take place in the machine, but they are rarely seen to be developing system solutions for automated feeding of the machines – some suppliers have developed solutions for feeding various machines, for instance by the use of conveyor belts, robots or similar (e.g. Marel, Baader, Cabinplant). There is also a need to handle fish, whole fish or pieces in the processing line. Good gripping solutions for fish (or meat), which is a non-rigid and undefined object, have not been available, impeding the development of good technical solutions. Due to this there are many operators in the fish processing industry still performing simple handling operations that could be expected to be automated. New concepts for gripping have been developed recently, which should have the potential to fill a need in the fish processing industry, but few systems of this sort have been put to use. In this area, further development is expected soon. Some new concepts for gripping have been outlined previously in the text. Cabinplant A/S, with headquarters in Denmark, has taken a completely new approach by combining food processing and packaging for fish (or shrimps) in an integrated and very compact process line. The new line is illustrated in Fig. 15.13. All process steps, such as cutting off heads and tails and the removal of viscera are performed during a combined pick-and-place operation. The robot moves individual products from the feed conveyor into the package (cans) by use of signals from the vision system, while controlling the filling weight. The idea of a single stroke procedure eliminates a number of handling steps required to link different processing stations (operations) by traditional methods. The number of robots is adapted to the line capacity requirements, and each robot is equipped with a multifunction tool. A cutting tool removes heads and tails, a suction tool removes viscera, and a gripper tool performs the pick-and-place operation. The new line is designed for small fish such as sardines, etc., and for shrimps. This is a very good example of new thinking and the use of robotics in seafood processing.

© Woodhead Publishing Limited, 2013

374 Robotics and automation in the food industry (a)

(b)

Fig. 15.11

I-Cut 300 portioning machine from Marel (a), and cut of white fish (b).

Fig. 15.12 A solution from Marel of pick-and-place operation of salmon portions.

15.4.2 Freezing and chilling operations Chilling and freezing of fish is necessary to avoid deterioration of quality. Out on the fish banks, pelagic fish are quickly chilled or frozen in plate coolers. Fish such as cod, pollock, saithe, haddock and similar, which are caught by trawlers far out at sea, are also frozen in blocks in plate coolers.

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing (a)

(b)

Fig. 15.13 The new processing line for sardines or shrimps from Cabinplant. (a) The robot is handling sardines by pick-and-place operations. (b) An overview of the line.

© Woodhead Publishing Limited, 2013

375

376 Robotics and automation in the food industry

Fig. 15.14

Process line from Marel where fish fillets are portioned, freezed and packed.

At processing plants onshore, chilling and freezing of processed produce will usually happen in designated areas (chilling or freezing tunnels) in the processing plant. Many such systems are equipped with conveyor belts for feeding produce into the machinery. In order to utilize the area of the conveyor belt, it is important that the placing of the produce on the belt is as compact as possible, without contact between the individual pieces of produce. Such operations are sometimes performed manually, but there are also some partially or fully automated solutions in existence. In other sectors of the food processing industry (e.g. chocolate and baking industry) such operations are solved by the use of fast-moving robots equipped with suitable grippers in combination with vision systems. A similar solution is developed by Marel for the handling of fish portions into a freezing tunnel.

15.4.3 Inspection and quality control In many branches of industry, many tasks related to inspection and quality control are still performed by operators along the production line. Humans are in possession of many senses and have the skill to evaluate different factors and make quick decisions. This often makes it difficult to replace them in such contexts. Nevertheless, there is a development whereby vision and possibly other sensor systems are being used for such tasks. The quality control of hen eggs has for many years been performed by such systems, and with a speed far exceeding human ability. In the automotive and electronics industries, such systems also gain ever more ground. Other branches of industry are also following this development. In the fish industry, such systems are not very common. Nevertheless, there are some good examples here as well. One example of successful automation was performed in a cooperation between AquaGen (supplier of genetic material for the aquaculture industry), the equipment supplier Maskon AS and the research institute SINTEF Fisheries and Aquaculture (Trondheim, Norway), resulting in a sorting machine for fertilized fish eggs. The machine identifies and removes eggs that do not satisfy the defined quality requirements. Previously, this operation was performed manually and required many operators. This new

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

377

machine performs the task with a very high quality and capacity (approximately 100 000 eggs/h) using a built-in vision system. Marel in Iceland has in recent years developed a system for the detection of bones in fish meat. The machine is called SensorX. The machine has two conveyor belts (for the right and the left fillet) for X-ray-based detection of bones. The machine detects fish bones of a thickness size down to 0.3 mm (Fig. 15.15). On the outflow side of the machine, a sorting device can be placed, which pushes fillets with bone remnants into separate stations for after-treatment. Here operators will deal with the individual fillet and at the same time study its X-ray image on a display in order to assist an easier and safer manual removal of remnant bones. The research and development project also included automated removal of pin bones in white fish. This is an example of positive technological development, which also can be put to use in other settings in the food processing industry. This development work was made possible by cooperation between Icelandic, Norwegian and Danish partners. Research work has also been undertaken where vision systems have been put to use in the inspection and sorting of different quality fish produce, as mentioned above in Section 15.2.3 on machine vision. It is to be expected that such systems in the future will be put to use to a much greater extent in the fish processing industry.

15.4.4 Further unit operations There are a variety of different unit operations inside a fish processing factory. A selection of operations has been mentioned in this chapter. Also, there are many stations and operations in relation to the farming of fish. In this sector there is quite a lot left to be desired in regard to the development of automated solutions.

15.5 Future trends Some future trends that will affect automation are briefly described in the following sections.

15.5.1 Pre-rigor processing Shortly after killing, fish enter a state of rigor mortis. In this state, processing (e.g. filleting) is rendered difficult, as the fish are not pliable. Conventional processing of, for instance, farmed salmon involves chilled storage of the salmon for approximately three days following the slaughter, to mature the meat. After this period, rigor mortis has ceased and pin bones are more easily removed from the fillets using mechanical processes. An alternative to the method outlined above is processing the fish prior to rigor mortis, a method depending on a low degree of handling stress at the time of slaughter, as high stress levels reduce the time window prior to rigor mortis. Processing fish prior to rigor mortis is called pre-

© Woodhead Publishing Limited, 2013

378 Robotics and automation in the food industry (a)

(b)

Fig. 15.15 The SensorX machine from Marel displaying x-ray images of fillets.

rigor processing. Pre-rigor processing of example farmed salmon (Salmo salar) allows for production of a super-fresh salmon fillet with firmer texture and generally improved quality, and simultaneously the time window for distribution and sale of the product increases. The transition to pre-rigor processing often leads to economical and environmental gain, for example, reduced costs related to chilled storage and transport, as well as a reduction of carbon dioxide emissions due to

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

379

decreased transport volumes (only fillets are transported). This transport is necessary due to the final processing often taking place near the end market. Pre-rigor processing of farmed salmon is in use at some fish processing companies, and this requires specialized methods. A high degree of automation in the processing plant, and processing lines with high uptime make it easier to perform pre-rigor processing of fish in the available time window. Pre-rigor processing is expected to become more widespread in the future, as the interest in Celsius super fresh fish raw material and fish products increases.

15.5.2 Super chilling Super chilling is a process where the product’s surface temperature is reduced to about 1–1.5 Celsius below freezing point (Nordtvedt and Stevik, 2009; Stevik et al., 2010). In the super-chilling process, small ice crystals are developed in the surface muscle tissue. The fish is pre-chilled in brine cooled with brash ice and thereafter super-chilled in a combined contact and convection freezer. During storage, the temperature in the product is equalized. The super-chilling process impedes the quality degradation processes and increases shelf life by several days. This increased shelf life can be used to regulate the stock of goods, or increase the number of days of sale. The increased level of coldness contributes to maintaining the product temperature even though the surrounding temperature is fluctuating. The slight increase of salt content in the product and the reduced temperature increases the strength of the fish muscle. The fish and fillets are therefore able to withstand (tolerate) more stress from the different unit operations compared to a fresh product, giving a higher yield. Traditional transportation of fresh fish and fish products requires use of ice in the transportation boxes to maintain low product temperature and to withstand changes in surrounding temperatures. Super-chilled products keep this cold as a reservoir and extra ice in the boxes is not necessary. Thereby, the weight of the transport goods is reduced by about 30% and this reduces transportation cost. The greatest challenge is to teach the fish market the advantages of superchilled fish products. Traditionally the boxes of fish contain ice when they arrive at the fish market. Boxes without ice indicate that the temperature has been too high and the shelf life of the products is reduced. The super-chilled products could therefore be rejected. The authorities and consumers are also sceptical of the prolonged shelf life. Increased knowledge of the process will eliminate this attitude. The consumers’ increasing interest of fresh seafood will be a driving force to establish value chains where super-chilling technology is an enabling technology.

15.5.3 Reconfigurable production lines to meet changing market trends Since early in the twentieth century, dedicated production lines have been used for mass production. In such lines, transfer line technology, with set tooling and automation, have been used. The goal was to cost-efficiently produce a specific

© Woodhead Publishing Limited, 2013

380 Robotics and automation in the food industry product in huge quantities maintaining an acceptable quality. Henry Ford (1863– 1947) was one of the first to use this kind of production line in the production of cars. This mentality is to a large extent in use also in the fish processing industry, especially in preliminary steps of the process where the fish are sorted and filleted, and unwanted elements are removed from the fillets. Conveyor belts are to a large extent used between different machines performing unit operations. Machines and transport devices form a coherent production line. By-products are separated from the main line and further processed on separate lines. After the fish have been processed into fillets and a variety of by-products, the task is to further process these raw materials into sales-friendly end products and suitably wrapped. The end products must satisfy demands and wishes from different customer segments. The markets for consumer products are to an ever increasing degree subject to unpredictability and quick changes – a fact that manufacturers of such goods must relate to. Globalization of the economy, ever more open markets, market fragmentation, faster access to new products/product varieties, and shorter product life spans are all parts of this picture. For the manufacturers this often means the need to produce an increased assortment of products, quick shifts between different products, and varied order sizes. In order to meet with the ever-changing market, manufacturers want more reconfigurable and flexible processing lines. In several countries, research regarding reconfigurable processing lines and which principles these should be based on has been conducted (Lutz and Sperling, 1997; Koren et al., 1999; Mehrabi et al., 2000). Presently this ability is to a moderate degree present in several of the fish processing companies. Flexibility is most often achieved through the employment of a considerable number of operators rather than by automation at present. However, it is possible to implement an increased flexibility in this area of production without being dependent on a huge staff of operators. This requires automated lines to be built according to new principles as reconfigurable production lines. Today, this is only seen to a very limited extent within fish processing companies. Interesting developments regarding these problems are expected in the years to come, and the fish processing plant of the future should take advantage of this development trend.

15.5.4 Enclosed production cells At processing plants where fresh food articles (e.g. fish) are being processed, a variety of machines and processing lines covering the need to produce the products in demand can often be found. The number of shifts per day operating these lines/machines can vary through the year, for instance due to seasonal variations. Through converting the lines/machines into enclosed cells where processing and cleaning/sterilisation can commence isolated from surrounding activities, these variations can more easily be brought about. These cells must be equipped with internal CIP or SIP systems, designed for automatic cleaning and disinfecting of process equipment without major disassembly and assembly work. Today it is usual to clean the whole production area at a time. This is not a restriction where enclosed production cells are used. Other factors, such as the possibility to control

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

381

the composition of the atmosphere (added gases, humidity, etc.) and temperature under closed processing of food articles, are also encouraging the development of machines/lines into encapsulated cells. Such cell design and control can provide a beneficial environment that maintains the quality of the food articles and prevents bacterial growth. In some situations, processing the foodstuffs at temperatures near freezing point, with defined atmospheric conditions, can be relevant. Increasing the degree of automation and robotisation will allow for such cells to perform the processing without personnel being in direct contact with the raw material, thus, under normal operation, eliminating the risk of contamination related to operators. Enclosed production cells will require higher investment at the time of procurement, but will bring advantages that will prove economically beneficial. This is a direction the design of production cells for fresh food stuff can be expected to take in some instances in the future.

15.5.5 Fully automated processing lines and integrated control systems Development of fully automated production lines is a challenging task due to the heterogeneous raw material. While automotive and mechanical industries are handling products with exact tolerance levels, the fish processing industries need technical solutions that are able to handle raw materials that differ in shape, size, texture, weight, and surface conditions, even though products are classified into suitable weight classes or product groups. Implementing automation solutions requires investment in technology, control systems and operators. When increasing the level of automation, the overall target is to reduce the production costs per unit. This economic profit is achieved through several potential benefits, such as increased yield and reduced wastage, improved efficiency, increased output rate, reduced labour turn over, and less difficulty recruiting workers. Historically, the financial justification has been based on labour savings. Companies are now realizing that more intangible benefits can far overweigh any labour savings. Replacing operators with robots, machines or other automated solutions gives several advantages, from increased production flexibility and output rate to increased uniform product quality and reduced costs. However, there are several challenges to overcome before implementing technical solutions, particularly related to the very heterogeneous raw material. Even though the raw material is sorted and graded before entering the production process, there are big differences in shape, size and texture. By changing over to more automated production lines, the companies free up the high flexibility of the operators that make them easily change between different operations. Operators carry out a variety of handling tasks adjusted to machine breakage within each operation, product flow, and quality control of product variations in texture, colour, shape and size. While the number of operators is reduced by increased automation level in the company resulting in reduced costs, there will be a shift in technical skills required because the operation and maintenance of robotic and automated systems will

© Woodhead Publishing Limited, 2013

382 Robotics and automation in the food industry require technical skills that did not previously exist within the company. The food industry has traditionally invested less in technical support than other industry sectors. In spite of many challenges in the seafood sector, the clear trend of more automated production lines in different industry sectors also will be visible in the seafood sector in the future.

15.5.6 Sustainable manufacturing Increased environmental awareness requires industry to meet more stringent environmental standards on a global basis. Manufacturing industries in general generate about 60% of annual nonhazardous waste (Ijomah et al., 2007). Significant opportunities exist for industry to reduce or prevent pollution through cost-effective changes in production, operation and raw material use (Wang et al., 2007). Focus on CO2 reduction is highlighted through the European Union which has the ambition by 2020 to reduce the CO2 emission by at least 20%, to increase the proportion of renewable energies in its energy mix to 20% and to reduce the energy consumption by 20%. While material productions are well documented, the production processes lack documentation in terms of environmental footprint. Optimization solutions may therefore be neglected when manufacturing processes are analysed in terms of ecological footprints (Duflou, 2010). Aspects of sustainable manufacturing will be more in focus in years to come, also in the seafood industry.

15.5.7 Traceability Traceability refers to full information about every step in a processing chain being accessible, for example in a processing operation with an appertaining distribution process. In recent years the authorities in different countries have introduced legislation which through a demand for systems with traceability provides documentation on the origin and process data, thus enhancing food safety. These systems are made on efficient through the use of modern technology, such as automation, bar codes, RFID technology and IT systems. Documentation on origin and all steps of the processing of seafood will be more and more relevant in the future.

15.6 Sources of further information and advice Some sources of further information are listed below: http://www.marel.com/ http://www.baader.com/ http://www.cabinplant.com/ http://www.stansas.no/ http://www.nordicsupply.no/ http://www.maskon.no/ http://www.trio.no/

© Woodhead Publishing Limited, 2013

Robotics and automation in seafood processing

383

http://www.aquagen.no/ http://www.simnet.is/basis/index.htm http://www.peruza.lv/eng/ http://www.cfs.com/ http://www.sealedair.com/default.aspx http://www.cryovac.com/en/default.aspx http://www.multivac.com/ http://www.kuka-robotics.com/

15.7 References Brett, P.N., A.P. Shacklock and K. Kodabandehloo (1991). “Research towards generalized robotic systems for handling non-rigid products.” IEEE Conference Proceeding 2 (2): 1530–1533. Chua, P.Y., T. Ilschner and T.G. Caldwell (2003). “Robotic manipulation of food products – a review.” The Industrial Robot 30 (4): 345. Davis, S., J.O. Gray and D.G. Caldwell (2008). “An end effector based on the Bernoulli principle for handling sliced fruit and vegetables.” Robotic and Computer-Integrated Manufacturing 24: 249–257. Duflou, J. (2010). “CO2PE cooperative efforts on process emission in manufacturing.” EREE Meeting, CIRP, PISA, August 22–28. Erzincanli, F. and J.M. Sharp (1997a). “A classification system for robotic food handling.” Food Control 8 (4): 191–197. Erzincanli, F. and J.M. Sharp (1997b). “Meeting the need for robotic handling of food products.” Food Control 8 (4): 185–190. Gjerstad, T.B., J.O. Buljo and T.K. Lien (2006). “Handle of non-rigid products using a compact needle gripper.” 39th CIRP IMS 2006. Ljubliana, June 7–9. Halveil, B. and Nierenberg, D. (2008). “Meat and seafood: The most costly ingredients in the global diet.” In State of the world. Innovation for a Sustainable Economy.” Harwes, R.J. (2004). “Improved prick and place gripper”. International publication number WP 2005/051812, The Worldwatch Institute. (Patent). Harwes, R.J. (2008). “Pick and place gripper device”. International publication number WP 2008/135720. (Patent). Haugen, J.E., E. Chanie, F. Westad, R. Jonsdottir, S. Bazzo, S. Labreche, P. Marcq, F. Lundby and G. Olafsdottir (2007). “Rapid control of smoked Atlantic salmon (Salmo salar) quality by electronic nose: Correlation with classical evaluation methods”. Sensors and Actuators B 116: 72–77. Heia, K., A.H. Sivertsen, S.K. Stormo, E. Elvevoll, J.P. Wold and H. Nilsen (2007). “Detection of nematodes in cod (Gadus morhua) fillets by imaging spectroscopy”. Journal of Food Science 72 (1): 11–16. Ijomah, W.L., C.A. McMahon, G.P. Hammond and S.T. Newman (2007). “Development of design for remanufacturing guidelines to support sustainable manufacturing.” Robotics and Computer-Integrated Manufacturing 23 (6): 712–719. Karakerezis, A., Z. Doulgeri, et al. (1994). “A robotic gripping system with consideration of grasping flat non rigid materials.” IEEE: 936–941. Koren, Y., U. Heisel, F. Jovane, T. Moriwaki, G. Pritschow, G. Ulsoy and H. Van Brussel (1999). “Reconfigurable manufacturing systems.” Annals of the CIRP, 48 (2): 527–540. Kuka (2011). Available from: http://www.hellotrade.com/kuka-robotics/wash-down-robot. html [Accessed 5 May 2011].

© Woodhead Publishing Limited, 2013

384 Robotics and automation in the food industry Lien, T.K., J.A. Hægstad and T.B. Gjerstad (2006). “A new thermal flow gripper for food products.” Annals of the CIRP 55 (1). Litzenberger, G. (2009). “World Robotics, Industrial Robots 2009”. International Federation of Robotics (IFR) Statistical Department. Lutz, P. and W. Sperling (1997) “OSACA – the vendor neutral Control Architecture.” Proceedings of the European Conference on Integration in Manufacturing IIM’97 Dresden, September 24–26. Mehrabi, M., G. Ulsoy and Y. Koren (2000). “Reconfigurable manufacturing systems: Key to future manufacturing.” Journal of Intelligent Manufacturing 11 (4): 403–419. Misimi, E. (2007). “Computer vision for quality grading in fish processing”. Trondheim: Norwegian University of Science and Technology, Faculty of Information Technology, Mathematics and Electrical Engineering, Department of Engineering Cybernetics (PhD-thesis). Natale, C.D., G. Olafsdottir, S. Einarsson, E. Martinelli, R. Paolesse and A. D’Amico (2001). “Comparison an integration of different electronic nose for freshness evaluation of cod-fish fillets.” Sensors and Actuators B 77: 572–578. Nordtvedt, T.S. and A. Stevik (2009). “Test av Superkjøling for filet produksjon.” Tekniske rapport ISBN 978-82-594-3480-7. SINTEF, Trondheim. Ozcelik, B., F. Erzincanli and F. Findik (2003). “Evaluation of handling results of various materials using a non-contact end-effector.” Industrial Robot 30 (4): 363–369. Rawal, B., V. Pare and T. Tripathi (2008). “Development of noncontact end effector for handling of bakery products.” The International Journal of Advanced Manufacturing Technology 38 (5): 524–528. Seliger, S., J. Stephan and K. Tripathi (2000). “Non-rigid part handling by new gripping device.” Proceedings of ICME. Stevik, A., A.S. Duun, T. Rustad, M. O’Farrell, H. Schulerud and S. Ottestad (2010). “Ice fraction assessment by near-infrared spectroscopy enhancing automated superchilled process lines.” Journal of Food Engineering 100: 169–177. Stone, R.S., P.N. Brett and B.S. Evans (1996). “An automated handling system for soft compact shaped non-rigid products.” IEEE International Conference on Systems, Manufacturing and Cybernetics: 3000–3005. Stone , R.S. , P.N. Brett and B.S. Evans (1996). “ An automated handling system for soft compact shaped non-rigid products.” Mechatronics 8 (2): 85–102. Stormo, S. K., A. H. Sivertsen, K. Heia, H. Nilsen and E. Elvevoll (2007). “Effects of single wavelength selection for anisakid roundworm larvae detection through multispectral imaging.” Journal of Food Production 70 (8): 1890–1895. Wallin, P.J. (1997). “Robotics in the food industry: An update.” Trends in Food Science & Technology 8 (6): 193–198. Wang, B., L. Jiang, J.W. Li, H.G. Cai and H. Liu (2005). “Grasping unknown objects based on 3d model reconstruction.” Proceedings of the 2005 IEEE/ASME: 461–466. Wang, L.K., Y.-T. Hung, H.H. Lo and C. Yapijakis (2007). Handbook of Advanced Industrial and Hazardous Wastes Treatment. Taylor & Francis Group, US. Wolf, A., S. Ralf and H. Schunk (2005). Grippers in Motion. Berlin Heidelberg, Springer-Verlag. Wögerer, C., G. Nittmann and P. Tatzer (2005). “Intelligent manipulation of non-rigid parts in industry applications.” Proceedings of the 2005 IEEE/ASME: 1120–1125. Zoller, Z., P. Zentay, Á.Meggyes and G. Arz (1999). “Robotical handling of polyurethane foam with needle grippers.” Periodica polytechnica, Mechanical Engineering 43 (2): 229–238.

© Woodhead Publishing Limited, 2013

16 Robotics and automation in the fresh produce industry N. Kondo, Kyoto University, Japan

DOI: 10.1533/9780857095763.2.385 Abstract: The first automated grading facilities for fruit and vegetables became available more than 10 years ago. Recently, machine vision and near infrared (NIR) technologies as well as mechatronics and computer technologies have been employed to make these facilities more sophisticated and have led to their use for many kinds of agricultural products. Robot technology has proved able to handle agricultural products delicately and with a high degree of precision, and to gather information to create a database of products every season. This information is then utilized as traceability data for consumers and as farming guidance for producers. Key words: fruit, grading, precision agriculture, robot, automation, information, traceability.

16.1 Introduction It is no exaggeration to say that of all agricultural operations, those carried out post-harvest have long employed the most automated equipment, following a great deal of research (Miller and Delwiche, 1991; Okamura et al., 1991; Rehkugler and Throop, 1986; Shaw, 1990; Tao et al., 1990; Lu and Ariana, 2002), because the correct environments for the introduction of automated instruments, such as machine vision, sensing systems, robots and PCs, are already in place. Since about ten years ago, packing robots and palletizing robots have been a frequent feature in fruit grading facilities (Njoroge et al., 2002), while grading robots (Kondo, 2003), which collect round-shaped fruits and inspect them using a machine vision system, are now being introduced in some East Asian countries. Mechanical systems

© Woodhead Publishing Limited, 2013

386 Robotics and automation in the food industry are more easily controlled within a facility than in the field, and harvested agricultural products are similarly more suited to automated processes than products on trees; as a result, automated systems are most frequently introduced within agricultural facilities. Automatic machines and robots have already removed the need for human operators to carry out heavy, dangerous and monotonous operations, have enhanced the market value of products, led to increased uniformity, improved hygienic/aseptic production conditions, and finally, have given future generations the possibility to achieve economic sustainability in small high value farming operations (Kondo and Ting, 1988). There has been worldwide concern about the quality of food supply because of recent problems with fruits, vegetables, and other processed foods: the problems include food poisoning by bacteria, illegal unregistered agricultural chemicals, lack of product authenticity, and so on. The impetus behind these trends may be attributed to an increase in consumer concern about health and well-being and a feeling that producers need to respond to these concerns by consistently providing products of guaranteed quality. Agricultural products present specific challenges in terms of quality inspection techniques that are not encountered with other industrial products: non-standard products must be inspected according to their appearance and internal quality, and in order for the product to be acceptable to consumers, only non-destructive methods can be used. Several sensors have been developed and used in assessing features relating to internal quality, including sugar content, acidity, rind puffing, rotten core, and other internal defects (Kawano, 2003; Ogawa et al., 2005). Automatic systems and robots used in agriculture then play another important role, as they are able to keep a precise record of their operations in databases. They then utilize that information for the next operation or store the data either for future use by the producer in decision-making or to provide traceability information for quality assured foods. This chapter describes the use of automated and robotized systems in postharvesting technologies of fruits and vegetables such as citrus fruits, deciduous fruits, and vegetables. Products are rarely the same color, size, or shape, with differences occurring even between examples of the same cultivar. Sensing systems for products with diverse and unpredictable characteristics are a particularly important type of automated technology in this respect. In this chapter, machine vision, one of the fundamental components of an automatic grading system, is explained first in general terms and then with specific reference to four automated post-harvesting systems for leeks, citrus fruits, eggplants, and deciduous fruits. Finally, a traceability system for food safety and security is described based on the technologies of data collection and utilization of grading facilities.

16.2 Machine vision system as a key technology Machine vision is an essential technology for inspecting agricultural products. Many machine vision systems have replaced the human eye in agricultural operations. Generally speaking, setting up a machine vision system requires the correct

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 387 lighting, hardware for image acquisition (usually a lens, TV camera, image capture board, PC, and cables between camera and capture board) and software for image acquisition and processing. Although many post-harvest systems may operate in controlled environments, some systems may be exposed to harsh environments that are not suitable for unprotected mechanical and electrical components. At fruit grading facilities in particular, precautions must be taken against fruit trichome and bloom of peaches and pears, which sometimes accumulate to depths of several millimeters during the course of a single day, in order to protect TV cameras and PCs. During summer, TV cameras must be able to operate in temperatures of up to 40°C in the facilities. Since agricultural products, such as fruit and vegetables, are extremely varied in nature, machine vision systems for post-harvest operations have also been developed to deal with more varied products. Many fruits and vegetables have glossy surfaces because of their cuticular layer. This may sometimes prove a barrier to the acquisition of high quality images of these products. The cuticle is a transparent layer impermeable to water and made from cutin and wax, which makes it glossy. Leaves, stems, fruits, and seeds have these layers, which play an important role in preventing water loss and evaporation through the surface. It has been observed that irregular fruit shapes can cause unexpected glares at certain points, or uneven illumination on the fruit surface, even when lighting conditions are perfectly adjusted for the fruit variety. The number of glares sometimes exceeds the number of light sources due to the irregular shape of the fruit. Eliminating glare and reflections from the surroundings, along with achieving uniform illumination conditions for the fruit, are the most important goals in the construction of a machine vision system. Kondo (2006) demonstrated a method that uses a polarizing (PL) filter in front of a halogen lamp, while protecting the PL filter from heat. Color cameras are the most common sensor in machine vision systems. Agricultural products exist in a wide variety of colors, which can provide information about their properties and condition. Color cameras create red, green, and blue (R, G, and B) component images and can convey colors similar to those perceived by human senses. Many cameras have been used in grading systems to measure the size, color, shape, and defects of fruits and vegetables and to look for evidence of germination in cell trays. Since the absorption band of plant parts that contain chlorophyll is around 670 nm, the R component often shows whether or not the product is mature. Photosynthesis requires relatively more R and B components, so more light is reflected from the leaves from the G component than from the R and B components. Images from a monochrome camera show the brightness of objects. The level of brightness intensity can indicate shine on the surface of fruits such as eggplants through the use of glare, which should be removed in color images (Kondo et al., 2007), while it can also measure product size and shape. Most monochrome charge coupled devices (CCD) are sensitive from the visible to the infrared region. Because the reflectance of agricultural products in the infrared region (700–1100 nm) is higher than that in the visible region (400–700 nm), it can be said that such

© Woodhead Publishing Limited, 2013

388 Robotics and automation in the food industry

Fig. 16.1 Rotten orange fruit images (right part is rotten): (a) black and white image; (b) fluorescent image.

CCD cameras also offer advantages in distinguishing agricultural products from other objects or from the background. Ultraviolet (UV) light is classified into three types according to wavelength: UV-C (100–280 nm), UV-B (280–315 nm), and UV-A (315–400 nm). It is claimed that UV-C light has disinfectant effects and reduces the amount of bacteria, viruses and mold. It is well known that UV-B light causes sunburn, and it is further hypothesized that it leads to the creation of vitamin D. UV-A is called black light and attracts insects, many of which have sensitivities in this range. In the UV-A range, some flower petals have high reflectivity in order to attract insects. Some agricultural products contain fluorescent materials and often fluoresce in the visible region due to excitation caused by the UV-A light. This often helps in the detection of defects in fruits, plants, and meats (Bodria et al., 2002; Kim et al., 2001). UV-A light sometimes excites fluorescent substances. It is well known that UV light at 365 nm causes damaged orange fruit skins to fluoresce (Uozumi et al., 1987) and the fluorescent images are used for detecting damaged fruits (Slaughter et al., 2008; Kondo et al., 2008) as shown in Fig. 16.1. X-rays are electromagnetic waves with wavelengths of 10–0.01 nm corresponding to frequencies of between 30 PHz and 30 EHz. They are shorter than UV rays and can be an effective way of inspecting the internal qualities of agricultural products due to the ease with which they can be transmitted. Although X-ray cameras may be used, another common means of obtaining images in this range is the use of a scintillation screen and a CCD camera. X-ray images are different from the images captured at other ranges, as shown in Fig. 16.2, because they provide transmissive, rather than reflective, information. The technology of computerized tomography (CT) has grown remarkably in recent years. CT is a non-destructive technique for capturing food images, which allows visualization of the internal structures or quality. To date, X-ray CT has been used for internal quality inspection: maturity of green tomatoes (Brecht et al., 1991), defects in melons and watermelons (Tollner 1993), internal qualities of peaches (Barcelon et al., 1997, 1999), stones in apricots (Zwiggelaar et al., 1997), measurement of water content of apples (Tollner et al., 1992), and

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 389

Fig. 16.2 Orange fruit images by color camera and X-ray camera: (a) black and white image; (b) transmissive image (top view); (c) transmissive image (side view).

Fig. 16.3 An apple damaged by a harmful insect: (a) CT image; (b) actual damaged apple.

evaluation of textural characteristics in nectarines exhibiting woolly breakdown (Sonego et al., 1995). Recently, a defect caused by an insect pest (Noctuoidea) in apples was investigated (Ogawa et al., 2003a) as shown in Fig. 16.3. Thus, machine vision technologies are used for inspecting agricultural products in a wide range of frequencies, from X-rays to infrared rays. Recently images using Terahertz waves – longer wavelength rays between far infrared and milliwaves – have started to be used for agricultural products in this region (Ogawa et al., 2003b, 2004), while it has been known that chemicals, including vitamins, sugars, amino acids, pharmaceuticals, and pigments, have specific absorption bands (Kawase et al., 2003; Walther et al., 2003; Wallace et al., 2004; Yamaguchi et al., 2005; Fukunaga et al., 2007). New findings in agricultural products or biological organs are expected, because as yet no studies have been carried out in this area.

16.3 Vegetable preprocessing and grading systems There are many kinds of fruits and vegetables. The grading operations and whole systems depend on their physical properties and producers’ strategies. Here, four

© Woodhead Publishing Limited, 2013

390 Robotics and automation in the food industry

Fig. 16.4 An image of a leek root just before cutting at the plate between stalk and roots.

different types of systems for leeks, oranges, eggplants, apples, pears, and peaches are introduced.

16.3.1 Peeling and grading of leeks It is essential that leeks are peeled before grading or shipping. The peeling operation is sometimes called a preprocessing operation. To properly peel the skin, the border plate between stalk and root must be cut. Because the plate is only a few millimeters thick, it is essential that the precise position of the plate is determined correctly before cutting. If it is cut too close to the stalk, the inner stalks are exposed and the market value drops because the inner stalks dry out during transportation. If it is cut too close to the root, the skin cannot be easily peeled, because the outer skin adheres to the plate. For this reason, a machine vision system that enables the precise detection of the position of the border plate between stalk and root in leeks is a key technology for leek peeling. Figure 16.4 shows an image of the plate between stalk and root. Figure 16.5 shows a leek preprocessing system currently in use in several Japanese agricultural facilities. Operators manually feed leeks onto the line of preprocessing machines, and once the position of the border plate is determined, the leeks are cut to a length of 60 cm. Before the position of the border plate position is determined, a blower is used to remove sandy soil attached to the lower stalk. In the second step, a three-fold root cutting operation is conducted because a single precise cut is not easy to achieve, due to the presence of soil and the irregular shape of the roots. In the last two cutting operations, two sets of monochrome cameras and color cameras are used. The monochrome cameras indicate the position of the border plate, while the color cameras detect the border between green leaves and white stalk for the third step. In the third step, the positions of the leeks on the line are adjusted so that the borders between green leaves and white stalks match with the positions of installed air nozzles. In the fourth step, high pressure air from a compressor is blown through nozzles and 20 leeks are peeled at a time.

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 391 Adjustment line

Leaf cutting

To grading

Peeling

Position adjustment

Root cutting (three times)

Feeding

Fig. 16.5 Leek preprocessing system (SI Seiko Co. Ltd).

Fig. 16.6 Acquired images of leek by color and monochrome cameras (SI Seiko Co., Ltd): (a) whole color image; (b) monochrome leaf image; (c) monochrome lower stem image.

In a recently used grading system for leeks, the grading operation is combined with the skin peeling operation described above. After peeling, several cameras inspect the length, width and bend of the white stalk as well as the number of leaves. Defects are identified from the color and monochrome images as shown in Fig. 16.6. Based on the results of the inspection process and on their weight, they are sorted into different grades. The most important evaluation indices are white stalk length and width, because the white part of the stalk is the edible part. A binding machine groups the leeks into bunches, and these are then packed into boxes for shipping.

16.3.2 Grading of oranges Orange grading operations have been automated for the last few decades. The system utilizes image-processing techniques and improved engineering design to convey fruits to the sensor, and then to detect fruit size, shape, color, maturity level, and taste. The system inspects each fruit with color CCD cameras stationed at six different angles to provide images of the fruit from all sides as shown in Fig. 16.7. Lighting is provided through lighting devices using halogen lamps or light-emitting diodes (LED) fitted with PL filters.

© Woodhead Publishing Limited, 2013

392 Robotics and automation in the food industry Judgement PC Sugar and acid PC X-ray image PC

Light interceptor

Light projector

X-ray camera

X-ray generator

Camera B in

Image A PC

Camera A in

Camera B out

Camera A Out Camera A top

Image B PC

DL Light

Camera B top

Spin, 180⬚C turn

Fig. 16.7 A schematic diagram of camera and lighting setup for citrus fruit grading. (Source: Njoroge et al., 2002.)

The labor-intensive process of hand-sorting fruit, traditionally considered the only accurate grading method, is completely eliminated by having a fully automated conveyance system that ensures that oranges are fed to the sensor at a constant speed. Unmanned line inspection has helped to solve the problem of labor shortage caused by an aging population and an insufficient number of successors, which has been experienced in most production areas. Specially designed rollers rotate the fruit 180° ensuring that the cameras can capture a complete view. Two basic inspection stages can be identified in this system: external fruit inspection and internal fruit inspection. In the external inspection stage, images from the CCD cameras set under random trigger mode are copied to the image grabber board fitted to the image-processing computer whenever a trigger occurs. The images are processed using specific algorithms for detecting color, size, bruises and shape. For internal fruit inspection, an NIR sensor determines the sugar content (brix equivalent) and acidity level of the fruits from light wavelengths received by spectrometers after light is transmitted through the fruit. The sensor photo-electrically converts the light into signals and sends them to the computer unit where they are processed and classified. In addition, the internal fruit quality sensor measures the granulation level of the fruit, which indicates its internal water content. Rind puffing, a biological defect that occurs in oranges, is inspected using an X-ray sensor. Output signals from image processing are transmitted to the judgment computer where the final grading decision is made based on features of the fruit and internal quality measurements. The speed of the conveyor in a citrus fruit grading system is usually 60 m/ min and several fruits are processed per second on each line. An important feature of the system design is that it can be adapted for the inspection of many other products, such as potato, tomato, persimmon, waxed apple, and kiwi fruit, with the only modifications required being adjustments to the processing codes. Multiple lines for orange fruit inspection, combined with high speed conveyors and microprocessors allow the system to handle large batches of fruit in a very short time. All the information collected, from the point when the fruit arrives at

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 393 Inspection system by TV cameras Monochrome Camera C Color Camera E Inspection system by TV cameras

Color Camera D

Monochrome Camera A Color Camera B Section of fruit feeding

Color Camera A NIR Camera

g tin ula g Sing eyor n i v d n ee Co it F or Fru nvey Co

Sorting Conveyor r Monochrome Camera D yo e v on Color Camera F gC

in ad Gr Turning over

Monochrome Camera B Color Camera C

g nin tur or Re nvey Co

Fig. 16.8 Inspection line of an eggplant grading system. (Source: Kondo et al., 2007.)

the collection site up to grading, packing, and shipping is stored in a database providing farmers with product-specific guidance (Njoroge et al., 2002).

16.3.3 Grading of eggplant Another type of grading system, designed for fruit with an elongated shape, such as eggplant, is also currently in use (Kondo et al., 2007). This inspects not only size, shape, color, and defects but also the gloss on the fruit surface. The grading line consists of feeding, singulating, grading, and sorting conveyors as shown in Fig. 16.8. Two sets of machine vision systems for grading fruit are installed in the grading conveyor, one before the fruit is rotated, the other after. Six color TV cameras provide images of the color, size, shape, and defects of the fruit (Chong et al., 2008a), while four monochrome cameras assign a degree of surface gloss to each fruit (Chong et al., 2008b) as shown in Figs. 16.9 and 16.10 respectively. Recently, the ability to inspect all sides of each fruit has become a highly desirable feature of a grading system. A rotary tray was developed, for use at a certain point on a grading line, in order to rotate the fruit while the color and monochrome machine vision systems carry out the quality inspection, as shown in Fig. 16.8. Each grading line consists of 348 rotary trays and runs at a speed of 38.1 m/min. Since this grading machine has six lines, it is capable of sorting a total of 504 000 fruits per day. This translates into an estimated potential processing capacity of 40.320 t per day.

16.3.4 Fruit grading robot A grading system using robots has been developed for use with deciduous fruits such as peaches, pears, and apples. This system automatically picks fruit from containers and inspects all sides of the fruit (Kondo, 2003). Figure 16.11 shows

© Woodhead Publishing Limited, 2013

394 Robotics and automation in the food industry

Fig. 16.9 Color images of eggplant: (a) original image; (b) binary image; (c) processed image (calyx, bruises and center line of fruit were detected).

Fig. 16.10 Monochrome images: (a) gloss surface fruit; (b) dull surface fruit.

the robot that picks the fruit from the containers (the providing robot) and the robot responsible for grading the fruit (the grading robot). A fruit container for 6 fruits × 5 lines (30 fruits), 6 fruits × 4 lines (24 fruits), or 5 fruits × 3 lines (15 fruits), depending on fruit size, is moved into the working area of the providing robot by a pusher (1). The providing robot has a three degrees-of-freedom (DOF) Cartesian coordinate manipulator and six suction pads as end-effectors. The suction pads come down and collect six (or five) fruits (2), which are then transferred to a halfway stage (3). Two providing robots work independently and move 12 (or 10) fruits to this halfway stage. A grading robot which consists of another 3 DOF manipulator (two prismatic joints and a rotational joint) and 12 suction pads picks them up again (4) and moves them to trays on the conveyor line. Images of the underside of the fruits are acquired as the grading robot moves over 12 TV color cameras, while four side images of fruits are acquired by rotating the fruits by 270°. The cameras turn for 90° following the movement of the grading robot (5). Once the image has been acquired, the robot releases the fruits into trays (7) and a pusher moves the trays to conveyor line (8). This grading robot’s maximum speed is 1 m/s and its stroke is about 1.2 m. It takes the robot 2.7 s to transfer 12 fruits to trays, 0.4 s to move down to the conveyor line, and 1 s to return once fruits have been released. 0.15 s are spent waiting for the next batch of fruit The total time to complete the operation is 4.25 s. This means that one set of robots can process approximately 10 000 fruits per hour and four robot sets can process 100 t in a day, if a fruit is assumed to weigh 350 g. This robot can be used in the grading of many round-shaped fruits. Figure 16.12 shows top images of 11 varieties of deciduous fruits – peaches, pears, and apples – which were graded by the grading robot.

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 395 Blower

x Grading robot Providing robot

z 5

4 3

6 7

Halfway stage TV camera

8

2

1

Pusher Lifter

Carrier Pusher

Fig. 16.11 A grading robot for deciduous fruits. (Source: Kondo, 2003.) Numbers in circles indicate: 1, pusher pushes 2 containers; 2, providing robot sucks 12 fruits together; 3, 12 fruits are spaced away from each other and are moved to halfway stage; 4, grading robot sucks 12 fruits together; 5, bottom images of 12 fruits are acquired by 12 TV cameras during moving from halfway stage; 6, 4 side images of 12 fruits are acquired during fruit rotation by 90 degrees turn down of the 12 TV cameras; 7, 12 fruits are released on carriers; 8, carriers are pushed to a grading line.

Fig. 16.12 Top views of apples, peaches and pears handled by the grading robot (upper images are apples, lower images are a peach and four pears).

A NIR inspector installed on each line measures the sugar content and internal qualities of the fruit after images have been captured by the machine vision system. Based on the inspection results, fruit is sorted into several grades and several sizes. After inspecting both the internal and external quality, another robot with 12 three DOF Cartesian coordinate manipulators packs fruits into a cardboard box. An inkjet printer can print the grade and size on the surface of the box on the line, and the box is then sealed. The boxes are palletized by an articulated robot before storage.

© Woodhead Publishing Limited, 2013

396 Robotics and automation in the food industry Production information Field management N, P, K, pH, EC SOM, MC, soil temperature, compactness.

Seedling production Crop management Operation record: Irrigation, fertilization, chemical spray transplanting, pruning

Harvesting

Grading

3D location of product Harvesting time and date

Plant variety, crop ID, size, color, disease

Crop ID, size, color, disease

Appearance (color, size, shape, bruise, disease), internal quality (sugar and acid contents, rotten core, inside defect), graded rank, chemical residue, ingredients, district name

Field ID (address, height above sea level), producer ID Climate, weather information (temperature, humidity, irradiation, precipitation)

Distribution information Transportation environmental condition (temperature, humidity) Package type and method, transporting method, time, and distance

Consumption information Selling price, selling time, information of arrival of goods, quantity to sell, opinions of consumers (on taste, freshness, etc.)

Fig. 16.13 Information accumulated in a database.

16.4 Information flow for food traceability and farming guidance As described in previous sections, in some grading facilities most operations are automated or robotized. That means that most of the information relating to grading can be stored in a database and be linked with other data such as producer ID, field ID, operation records kept by producers, weather information, geographical information, and so on. If the other operations were also robotized and automated, many kinds of information could then be obtained by the robots or automated systems. Operations at the production stage may be classified into several categories: field management, tree management, harvesting, and grading operations. Figure 16.13 shows the operations and possible information obtained on each operation at the production stage as well as at the distribution and consumption stages. Figure 16.14 shows a flow of data collection from field management in the production stage up to the consumption stage, with particular emphasis on information gathered by the grading robot, when data from the robotic system were stored in a database. Containers from producers have individual barcodes, which provide the producer ID, field ID, fruit variety and number of fruits in the containers to a PC. The data relating to each fruit processed by the grading robot are automatically stored in a PC and immediately sent to each Radio Frequency Identification (RFID) in the carrier through an antenna and a ROM-writer. Based on the data, each fruit in the carrier is sorted and packed into a box. In the box,

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 397

Grading robot

Sorting ROMWriter Grading ID Size Color Shape Defect

Barcode reader Reception ID Producer ID Field ID Received date Variety Container number Fruit number

Packing robot

IC memory

Product ID

PC

C

AB

123 Product ID issued date and time Packing robot No. Grading and reception information Transportation data Environmental condition Sales data Price, quantity

Soil sensor Field information Farming operation records Chemicals Fertilizers

http://www.***

Tracebility database

Product ID consumer

Producer

Fig. 16.14 A flow of data from grading robot for traceability information.

the fruit can still be matched with the fruit data stored thanks to the position of the fruit in the PC as long as this position is maintained. All the fruit data are collected in a traceability database, and are linked to the other production data shown in Fig. 16.13. The other data may be inputted to the traceability database in another way. If a problem occurs, rendering it necessary to trace all the data up to the production stage, it is possible to check the data of any fruit in a box at the distribution or consumption stage by typing the product ID and the fruit position in the box into a PC. There is another useful application of this information: it is obvious that the relationship between farming operations and the grading data collected in the database every year can be used by farmers as a tool for achieving precision agriculture. Thus, the automated grading systems and robots can assist farmers through their accurate collection of many kinds of information. That means that the data accumulated in a database can be used for traceability to ensure food safety and security foods and for the provision of farming guidance to producers. The most significant problem is then the question of who can manage such a large amount of data as that collected by the robots, sensors and producers. Figure 16.15 shows a solution whereby the management of the data is undertaken by an agricultural cooperative association (or company) or local farmers’ group. The agricultural association may establish a soil analysis center or product information center with inspection and analysis devices and sensors, as well as a database for conducting precision agriculture in the local area. The association can make use of many kinds of data for producers, because field information from a soil sensor or other methods, product information from the grading robot, and producers’

© Woodhead Publishing Limited, 2013

398 Robotics and automation in the food industry Residue Carbonization

Intelligent farming Precision farming

Soil sensor

Fresh product Information oriented fields

Operation records Sensing information

Farming guidance DSS for farmers

Field information

Grading robot Information added product Product information

Agricultural Cooperative Association

Transport

Soil analysis center Product information center

Biomass Re-uses

Variable distribution channel Marketing route

GIS, DSS Analysis of soil and chemicals Voice of consumer

BRAND

BRAND

Consumer

Quantity and marketing value Market (Distribution)

Flow of product and information

Fig. 16.15 Data management by agricultural cooperative association.

operation records are stored in the database. All the data are linked and analyzed on a geographical information system (GIS) and each producer is informed of the operations necessary for improved yields and higher quality products by a decision support system (DSS). From a different perspective, all the information can be made available to consumers and distributors through a web site. A trusted high quality product with precise traceability data may increase the market value as well as producing a local regional brand with a good reputation. If there is a problem with any of the products, quick action to resolve the issue is possible based on the opinions of consumers or distributors and on linked data in the database. This risk management is also an important role of the traceability system.

16.5 Conclusion In recent years, problems with the quality of food supply are becoming increasingly significant worldwide, with particular concerns surrounding, imported or exported foods due to the different criteria and unfamiliar food culture in many countries. To solve these problems, a food traceability system, which can show the history of the food during production, distribution and consumption, is desirable. Bioproduction robots and automated systems have replaced human labor

© Woodhead Publishing Limited, 2013

Robotics and automation in the fresh produce industry 399 and have contributed to increasing the market value of products and producing uniform products. In addition, recent experiences show that these robots and automated systems can also be useful for gathering and storing information on agricultural products and operations. Since the product information and operation history leads to foods with a higher level of safety and security the data can be useful to consumers in terms of food traceability. Furthermore, if information on postharvesting operations is linked with that from field operations, the resulting data can provide farming guidance to producers and can offer a valuable contribution to the development of precision agriculture systems.

16.6 References Barcelon, E. G., S. Tojo and K. Watanabe, 1997. X-ray CT scanner for detecting internal changes in peach. Proceedings of the International Symposium on Agricultural Machinery and Automation, Taipei, Taiwan, 227–232. Barcelon, E. G., S. Tojo and K. Watanabe, 1999. X-ray computed tomography for internal quality evaluation of peaches, Journal of Agricultural Engineering Research, 73, 323–330. Bodria, L., M., Fiala, R. Oberti and E. Naldi, 2002. Chlorophyll fluorescence sensing for early detection of crop’s diseases symptoms, ASAE Paper Number 021114. Brecht, J. K., R. L. Shewfelt, J. C. Garner and E. W. Tollner, 1991. Using X-ray-computed tomography to nondestructively determine maturity of green tomatoes. HortScience, 26, 45–47. Chong, V. K., N. Kondo, K. Ninomiya, T. Nishi, M. Monta, K. Namba and Q. Zhang, 2008a. Features extraction for eggplant fruit grading system using machine vision. Applied Engineering in Agriculture, ASABE, 24 (5), 675–684. Chong, V. K., T. Nishi, N. Kondo, K. Ninomiya, M. Monta, K. Namba, Q. Zhang and H. Shimizu, 2008b. Surface gloss measurement on eggplant fruit. Applied Engineering in Agriculture, 24 (6), 877–883. Fukunaga, K., Y. Ogawa, S. Hayashi and I. Hosako, 2007. Terahertz spectroscopy for art conservation. IEICE Electronics Express, 4, 258–263. KAWANO, S., 2003. Handbook for Food Non-destruction Measurement, Science Forum Inc. Kawase, K., Y. Ogawa, Y. Watanabe and H. Inoue, 2003. Non-destructive terahertz imaging of illicit drugs using spectral fingerprints. Optics Express, 11, 2549–2554. Kim, M. S., Y. R. Chen and P. M. Mehl, 2001. Hyperspectral reflectance and fluorescence imaging system for food quality and safety. Transactions of the ASAE, 44 (3), 721–729. Kondo, N., 2006. Machine vision based on optical properties of biomaterials for fruit grading system. Environment Control in Biology, 44 (3), 3–11. Kondo, N. P., P. Ling, M. Kurita, P. D. Falzea, T. Nishizu, M. Kuramoto, Y. Ogawa and Y. Minami, 2008, A double image acquisition system with visible and UV LEDs for citrus fruits. Proceedings of Food Processing Automation Conference on CD-ROM, ASABE, June 28 2008, Rhode Island, USA. Kondo, N. and K. C. Ting, Eds., 1998. Robotics for Bioproduction Systems. ASAE, St. Joseph, MI, 325p. Kondo, N., 2003. Fruit grading robot. In Proceedings of IEEE/ASME International Conference on Advanced Intelligent Mechatronics on CD-ROM, July 20–24 2003, Kobe, Japan. Kondo, N., K. Ninomiya, J. Kamata, V. K. Chong, M. Monta and K. C. Ting, 2007. Eggplant grading system including rotary tray assisted machine vision whole fruit inspection. Journal of the JSAM, 69 (1), 68–77.

© Woodhead Publishing Limited, 2013

400 Robotics and automation in the food industry Lu, R. and D. Ariana, 2002. A near-infrared sensing technique for measuring internal quality of apple fruit. Applied Engineering in Agriculture, 18, 585–590. Miller, B. K. and M. J. Delwiche, 1991. Peach defect detection with machine vision. Transactions of ASAE, 34 (6), 2588–2597. Njoroge, J., K. Ninomiya, N. Kondo and H. Toita, 2002. Automated fruit grading system using image processing. In Proceedings of SICE Annual Conference 2002, Osaka, Japan, MP18-3 (CD-ROM). Ogawa, Y., N. Kondo and S. Shibusawa, 2003a. Inside quality evaluation of fruit by x-ray image. In Proceedings of the 2003 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM2003) on CD-ROM, 1360–1365. Ogawa, Y., K. Kawase, H. Tashiro and M. Yamashita, 2003b. Water content measurement in plants by THz-imaging, THz2003 Technical Digest: 93. Ogawa, Y., K. Kawase, M. Mizuno, M. Yamashita and C. Otani, 2004. Nondestructive and real-time measurement of moisture in plant. IEEE Journal of Transactions on ELS, 124 (9), 1672–1677 (in Japanese). Ogawa, Y., N. Kondo and S. Shibusawa, 2005. Internal quality evaluation of fruit with X-ray CT. Journal of Society of High Technology in Agriculture, 17 (2), 75–83. Okamura, N. K., M. J. Delwiche and J. F. Thompson, 1991. Raisin grading by machine vision. ASAE Paper No. 91–7011. Rehkugler, G. E. and J. A. Throop, 1986. Apple sorting with machine vision. Transactions of ASAE, 29 (5), 1388–1397. Shaw, W. E., 1990. Machine vision for detecting defects on fruits and vegetables. Food Processing Automation, Proceedings of the 1990 Conference, ASAE: 50–59. Slaughter, D. C., D. M. Obenland, J. F. Thompson, M. L. Arpaia and D. A. Margosan, 2008. Non-destructive freeze damage detection in oranges using machine vision and ultraviolet fluorescence, Postharvest Biology and Technology, 48, 341–346. Sonego, L., R. Ben-Arie, J. Raynal and J. C. Pech, 1995. Biochemical and physical evaluation of textural characteristics of nectarines exhibiting woolly breakdown: NMR imaging, X-ray computed tomography and pectin composition. Postharvest Biology and Technology, 5 (3), 187–198. Tao, Y., C. T. Morrow, P. H. Heinemann and J. H. Sommer, 1990. Automated machine vision inspection of potatoes. ASAE Paper No. 90–3531. Tollner, E. W., 1993. X-ray technology for detecting physical quality attributes in agricultural produce. Posthavest News and Information, 4 (6), 149N–155N. Tollner, E. W., Y. C. Hung, B. L. Upchurch and E. E. Prussia, 1992. Relating X-ray absorption to density and water content in apples. Transactions of the ASAE, 35, 1921–1928. Uozumi, J., S. Kawano, M. Iwamoto and K. Nishinari, 1987. Spectrophotometric system for the quality evaluation of unevenly colored food. Journal of the JSFST, 34 (3), 163–170. Wallace, V. P., P. F. Taday, A. J. Fitzgerald, R. M. Woodward, J. Cluff, R. J. Pye and D. D. Arnone, 2004. Terahertz pulsed imaging and spectroscopy for biomedical and pharmaceutical applications. Faraday Discussions, 126, 255–263. Walther, M., B. M. Fischer and P. Uhd Jepsen, 2003. Noncovalent intermolecular forces in polycrystalline and amorphous saccharides in the far infrared. Chemical Physics, 288, 261–268. Yamaguchi, M., F. Miyamaru, K. Yamamoto, M. Tani and M. Hangyo, 2005. Terahertz absorption spectra of L-, D-, and DL-alanine and their application to determination of enantiometric composition. Applied Physics Letters, 86, 053903. Zwiggelaar, R., C. R. Bull, M. J. Mooney and S. Czarnes, 1997. The detection of soft materials by selective energy X-ray transmission imaging and computer tomography. Journal of Agricultural Engineering Research, 66 (3), 203–212.

© Woodhead Publishing Limited, 2013

17 Robotics and automation for packaging in the confectionery industry J. S. Dai, King’s College London, UK

DOI: 10.1533/9780857095763.2.401 Abstract: The confectionery industry is characterised by short runs and small batch production, thus confectionery packaging is frequently accomplished manually, resulting in wastage, injuries and hygiene issues. Automation of packaging in the confectionery industry would reduce these problems; and the frequent change-overs due to the large variety of products under manufacture necessitate an extremely versatile automated packaging system. This chapter describes market trends in the confectionery industry and the industry’s packaging needs and reviews reconfigurable mechanisms and the potential for flexible packaging automation in the confectionery industry. A case study of a reconfigurable demonstration system is outlined. Key words: packaging, automation, reconfigurable mechanism, robotics, confectionery.

17.1 Introduction The confectionery market is a highly seasonal and competitive market, with customers demanding unique products supplied in innovative packaging with a high on-shelf impact. Significant characteristics of the industry, especially in the production of seasonal products and other gifting products, are the range of packaging formats in existence, the speed with which they are changed, and the small production runs required to produce different products for different customers. Sales in the confectionery market are also increasingly driven by promotional activities including BOGOF (Buy One Get One Free) and other price promotions. The costs of these promotional activities are borne by the manufacturers, whose margins are being squeezed year on year. To be competitive, manufacturers are

© Woodhead Publishing Limited, 2013

402 Robotics and automation in the food industry being forced into using lean production, increasing productivity, reducing waste, utilising labour efficiently, and reducing machinery downtime. While some pickand-place operations are automated by using robots such as delta robots, endof-line packaging is still often achieved using manual work, due to the variety of products manufactured (Dai, 1996a, 2005). This is a more expensive process, though, which adds to the cost of the final product, particularly when an expensive seasonal labour supply is required. The above suggests that what is required is automatic reconfigurable confectionery handling and packaging in appropriate forms. This would enable manufacturers to exploit the seasonal confectionery market, allow leaner production with reduced rejects and downtime, and increase market opportunities. This chapter considers what is required for such a reconfigurable system and describes a demonstrator system that was created to illustrate what can be achieved.

17.2 The confectionery market and its business requirements The confectionery market is one of the major markets in the food sector and presents a major market for food packaging. This section describes the confectionery market in more detail, in particular the specific requirements of this market.

17.2.1 Market size In the confectionery market, France consumed 597 000 tonnes in 2000 and over 667 000 tonnes in 2006; Germany consumed 792 000 tonnes in 2000 and 833 000 tonnes in 2006. The United Kingdom has its place as one of the world’s major consumers of confectionery. Overall consumption of packaged confectionery in the United Kingdom increased by 6.5% to 992 000 tonnes in 2006 (see Fig. 17.1), with unit sales of food by carton approximately 2.4 billion units in 2001, and 2.6 billion units in 2006. According to Research and Markets (2007, 2009) and Plimsoll Analysis Confectionery (2011), the UK confectionery market achieved year-on-year growth to reach a value of £4.83bn in 2009. The confectionery market can be divided into two broad sectors: chocolate confectionery (including countlines, blocks, boxed chocolates and bite-size products), and sugar confectionery (including fruit sweets, mints and chewing gum). Chocolate confectionery accounts for nearly three-quarters of sales by value. Countlines continue to account for the largest share of the sector but boxed assortments are showing the fastest growth.

17.2.2 Use of packaging in the confectionery industry Confectionery represents a major market for food packaging, especially folded cartons. Examples of confectionery packaging can be found in Fig. 17.2. The total value for carton-boxed confectionery in the sales during the 16-week period

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 403

1000 tones

1000 950 900 850 800 750 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

Fig. 17.1 UK consumption of packaged confectionery. (Source: Packaging Machinery Technology, 2007.)

Fig. 17.2 Variety of confectionery packs.

before Christmas was 40% of the total. The value of carton-boxed confectionery from one factory alone was £12 million for that period in 2007. The manufacture of products to be sold during this period takes 4–5 months and the manufacture of products for Easter sales takes 3 months. Figure 17.3 indicates that confectionery represents a major market for cartons in the UK food market, with sales of about 2.4 billion units in 2001, which grew by 8% to 2.6 billion in 2006. The usage of other packaging methods, for instance bags, however fell by 4%. On the other hand, the usage of films continued to grow in this period by 14%.

17.2.3 Business requirements and commercial viability Seasonal products and other gifting products, such as those supplied to hotel chains, give the manufacturer and retailer a higher margin per kilo product than all-year-round (AYR) products. Consumers appear willing to pay a premium for these products because they are seen as unique and special purchases. However, as these products are seasonal and seen by the consumers as unique, their pack-

© Woodhead Publishing Limited, 2013

404 Robotics and automation in the food industry 3000 Food carton Films

2500 Million units

Rigid plastic tub 2000 1500 1000 500 0 1996

1998

2000

2002

2004

2006

Fig. 17.3 UK market for confectionery usage of packaging by product sector. (Source: Packaging Machinery Technology, 2007.)

aging is changed regularly. Different packaging formats are required for different customers and therefore production runs tend to be small. In the confectionery industry, manufacturers are limited in the range of carton types that can be used by the carton erecting and packaging machinery they possess. Usually this machinery is limited in the types of carton format it can erect and hence all new cartons require the development and manufacture of new machines or new tooling for existing machinery. New tooling is also required for each different pack size and format used. The development and manufacture of such tooling can be very expensive and increases the confectionery manufacturer’s lead time for introducing new products to market. This therefore reduces the manufacturer’s ability to react to changes in the demands of their customers. Tooling change-over, when changing production from one packaging format type to another, also adds cost to the confectionery manufacturer. In conventional machines, even dealing with a size change, more than 40 adjustments are required and can only be done manually (Dai, 1996a, 1996b; Cannella and Dai, 2006). During any change-over the machine is unproductive, and in most cases after the change-over, the machine requires a tuning period before full production can commence. This again reduces productivity (Dai, 1996a, 2005). Adding to the problem, most packaging machines are based on the 1950s and 1960s technology, largely using pneumatic actuators that are difficult to reconfigure as in Fig. 17.4. The alternative to use of machines is manual erection of cartons. This is a more expensive process, adding cost to the final product, particularly when an expensive seasonal labour supply is required. These manual workers must go through an expensive induction and training programme to teach them to erect a carton. Training and practice are required for each different pack type and the nature

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 405

Fig. 17.4 A conventional packaging machine.

of short runs and small batch production means that re-learning is required frequently for relatively short periods. The speed and quality of the carton erection process is reduced when compared to that of a dedicated machine, and waste is often increased greatly. As such, the manufacturer generally expects approximately 2% packaging reject from a carton erecting machine, but the initial stages of training operatives to manually erect complex cartons is likely to reach up to 35% material wastage. After training, this figure is likely to settle down to between 10% and 15% wastage. As packaging is, on average, 35% of the finished product cost, this adds a significant sum to the production costs for the confectionery manufacturer. A number of other problems are associated with the use of manual labour as follows: • Social problems: production is often geared to meet seasonal requirements (e.g. Easter eggs and Christmas production) and this results in recruitment problems and subsequent unemployment. • Labour injuries: repetitive motion causes injuries to human fingers and wrists, and the sharp edges of paper and carton material cause cuts to human fingers and hands. • Hygiene: this is a subsequent problem – for example if labour injuries lead to bleeding into the production area then the producer is required to carry out a complete deep clean of the production line.

© Woodhead Publishing Limited, 2013

406 Robotics and automation in the food industry It has been verified that in some industries such as the confectionery industry (Dai, 1996a; Dunn, 2006) reconfigurable packaging machines (Dai, 2006b; Dubey and Dai, 2006, 2007) are the means of achieving a flexible production to reduce the expensive costs of seasonal employment and to tackle the aforementioned problems. The machines are expected to be reconfigurable in order to adapt to different types of cartons and to follow different procedures for various closing methods. It has been suggested (Dai, 1996b; Dai and Caldwell, 2010) that this reconfigurability can be spread into a number of stations so that both productivity and utilisation can be increased. In this arrangement, with a higher productivity and reduced complexity resulting in reduced cost, the benefit from this technology is overwhelming. The cost saving can be illustrated from a case study. The cost of employment of ten seasonal workers is around £180 000 with annual inflation rate of 2.5% for subsequent years. In comparison, the cost of a reconfigurable machine that can complete the same task as ten workers is estimated as approximately £180 000. The payback is 10 months. The other quantifiable benefits include employment, wastage reduction, health and safety, and hygiene. Of course one of the difficulties in designing any form of cartoning machinery is that cartonboard is a highly non-linear material (Dai and Cannella, 2008; Beex and Peerlings, 2009). It is possible to model some forms of carton behaviour (Hicks et al., 2001; Dai and Rees Jones, 2002). Nonetheless the usual design strategy is to try to ensure that the worst cases are catered for. A number of researchers have considered various forms of board folding. This includes paper folding (Song and Amato, 2004; Balkcom and Mason, 2008), as well as origami carton folding (Dai and Cannella, 2008; Dai and Caldwell, 2010). Typical confectionery packs such as origami packs, which do not require the use of solvents in packaging, need a dexterous and reconfigurable machine with robotic fingers (Dubey and Dai, 2006; Luo and Dai, 2006; Dai and Caldwell, 2010) to produce packs with sufficient variety and complexity to attract customers. The demonstrator system in this chapter makes use of two forms of mechanism. There are basic folder units for making the folds around pre-defined creases (Dai et al., 2009a; Yao et al., 2010), and finger mechanisms for guiding carton faces in the appropriate direction. A number of finger mechanisms have appeared in the literature (e.g., Liu and Dai, 2003; Birglen and Gosselin, 2006; Luo and Dai, 2006; Wei and Dai, 2009) with varying numbers of degrees of freedom and hence potential application areas. Constraint modelling techniques were used for part of the design and simulation work for the demonstrator. These allow the design to be built up from what is known initially about the constraints, with the model being expanded as the designer’s confidence increases and knowledge about the design task becomes greater (Mathews et al., 2006; Mullineux et al., 2010). Constraintbased design is frequently used in this way (Mullineux, 2001; O’Sullivan, 2002; Hoffmann, 2005; Mattthews et al., 2006). The constraint modelling environment that was used makes use of optimisation techniques to resolve constraints (Hicks et al., 2006), although other approaches are possible (Hoffmann, 2005; Michelucci and Foufou, 2006).

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 407

17.3 Reconfigurable mechanism technology Reconfigurable mechanism technology is particularly applicable to the confectionery industry because of the variety of confectionery products that are packaged and the need for quick change-overs between different products. The extended reconfigurability and the versatility of machines using this technology would enable the machines to adapt to different packaging requirements and ensure the machines are fully utilised. 17.3.1 Introduction Reconfigurable mechanisms have been a research trend since the 1990s (Dai et al., 2009b; Dai et al., 2012). Mechanisms with a changeable topology (Kuo et al., 2009; Zhang and Dai, 2009) and a metamorphic process (Dai and Rees Jones, 1999; Gan et al., 2010; Zhang et al., 2010) have been investigated. Confectionery packaging presents a new scope for reconfigurable mechanisms and their study (Dai et al., 2009b; Dai et al., 2012). The development of machine reconfiguration for the confectionery industry requires research into carton-folding techniques and machine adaptability. Lu and Akella (2000) at the Beckham Institute used fixture techniques to describe carton folding based on a conventional cuboidal carton and used the similarity between carton motion sequence and robot operation sequence for operation. But the technology only focused on rectangular cartons. Though some machine prototypes have been produced to handle more complex cartons, these machines often only cover a single type of carton (or possible very similar types) and so lack reconfigurability. Dai and Rees Jones (2002) took a new approach and used an equivalent mechanism to describe a cardboard carton, this generates carton manipulation. Liu and Dai (2002) subsequently investigated carton manipulation and developed robotic fingers for carton folding. Dai and Cannella (2008) further revealed the characteristics of carton panels and creases during the folding motion that were used for the development of a novel packaging station. Neale et al. (2009) applied the constraint-based approach to the modelling and analysis of packaging machinery and to the creation of a new packaging machine. There have been opportunities for research on reconfigurable confectionery handling and packaging technology using the robotic finger principles (Yao and Dai, 2008) and constraint modelling (Mullineux and Matthews, 2010) to meet the business needs for confectionery handling and packaging.

17.3.2

Stages in the design and creation of reconfigurable mechanisms for confectionery packaging The following stages were used for design and creation of reconfigurable mechanisms for confectionery packaging (Dai et al., 2009a). Work analysis The motion of packaging cartons for confectionery is critical for folding and erecting mechanisms and for automation. This needs a generic analysis (Dai and

© Woodhead Publishing Limited, 2013

408 Robotics and automation in the food industry Rees Jones, 2002; Döring et al., 2006; Hicks et al., 2007; Yao and Dai, 2008; Wurdemann et al., 2010a) of the geometry, kinematics and sequence. The motion analysis generates a profile that covers similar types of cartons and differentiates the cartons by their categories. This analysis leads to the design and development of reconfigurable mechanisms. Modularity The system requires modular construction for diverse carton-folding functions including poking, tucking and squeezing operations using robotic end-effectors (Dubey and Dai, 2006; Wurdemann et al., 2010b). The modules, in the form of either folders or fingers, can be designed through the common motion and manipulation of carton folding. In particular, the packaging unit needs to be modular in order to be arranged on the production line. Reconfigurability The system needs to be capable of changing its modules by repositioning and reorienting them in the reconfigurable space, resulting in adaptability and versatility to suit multiple tasks. This can be realised by programmed motion and progress control and by mechanism reconfiguration. The position and orientation of modules varies with the shapes of cartons and their folding sequences (Dai, 1996b; Dai et al., 2009a; Zhang and Dai, 2009; Aminzadeh et al., 2010). Requirements The system needs to be designed to meet the requirements of customers and the constraints of the manufacturing environment. An automatic packaging process involves many subtasks. Generating physically valid folding sequences is the first to be considered (Liu and Dai, 2002; Song and Amato, 2004). The carton function is decomposed and the information is obtained from carton-folding analysis so that robotic fingers and folders can be designed. These modules are arranged in a lay-out by changing their positions and orientation for different cartons and folding functions. The flow chart of the analysis and synthesis is shown in Fig. 17.5.

17.4 Case study of a reconfigurable system for carton folding Reconfigurable packaging machines are the means to achieve flexible production and to reduce the cost of seasonal employment. A reconfigurable system (Dai et al., 2009a; Yao et al., 2010) was designed and produced to demonstrate the concept of reconfigurability and adaptability in food handling and packaging.

17.4.1 Reconfigurable demonstrator system The demonstrator system was created to handle a particular form of carton. This is a tray carton whose net is shown in Fig. 17.6. The main base panel is considered

© Woodhead Publishing Limited, 2013

© Woodhead Publishing Limited, 2013

Opening

Kinematic model

Analysis

Folding

Modelling

Motion sequences

Tucking

Reconfigurable robotic system

Simulation

Feeder

Folders

Modules assembly on positions and orientation

Fingers

Fig. 17.5 Design flow charts for reconfigurable packaging systems.

Pushing

Trajectory planning

Folding motion

Manipulation

Carton functional decomposition

Requirement

Conceptual design

Locking mechanisms

Synthesis

410 Robotics and automation in the food industry

Fig. 17.6 Folding trajectories of the faces and gussets of the tray carton.

as being fixed during the erection process. The four main faces need to be folded into the vertical position. Between these faces are four pairs of ‘gusset’ faces which move with the four main faces. The two end-faces are each attached to an additional rectangular face. The final stage in the erection process is to fold these over and into the tray. These then capture and retain the folded gussets and form double walls at the end of the completed tray. Tabs at the sides of the additional faces fit into cut-outs on the longer side faces and ensure the completed tray remains erect without the need to apply glue. Trajectories of four gussets and of the main faces are shown in Fig. 17.6. Based on carton folding and common-motions analysis, the reconfigurable system was designed for the carton tray folding. What was found was that certain faces of the carton net needed to be actively folded. Other faces followed passively, but these needed to be guided to ensure that they moved in the correct direction. Modules were created to show the reconfigurability of a robotic system for carton packaging. There were two types of modules designed to help fold the faces of the carton: robotic fingers (Yao and Dai, 2008; Yao et al., 2011) and folder mechanisms (Sirkett et al., 2007; Mullineux et al., 2010). The arrangement of the system is shown in Fig. 17.7. The end-panels of the folder mechanisms make contact with the faces of the carton net and cause them to turn when required. The panels can be changed depending on the size of the corresponding carton face. Suction cups within the end-panels grip the carton face to ensure that no slippage occurs during folding. Suction cups are also used to hold the base of the tray fixed while the other faces are folded around it. The robotic fingers are used to guide the passive faces of the carton. An end-effector, which is essentially a thin blade, pushes on the appropriate crease and moves with it as the carton erects around it. These modular fingers and folders are arranged on a base and can be moved appropriately depending on the size and design of the tray carton to be erected. 17.4.2 Prototype of reconfigurable system for folding carton trays As part of the design process for the reconfigurable demonstrator, prototypes of the basic robotic finger and folder mechanism were created and tested. This

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 411

Finger Carton Motor

Double 4-bar folder

Folder motor

Fig. 17.7 Arrangement of modules for tray folding.

enabled the appropriate motors to be selected and tested. It also ensured that satisfactory speeds could be obtained and that the torques created were sufficient to deal with the expected range of cartons. The fingers were designed to achieve planar motion with two degrees of freedom. This means the motors need to move together in a pre-determined relation. The folder consists of two folder sub-units. The sub-units are essentially the same and allow rotation about a single (virtual) axis. One sub-unit is sufficient to fold the long-side faces of the tray carton. Mounting one sub-unit on another allows the two folding operations on the end-faces to be achieved. In the final demonstrator system, the modules are mounted on an aluminium structure system that was built from extruded aluminium beams (of modular design) together with framework connectors and sliding elements. The system can be assembled quickly and economically using only hand tools. The structure also enables additional components to be easily integrated, such as pneumatic cylinders, sensors and cables.

17.4.3 Motion control for the demonstrator system The demonstrator system made use of servo-motors and there is a need to establish the control signals for these to achieve the appropriate motions (Sahinkaya et al., 2007). An experiment was undertaken to make sure the timings of the various operations were correct. Since vacuum cups were used to hold several of the faces of the carton net, sufficient time delays needed to be in place to ensure sufficient grip was obtained before folding took place. The experiment was designed to test the performance of the automatic packaging system. Figure 17.8 shows the flow chart of the process.

© Woodhead Publishing Limited, 2013

412 Robotics and automation in the food industry

Finger and folder in place Tuning motors

Start

Control maps linking

Suck carton blank

Motion outputs

End

Test

Lay-out

Fig. 17.8 Flow chart of the experiment.

π

Angle (rad)

Folder upper motor 3π/4 Folder lower motor π/2

Finger right motor

π/4

0

Finger left motor

0

10

20

30 40 50 Time (ms)

60

70

80

Fig. 17.9 Timing diagram from control simulation of fingers and folders.

At the start of this experiment the carton was sucked onto the base platform at the centre of the system. Fingers stand below the gusset corners and the orientation of the fingertips is along the creases of the gusset corners. Folders are arranged along the creases of the wall-panels. Figure 17.9 shows the timing sequence for tray cartons. The cycle time was 0.81 s, with a speed of 72 packs per minute. Synchronisation between the motors was another important factor, particularly for the finger modules where the position of the end-effector is determined by the position of its two motors. Each finger should stand under the base of the tray. The first step of the fingers should be to move up and break the crease of the gusset corner, in the interval 0–0.21 s. This manipulation is controlled by one motor. The second step is to push the broken gusset corner into the right position. This manipulation is controlled by the other motor. At the same time the ‘lower folders’ rotate 90° to fold and support the four side faces. After that the ‘upper folders’ rotate 180° to fold the double wall into its final position, requiring 0.4 s. Peaks occur during the movement of the end-effectors of two fingers from the end position to the start point. The design of the fingers means that these manipulations are more difficult than the folder’s manipulation.

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 413 Conveyor

Control unit

Pick-and-place unit

Folders

Fig. 17.10 A demonstrator in an industrial environment.

17.4.4 Testing of the demonstrator system An experiment was carried out in a demonstrator constructed (Dai et al., 2009a) based on Fig. 17.8 with reconfigurable modules. The general arrangement is shown in Fig. 17.10. Inductive sensors are used to mark the zero positions of the motors in order to avoid position errors. The folding process is shown in stages in Fig. 17.11. In Fig. 17.11a, the pickand-place unit extracts the next carton net from the supply stack. It moves it across to the erection unit (Fig. 17.11b), places it there and then moves away (Fig. 17.11c). In the version of the system shown, fixed ‘crease breakers’ are used instead of the fingers to initiate the breaking of the gusset creases as the net is placed onto the erector. The folder units act to erect the carton tray, and the erected form is seen in Fig. 17.11d. The pick-and-place unit then approaches the erected tray (Fig. 17.11e), contacts it (Fig. 17.11f) and then moves it to a conveyor. The total cycle is 30 units per minute, which compares well with the time achieved for manual erection. The model-based approach is used for co-operating two folders and two double-folders (Liu et al., 2008). The system is able to fold tray cartons of different sizes (with the same basic geometry) by changing the positions of the mechanism’s modules. The advantages of the system include that the system enables more reconfigurable and flexible solutions, especially for seasonal production, and reduces the cost for overcharges on packaging assembly line with its modular design. In this process the velocity and acceleration for every position segment is generated automatically (Quin Systems, 2001). The speed is limited by the maximum speed of the system. These maps are extracted from the timing diagram for the angular motion of fingers. This control was successfully demonstrated in the test rig and proves the manipulation planning strategy. The strategy consists of four steps: modelling the carton and generating the motion trajectory, manipulating the trajectory in the interactive space,

© Woodhead Publishing Limited, 2013

414 Robotics and automation in the food industry

Fig. 17.11 Folding cartons by the demonstrator in an industrial environment.

establishing a timing diagram of the motors, and linking the motion trajectories to controllers by maps. It has been verified in confectionery manufacture that the reconfigurable packaging robotic system is able to reduce the expensive costs of seasonal employment and tooling change-over. The cost saving can be illustrated from a case study in a manufacturer in the United Kingdom (Yao et al., 2010) as in Section 17.3.

17.5 Future trends The future trends are for flexibility and reconfigurability in the context of small batch production, which entails short runs and frequent change-over. This presents scope for research and development of reconfigurable mechanisms and machines (Dai et al., 2009b; Dai et al., 2012).

17.5.1 Reconfiguration Research and development work has indicated the potential of reconfigurable systems with respect to confectionery packaging. The automatic reconfigurable confectionery handling and packaging system arising from the work described in this chapter can enable the confectionery industry to further exploit the seasonal confectionery market. It also allows leaner production, reduces rejection rates and downtime and increases market opportunities. Between 10% and 30% of packaging waste would be reduced if the manual erecting of packaging could be replaced with a reconfigurable system. A reconfigurable system could be exploited in several ways. One way would be by providing a means whereby end-users can perform the reconfiguration for themselves (e.g. via software). An alternative would be to provide a service for end-users in which the specification for reconfiguration

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 415 is provided based on the details of the packaging design. There is scope for providing these details electronically and/or via internet communications. As mentioned above, reconfiguration techniques particularly benefit industries where variety and seasonal changes are required. In cases where the manual erection and closure of cartons is not economically viable and existing machines cannot adapt to the new product, the techniques developed and the machine prototype produced can provide a means of creating new products for growing markets. An example from the confectionery industry is the supply of confectionery to hotel chains where mini-bars require special confectionery products with forms of packs and assortment changing frequently. The use of reconfigurable-machine technology could potentially result in an increase in output/sales by confectionery manufacturers. In general, the current research and development of reconfigurable machines benefits the industry from machine making to food production and to food distribution. The technology is capable of being extended to other food manufacturing sectors and the idea of reconfigurability provides an opportunity to enhance food manufacturing. This will lead to improvements in working conditions and levels of employment and so help social and economic sustainability. Further, the concept of reconfigurability would lead to future development of machines based on the new technology. The potential market in the United Kingdom for the machine can be envisaged to be 5–10 machines per year, which represents up to £1.5 million. Also, possible savings on waste can be somewhere between £3000 and £10 000 per packing line per year.

17.5.2 Reduction of environmental impacts Since the research described above is aimed at the food and confectionery industry, most of the targeted packaging has a functional quality that is more important than its appearance or suitability as a gift package. Its chief function is to get the product to market in good condition, thereby reducing wastage. The research into reconfigurable-machine technology should have no negative environmental impacts. On the contrary, there are some other significant environmental benefits. These include: • Reduced level of material wastage compared with hand erected packaging. • Use of fewer machines than alternative solutions, and hence power savings. • Better transport/storage costs. Pre-glued hand-assembled cartons take up more space than flat machine-erect blanks that are used in this project. • If glues that contain solvents are avoided, it is easier to recycle packages. Flat blank machine-erect-cartons usually need a smaller area of carton board than pre-glued cartons. • Improved health and safety aspects. Less personnel involvement automatically reduces the accidents, RSI potential, burn injuries associated with working with hot glues and cut injuries from sharp edges of cartons. • The quicker and more efficient size-changes also reduce material wastage.

© Woodhead Publishing Limited, 2013

416 Robotics and automation in the food industry • Generally, packaging accounts for between 10% and 35% of the total cost of the product in the confectionery sector. This is already a huge cost to the manufacturer who is trading in a market driven by price promotion. The incentive is for the manufacturer to reduce the packaging cost, rather than to go for overpackaging of items.

17.6 Conclusion The confectionery industry is a highly competitive market where reconfiguration technology is needed to meet the demanding requirements of short runs, small batch production and frequent change-overs for a vibrant and demanding consumer market. Automation and robotics technology are hence leading the research in this industry particularly in the end of production lines for packaging of packaged confectionery products. This provides wide scope for researchers to use the reconfigurable mechanisms and robotics technology and to develop reconfigurable machines. This chapter further presents this challenging market and the pressing task associated with developing and using reconfigurable-machine technology; it also describes a reconfigurable robotic demonstrator system with a design based on the strategy of reconfigurability and developed in a factory. This reconfigurability technology helped to create a reconfigurable production line that meets the challenges faced in the confectionery and food industry.

17.7 Acknowledgements The demonstrator system described in this paper was created as part of a research project funded by the Department of Environment, Food and Rural Affairs (Defra) and the Engineering and Physical Sciences Research Council (EPSRC) of the UK under its Advanced Food Manufacturing LINK Programme in a collaboration with the University of Bath (Professors Tony Medland and Glen Mullineux) and a group of industrial companies including Bendicks Mayfair Ltd (Ms Evelyn Weddell), Quin Systems (Mr Mike Webb), and Marks and Spencer (Dr Mark Caul). This funding and support are gratefully acknowledged. Contribution to the demonstrator from Dr Wei Yao, Dr Ferdinando Cannella, Dr Lei Cui and Dr Guowu Wei of King’s College London is gratefully acknowledged. The author further acknowledges the financial support of the European Seventh Framework ECHORD project DEXDEB under grant number 231143 and in particular thanks Prof. Glen Mullineux for his friendly check of the manuscript.

17.8 References Aminzadeh, V., Wurdemann, H.A., Dai, J.S., Reed, J. and Purnell, G. (2010) “A new algorithm for pick and place operations,” Industrial Robot: An International Journal, 37 (6), 527–531. © Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 417 Balkcom, D. J. and Mason, M. T. (2008) “Introducing robotic origami folding,” The International Journal of Robotics Research, 27(5), 613–627. Beex, L. A. A. and Peerlings, R. H. J. (2009) “An experimental and computational study of laminated paperboard creasing and folding,” International Journal of Solids and Structures, 46, 4192–4207. Birglen, L. and Gosselin, C. M. (2006) “Kinetostatic analysis of underactuated fingers,” The International Journal of Robotics Research, 25(10), 1033–1046. Cannella, F. and Dai, J. S. (2006) “Crease stiffness and panel compliance of carton folds and their integration in modelling,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 220, 847–855. Dai, J. S. and Cannella, F. (2008) “Stiffness characteristics of carton folds for packaging,” Transactions of the ASME: Journal of Mechanical Design, 130(2): 022305_1–7. Dai, J. S. (1996a) “Survey and business study of the dextrous reconfigurable assembly and packaging system: DRAPS – CREF reconfigurable production plant,” Research and Project Report , PS96○321, Research. Dai, J. S. (1996b) “Conceptual design of the dextrous reconfigurable assembly and packaging system: DRAPS – CREF reconfigurable production plant,” Research and Project Report,PS96○322,Research. Dai, J. S. (2005) “Automatic reconfigurable confectionery handling and packaging system,” Food Link News, 53, 10–11. Dai, J. S. and Caldwell, D. G. (2010) “Origami-based robotic paper-and-board packaging for food industry,” Trends in Food Science & Technology, 21, 153–157. Dai, J. S. and Cannella, F. (2008) “Stiffness characteristics of carton folds for packaging,” Transactions of the ASME: Journal of Mechanical Design, 130, 022305:1–7. Dai, J. S., Medland, A. J. and Mullineux, G. (2009a) “Carton erection using reconfigurable folder mechanisms,” Packaging Technology and Science, 22, 385–395. Dai, J. S., Zoppi, M. and Kong, X. W. (2009b) “Editorial preface in reconfigurable mechanisms and robots,” Proceeding of the ASME/IFToMM International Conference on Reconfigurable Mechanisms and Robots (ReMAR 2009), KC Edizioni, June. Dai , J. S. , Zoppi , M. and Kong , X. W. (2012) “Preface,” Reconfigurable Mechanisms and Robots I, Springer, London, July. Dai, J. S. and Rees Jones, J. (1999) “Mobility in metamorphic mechanisms of foldable/ erectable kinds,” Transactions of ASME: Journal of Mechanical Design, 121 (3): 375–382. Dai, J. S. and Rees Jones, J. (2002) “Null-space construction using cofactors from a screw-algebra context,” Royal Society of London Proceedings Series A, 458, 1845–1866. Döring, U., Brix, T. and Reeβing, M. (2006) “Application of computational kinematics in the digital mechanism and gear library DMG-Lib,” Mechanism and Machine Theory, 41, 1003–1015. Dubey, V. N. and Dai, J. S. (2006) “A packaging robot for complex cartons,” Industrial Robot: An International Journal, 33 (2), 82–87. Dubey, V. N. and Dai, J. S. (2007) “Complex carton packaging with dexterous robot hands.” In Huat, L.K., ed., Industrial Robotics: Programming, Simulation and Applications. Mammendorf, Germany: pro Literatur verlag Robert Mayer-Scholz/Advanced Robotics Systems International, pp. 583–594. Dunn, J. (2006) “It’s chocs away as flexibility takes flight – new research aims to develop packaging machinery that combines automation with flexibility,” Food Manufacture, 29 March 2006, in the news, also available in http://www.foodmanufacture.co.uknews/ fullstory.php/aid/3059/It’s_chocs_away_as_flexibility_takes_flight.html Gan, D. M., Dai, J. S. and Liao, Q. Z. (2010) “Constraint analysis on mobility change in the metamorphic parallel mechanism,” Mechanism and Machine Theory, 45, 1864–1876. Hicks, B. J., Medland, A. J. and Mullineux, G. (2001) “A constraint based approach to the modelling and analysis of packaging machinery,” International Journal of Packaging Technology and Science, 14 (5), 209–225. © Woodhead Publishing Limited, 2013

418 Robotics and automation in the food industry Hicks, B. J., Medland, A. J. and Mullineux, G. (2006) “The representation and handling of constraints for the design, analysis, and optimization of high speed machinery,” Artificial Intelligence for Engineering Design, Analysis and Manufacturing, 20, 313–328. Hicks, B. J., Mullineux, G., Berry, C., McPherson, C. J. and Medland, A. J. (2007) “Energy method for modelling delamination buckling in geometrically constrained systems,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 217, 1015–1026. Hoffmann, C. M. (2005) “Constraint-based computer-aided design,” Journal of Computing and Information Science in Engineering, 5, 182–187. Kuo, C., Dai, J. S. and Yan, H. (2009) “Reconfiguration principles and strategies for reconfigurable mechanisms,” ASME/IFToMM International Conference on Reconfigurable Mechanisms and Robots (ReMAR 2009), 22–24 June, London, UK. Liu, H. and Dai, J. S. (2002) “Carton manipulation planning using configuration transformation,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 216 (5), 543–555. Liu, H. and Dai, J. S. (2003) “An approach to carton-folding trajectory planning using dual robotic fingers,” Robotics and Autonomous Systems, 42 (1), 47–63. Liu, H., Dai, J. S. and Seneviratne, L. D. (2008) “A model-based approach to cooperative operation of multirobot systems,” Industrial Robot: An International Journal, 35(1), 37–45. Lu, L. and Akella, S. (2000) “Folding cartons with fixtures: A motion planning approach,” IEEE Transactions on Robotics and Automation, 16 (4), 346–356. Luo, Z. and Dai, J. S. (2006) “Geometric analysis and characteristics of a three-fixed-pivoted multi-phalanx robotic finger,” Journal of Mechanical Engineering Science, 220, 1075–1082. Matthews, J., Singh, B., Mullineux, G. and Medland, A. J. (2006) “Constraint-based approach to investigate the process flexibility of food processing equipment,” Computers & Industrial Engineering, 51, 809–820. Michelucci, D. and Foufou, S. (2006) “Geometric constraint solving: The witness configuration method,” Computer-Aided Design, 38, 284–299. Mullineux, G. (2001) “Constraint resolution using optimisation techniques,” Computers & Graphics, 25(3), 483–492. Mullineux, G., Feldman, J. and Matthews, J. (2010) “Using constraints at the conceptual stage of the design of carton erection,” Mechanism and Machine Theory, 45, 1897–1908. Mullineux, G. and Matthews, J. (2010) “Constraint-based simulation of carton folding operations,” Computer-Aided Design, 42, 257–265. Neale, G., Mullineux, G. and Medland, A. J. (2009) “Case study: Constraint-based improvement of an overwrapping machine,” Journal of Engineering Manufacture, 223, 207–216. O’Sullivan, B. (2002) Constraint-Aided Conceptual Design, Professional Engineering Publishing Limited, London. Quin Systems (2010), Putting Imagination into Motion, available at: http://www.quin. co.uk/, accessed February 2011. Packaging Machinery Technology (2007) “Confectionery,” Larger Confectionery Pavilion to Make Pack Expo in 2007, Dusseldorf. Plimsoll Analysis Confectionery (2011) “A comprehensive study of the UK Confectionery market in 2010,” Plimsoll Publishing Ltd. Research and Markets (2007) “Snapshots UK Confectionery 2007,” Snapdata International Group. Research and Markets (2009) “Confectionery Market Report Plus 2009,” Key Note Publications Ltd. Sahinkaya, M. N., Rayner, R. M. C., Vernon, G., Shirley, G. and Aggarwak, R. K. (2007) “Synthesis of demand signals for high speed operation of a packaging mechanism,”

© Woodhead Publishing Limited, 2013

Robotics and automation for packaging in the confectionery industry 419 Proceedings of 2007 ASME International Design Engineering Technical Conferences & Computers and Information in Engineering Conference, IDETC/CIE, Las Vegas, paper DETC2007–35444. Sirkett, D. M., Hicks, B. J., Singh, B., Mullineux, G. and Medland, A. J. (2007) “The role of simulation in predicting the effect of machine settings on performance in the packaging industry,” Proceedings of the Institution of Mechanical Engineers, Part E: Journal of Process Mechanical Engineering, 221, 163–176. Song, G. and Amato, N. M. (2004) “A motion-planning approach to folding: From paper craft to protein folding,” IEEE Transactions on Robotics and Automation, 20, 60–71. Wei, G. and Dai, J. S. (2009) “Geometric and kinematic analysis of a seven-bar three-fixed-pivoted compound-joint mechanism,” Mechanism and Machine Theory, 45, 170–184. Wurdemann, H. A., Aminzadeh, V., Dai, J. S., Reed, J. and Purnell, G. (2010a) “Category-based food ordering processes,” Trends in Food Science & Technology, 22, 14–20. Wurdemann, H. A., Aminzadeh, V., Dai, J. S., Reed, J. and Purnell, G. (2010b) “Introducing a new 3D ordering process for discrete food products using food categorisation,” Emerald Group Publishing Limited, Industrial Robot: An International Journal, 37 (6), 562–570. Yao, W. and Dai, J. S. (2008) “Dexterous manipulation of origami cartons with robot fingers based on the interactive configuration space,” Transactions of the ASME: Journal of Mechanical Design, 130, 022303:1–8. Yao, W., Dai, J. S., Medland, T. and Mullineux, G. (2010) “A reconfigurable robotic folding system for confectionery industry,” Industrial Robot: An International Journal, 37 (6), 542–551. Yao, W., Cannella, F. and Dai, J. S. (2011) “Automatic folding of cartons using a reconfigurable robotic system,” Journal of Robotics and Computer-Integrated Manufacturing, 27 (3), 604–613. Zhang, K., Dai, J. S. and Fang, Y. (2010) “Topology and constraint analysis of phase change in the metamorphic chain and its evolved mechanism,” Transactions of the ASME: Journal of Mechanical Design, 132, 121001. Zhang, L. and Dai, J. S. (2009) “Reconfiguration of spatial metamorphic mechanisms,” Transactions of the ASME: Journal of Mechanisms and Robotics, 1 (1), 011012_1–8.

© Woodhead Publishing Limited, 2013

18 Automatic control of batch thermal processing of canned foods R. J. Simpson, S. F. Almonacid, Universidad Técnica Federico Santa María, Chile and Centro Regional de Estudios en Alimentos Saludables (CREAS), Chile and A. A. Teixeira, University of Florida, USA

DOI: 10.1533/9780857095763.2.420 Abstract: This chapter describes theoretical, practical and efficient strategies for on-line correction of thermal process deviations during retort sterilization of canned foods. It presents a review of recent developments over the past 10 years that have further advanced the state of the art in improving food safety, quality and manufacturing efficiency in the canned food industry worldwide. The focus is on retort control systems, and the various approaches that have been taken to help canned food processors accomplish on-line correction of unexpected process deviations, the major cause of lost productivity. Important features of each approach are discussed, along with suggested industry applications that would be appropriate for each method. The chapter concludes with a discussion of future trends to be expected in the industry. Key words: thermal processing, on-line control, batch processing.

18.1 Introduction1 Thermal processing is an important method of food preservation in the manufacture of shelf stable canned food. The basic function of a thermal process is to inactivate food spoilage microorganisms in sealed containers of food by using heat treatments at temperatures well above the ambient boiling point of water in pressurized steam retorts (autoclaves). Excessive heat treatment should be avoided because it is detrimental to food quality and underutilizes plant capacity. Food processing, and 1

Section 18.2.1 is reprinted from Simpson et al. (2007b); Section 18.2.2 is reprinted from Simpson et al. (2007a); Section 18.2.3 is reprinted from Simpson et al. (2007c); Section 18.3 is reprinted from Simpson et al. (2006). All material is reprinted with permission from Elsevier.

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 421 thermal processing in particular, is a strongly competitive global industry, driving continuous innovation and improvement of processing procedures and facilities. Control of thermal process operations in food canning factories has consisted of maintaining specified operating conditions that have been predetermined from product and process heat penetration tests, such as the process calculations for the time and temperature of a batch cook. Sometimes unexpected changes can occur during the course of the process operation such that the pre-specified processing conditions are no longer valid or appropriate, and off-specification product is produced that must be either reprocessed or destroyed at appreciable economic loss. These types of situations are known as process deviations. Because of the strong emphasis placed on the public safety of canned foods, processors must operate in strict compliance with the US Food and Drug Administration’s Low-Acid Canned Food (FDA/LACF) regulations (Teixeira and Manson, 1982). Among other things, these regulations require strict documentation and record-keeping of all critical control points in the processing of each retort load or batch of canned product. Particular emphasis is placed on product batches that experience an unscheduled process deviation, such as when a drop in retort temperature occurs during the course of the process, which may result from unexpected loss of steam pressure. In such a case, the product will not have received the established scheduled process, and must be either fully reprocessed, destroyed, or set aside for evaluation by a competent processing authority. If the product is judged to be safe then batch records must contain documentation showing how that judgment was reached. If judged unsafe, then the product must be fully reprocessed or destroyed. Such practices are costly. Commercial systems currently in use for on-line retort control accomplish on-line correction of process deviations by extending process time to that which would be needed had the entire process been carried out at the retort temperature reached at the lowest point in the deviation. In addition to the versatility and relative simplicity of this control strategy, it is also clear that it will always result in a safe correction, but is by no means optimal or efficient (Alonso et al., 1993; Simpson et al., 2007a).

18.2 On-line control strategies On-line control strategies can be used for the correction of thermal process deviations during retort sterilization of canned foods based on mathematical models.

18.2.1 Mathematically modelled foods Most mathematical models for the prediction of time-temperature histories in food products at a given point normally need to assume one of the basic modes of heat transfer. Two extreme cases have their own analytical solutions: (i) perfect mixing of a liquid (forced convection), and (ii) homogeneous solids (pure conduction). Most foods are an intermediate case, but these extreme cases have analytical solutions, then on-line control strategies can be designed for mathematically modeled foods.

© Woodhead Publishing Limited, 2013

TRT Prescheduled

T Prescheduled

TRT Corrected

T Corrected

Fo Prescheduled

Fo Corrected

140

30

120

25

100

20

80 15 60 10

40

Fo (min)

Temperature (°C)

422 Robotics and automation in the food industry

5

20 0

0 0

10

20

30

40 50 60 Time (min)

70

80

90

Fig. 18.1 Pure conduction simulation for on-line correction of an unexpected retort temperature deviation occurring approximately midway into the scheduled process for a cylindrical can of conduction-heated food Fo, accumulating lethality.

A control strategy for on-line correction without a change in process time The objective of this control strategy is to accomplish an on-line correction of an unexpected retort temperature deviation without changing the originally scheduled retort process time (see Fig. 18.1). Fundamental to the strategies presented in Sections 18.2.1 and 18.2.2 is the understanding that the retort control system would include a computer that is running the software containing the appropriate mathematical heat transfer model (most mathematical models for the prediction of timetemperatures histories in food products at a given point normally need to assume one of the basic modes of heat transfer) and that the computer continually reads the actual retort temperature from a temperature-sensing probe through an analogue/ digital (A/D) data acquisition system. This continual reading of retort temperature would be used as real-time input of dynamic boundary conditions for the mathematical heat transfer model. The model, in turn, would be accurately predicting the internal product cold spot temperature profile as it develops in response to the actual dynamic boundary condition (retort temperature), just as described by Datta et al. (1986). As the predicted cold spot temperature profile develops over time, the accumulating lethality (Fo) would be calculated by the General Method and known at any time during the process. Should a deviation occur during the process a simulated search routine would be carried out on the computer to find the combination of process conditions for the remainder of the process that would result in meeting the final target lethality without overextending processing time. The key in this strategy is to identify the retort temperature as the control variable to be manipulated during the remainder of the process (rather than process time). Therefore, upon recovery of the deviation, the search routine would find the new higher retort temperature to be used for the remainder of the process, and send the appropriate signals through the data acquisition system to readjust the

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 423 retort temperature accordingly. This entire procedure would be triggered whenever retort temperature falls below that specified in the FDA process registration file. Increasing retort temperature cannot be accomplished without increasing steam pressure correspondingly, which dictates a practical upper limit to choice of higher retort temperature. An upper limit of 135°C was chosen for the simulations carried out in this work. This upper limit comes into play when the deviation occurs near the end of the process, when the little time remaining forces the simulation search routine to choose the upper limit for retort temperature. In these cases, the safety requirement for reaching the final target lethality (Fo) must take priority overcompromising process time. This will inevitably require some extension in process time, but it will be an absolute minimum, that would not be likely to upset scheduling routines. Examples of this control strategy being applied for correcting deviations in the process (both early and late), were simulated for the case of thermal processes typical of still-cook conduction heating in a cylindrical can or retort pouch, and forced convection heating in a cylindrical can under mechanical agitation (Simpson et al., 2006). A control strategy for on-line correction of variable retort temperature process without compromising quality The objective of this control strategy was to accomplish an on-line correction of an unexpected retort temperature deviation while maximizing the final quality retention in the product, and to demonstrate the application of this strategy to dynamic (time-varying) retort temperature processes (see Figs 18.2 and 18.3). This strategy is meant to apply to those retort processes that were designed specifically for retaining maximum quality, such as nutrient retention, in the product. Canned foods in this category are usually slow conduction-heating products in relatively large containers in which significant non-uniform temperature distributions exist throughout the thermal retort process. These non-uniform temperature distributions provide opportunity to improve levels of quality retention by use of dynamic or time-variable retort temperature processes over traditional constant retort temperature processes. Results from process optimization studies have shown that most optimum retort temperature profiles for this purpose involve a ramp increase to a maximum retort temperature followed by a ramp decrease to a minimum prior to cooling (Teixeira et al., 1975; Saguy and Karel, 1979). For these types of products the heat transfer model must also be able to accurately calculate the quality degradation caused by the thermal process. The thermal degradation kinetics of most quality factors are such that these calculations are very sensitive to the non-uniform temperature distribution in such conduction-heated foods. Therefore, the conduction heat transfer models to be used for this purpose must be shape-specific for the shape of product container to be simulated. This is not necessarily the case when only lethality at the cold spot is of concern. The focus at this point is on use of the model’s ability to calculate volumeintegrated mass-average nutrient retention (as described by Teixeira et al., 1969) in response to time-varying boundary conditions (variable retort temperature). In the case of an unexpected retort temperature deviation occurring during the preprogrammed ramp-up or ramp-down of a variable retort temperature process the

© Woodhead Publishing Limited, 2013

424 Robotics and automation in the food industry TRT Prescheduled

T Prescheduled

TRT Corrected

T Corrected

Fo Prescheduled

Fo Corrected

140

30

120

25 20

80 15 60

Fo (min)

Temperature (°C)

100

10

40

5

20

0

0 0

10

20

30

40

50

60

70

80

Time (min)

Fig. 18.2 Dynamic process simulation for on-line correction of an unexpected retort temperature deviation occurring during the ramp-up portion of the scheduled dynamic retort process, for a cylindrical can of conduction-heated food. TRT Prescheduled

T Prescheduled

TRT Corrected

T Corrected

Fo Prescheduled

Fo Corrected

140

30

120

25 20

80 15 60

Fo (min)

Temperature (°C)

100

10

40

5

20

0

0 0

10

20

30

40 Time (min)

50

60

70

Fig. 18.3 Dynamic process simulation for on-line correction of an unexpected retort temperature deviation occurring during the ramp-down portion of the scheduled dynamic retort process, for a cylindrical can of conduction-heated food.

strategy would call for completing the process after recovery of the deviation at the optimum constant retort temperature-process time combination that would maximize final nutrient retention. Just as in the previous case in Section 18.2.1 the deviation would trigger a simulation search routine of finding a family of remaining retort

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 425 temperature–time combinations that would deliver the same final target lethality, and choosing the one that yielded maximum nutrient retention. Examples of this control strategy being applied for correcting deviations in the process (during rampup and ramp-down), have been simulated for the case of a thermal process typical of a still-cook conduction-heating product in a cylindrical can (Simpson et al., 2006). 18.2.2

A control strategy for on-line correction of retort temperature with minimum extended process time at the recovered retort temperature The objective here is to develop a strategy to correct the process deviation by an alternative ‘proportional-corrected’ process that delivers no less than final target lethality, but with near minimum extended process time at the recovered retort temperature. In addition, the strategy performance is demonstrated by comparing ‘proportional-corrected’ with ‘commercial-corrected’ and ‘exact-corrected’ process times. Finally, the aim is to demonstrate consistent safety of the strategy by exhaustive searches over an extensive domain of product and process conditions to find a case in which safety is compromised (Simpson et al., 2007b). Proportional correction strategy development The objective of this strategy is to accomplish on-line correction of an unexpected retort temperature deviation by an alternative process that delivers final target lethality, but with minimum extended process time (Fig. 18.4) at the recovered retort temperature. This would be accomplished with use of the same alternative process ‘look-up tables’ that would normally be used with currently accepted methods of online correction of process deviations, but with a ‘proportional correction’ applied to the alternative process time that would reduce it to a minimum without compromising safety. In order to fully understand this strategy, it will be helpful to first review the currently accepted method that is in common practice throughout the industry. Commercial systems currently in use for on-line correction of process deviations do so by extending the process time to that which would be needed to deliver the same final lethality had the entire process been carried out with an alternative lower constant retort temperature equal to that reached at the lowest point in the deviation. These alternative retort temperature–time combinations that deliver the same final process lethality (Fo) are called equivalent lethality processes. When these equivalent time–temperature combinations are plotted on a graph of process time versus retort temperature, they fall along a smooth curve called an equivalent lethality curve. These curves are predetermined for each product from heat penetration tests and thermal process calculations carried out for different retort temperatures. In practice, the new process times obtained from these curves at such low alternative temperatures can be as much as two or three times longer than the originally scheduled process time required to reach the same final target lethality, resulting in considerable quality deterioration and costly disruption to scheduled retort operations. Nonetheless, these systems are versatile because they are applicable to any kind of food under any size, type or container shape, as well as mode of heat transfer (Larkin, 2002). These consequences are particularly painful when, as in most cases the deviation recovers quickly, and the alternative

© Woodhead Publishing Limited, 2013

426 Robotics and automation in the food industry TRT Commercial correction

TRT Proportional correction

TRT Exact correction

TRT Scheduled process

140

Temperature (°C)

120 100 80 60 40 20 0 0

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 Time (min)

Fig. 18.4 Pure conduction simulation for on-line correction of an unexpected retort temperature deviation occurring early in the scheduled process for a cylindrical can of solid food under still-cook.

extended process time is carried out at the recovered original retort temperature. Canned food products subjected to such ‘corrected’ processes become severely overprocessed, with final lethalities far in excess of that required and quality deterioration often reaching levels below consumer acceptance (a safe correction, but by no means optimal or efficient – Alonso et al., 1993; Simpson et al., 2007a). To avoid these corrections, processors normally operate at retort temperatures 3–4°C over the registered retort temperature. The ‘proportional correction’ strategy presented in this section significantly avoids such excessive overprocessing by taking advantage of the short duration of most recovered retort temperature deviations, and the lethality delivered by carrying out the corrected process at the recovered retort temperature. The strategy will calculate the corrected process time (tD) as a function of the temperature drop experienced during the deviation, but also takes into account the time duration of the deviation. The following expression illustrates mathematically how this ‘proportional corrected’ process time would be calculated for any number (n) of deviations occurring throughout the course of a single process: n

tD

t

+ ∑ (tD i =1

⎛ Δt ⎞ tTRT ) ⎜ i ⎟ ; tD ≥ tTRT ⎝ tTRT ⎠

[18.1]

where n, number of deviations occurring during the process; tD, corrected process time; tTRT, pre-established process time at retort temperature TRT; Δti, duration of deviation i; tDi, process time at the deviation temperature TRTi; TRTi, lowest temperature during the deviation i; TRT, retort temperature. For example, in the case of a single deviation, the corrected process time would be calculated by first finding the alternative process time that would be

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 427 required to deliver the same final lethality had the entire process been carried out with an alternative lower constant retort temperature equal to that reached at the lowest point in the deviation (tD). This would be done by use of the equivalent lethality process curve or look-up table described earlier. The difference between this longer alternative process time and the originally scheduled process time (tD − tTRT) is the extra time that would normally be added to the original process time to correct the process according to current industry practice. However, in the new strategy proposed here, this extra process time differential (tD − tTRT) is multiplied by a proportionality factor consisting of the ratio of the time duration of the deviation to the originally scheduled process time (Δt/tTRT). This proportionality factor is always less than or equal to 1, and always results in a corrected process that delivers no less than the final target lethality specified for the original process, but with near minimum extended process time. The logic behind this ‘proportional correction’ strategy stems from the following rationale: • The current industry practice is necessary only when the deviation fails to recover, and retort temperature remains at the lowest point for the duration of the process. • This practice is unnecessary when the deviation recovers, and processing resumes at the original scheduled retort temperature over the ‘corrected’ extended time. • If the extended process time is chosen to be in proportion to the duration of the deviation as a fraction of original scheduled process time, we are making the assumption that the amount of lethality lost during the deviation duration time is the amount that would have accumulated at retort temperature. • In reality, this amount of lost lethality is much less, since the actual retort temperature had fallen during the deviation to some lower level where lethality would still continue to accumulate, only at a slower rate. • Therefore, the ‘proportional correction’ should always deliver total final lethality greater than that originally specified for the scheduled process. • With the implementation of this novel and efficient on-line strategy it will be unnecessary for processors to operate at higher retort temperatures.

18.2.3

On-line correction of process deviations without extending process time and without computer simulation software This section describes a method derived from the ‘proportional corrected process’ method of Simpson et al. (2007a) for accomplishing the same objective without the need of any computer simulation software (Simpson et al., 2007c). What is needed instead is a look-up table or curve on a graph showing alternative retort temperature–time combinations that were predetermined to deliver the same target lethality (iso-lethality curves) for each product. Normally, these are generated from heat penetration test data, and stored in a file for immediate access when needed. In this application the strategy involves finding the new retort temperature

© Woodhead Publishing Limited, 2013

428 Robotics and automation in the food industry to which the retort must be elevated for the process time remaining or at a minimum for a time period equal to the duration of the deviation. This temperature will be obtained from the iso-lethality process curve or look-up table of equivalent processes, once the equivalent process time associated with this temperature is determined. Mathematically, this equivalent process time can be calculated from the anatomy of the recovered process deviation and original process conditions as follows: tH =

2 tTR T tD

[18.2]

where tH is the equivalent process time at required higher temperature, tTRT is the original scheduled process time, tD is the equivalent process time at the lower deviation temperature. Derivation of Equation [18.2] stems from the ‘proportional corrected process’ method of Simpson et al. (2007a) by considering the fact that the lethality lost during an unexpected time period of lower than normal retort temperature (process deviation) can be recovered later in the process by an equal time period at a higher than normal retort temperature. Consider a process in which two deviations occur in sequence. The normal retort temperature is TRT, the first deviation occurs over a time interval Δt at lower than normal retort temperature TRTD and the second occurs later over an equal time interval Δt at a higher than normal retort temperature TRTH. If a ‘proportional corrected process’ is applied to each one of the deviations as described in Simpson et al. (2007a), the mathematical expressions for each correction will be as follows: Correction1 =

tTRT − tH Δt tTTRT

[18.3]

Correction2 =

tD − tTTRT Δt tD

[18.4]

where Δt is the duration time of the process deviation, tTRT is the originally scheduled process time at the original scheduled retort temperature, TRT, tH is the equivalent process time for same lethality at the higher retort temperature, TRTH, tD is the equivalent process time for same lethality at the lower deviation temperature, TRTD. If we assumed now that TRTH is selected so both corrections are equivalent, we can equate both terms: tTRT tH t t Δ t = D TRT Δ t tTRT tD

© Woodhead Publishing Limited, 2013

[18.5]

Automatic control of batch thermal processing of canned foods 429 tTRT tH tD tTRT = tTRT tD 1

tH tTTRT tH

1−

tTTRT tD

=

tTRT tD

⇒ tH =

2 tTR T tD

tTRT

[18.6] [18.7]

[18.8]

[18.9]

Experimental validation: case study The utility of this approach to on-line correction of process deviations was demonstrated experimentally as a means of preliminary validation (Simpson et al., 2007c) (see Fig. 18.5). Cylindrical cans (0.075 m diameter, 0.113 m height) containing a commercially prepared food product (Centauri Ravioli, 350 g) were thermally processed in a vertical still-cook retort under saturated steam with maximum working pressure of 40 psig (377.1 kPa) at 140°C (Loveless, Model 177). Both retort temperature and internal product cold spot temperature were monitored with K-type thermocouples, and recorded with an Omega 220 data logger and modem with COM1 connection port. Cans were processed under different combinations of retort temperature and process time, with the temperatures recorded every 2 s. Each normal process was defined with come-up-time (CUT) of 7 min, during which the retort temperature increased linearly, followed by a period of constant retort temperature and a cooling cycle. Deviations during the process were deliberately created by manually shutting off the steam supply to the retort control system. Experiments were carried out in the Food Laboratory pilot plant of the Universidad Técnica Federico Santa Maria in Valparaiso, Chile. In order to construct a iso-lethality curve it is necessary to use the data obtained from heat penetration tests conducted at well-controlled constant retort temperature, from which the process time needed to achieve a specified target lethality at any given retort temperature can be calculated. Therefore, using data from the constant-temperature heat penetration tests carried out in this study, an equivalent process lethality curve (for a target lethality of Fo = 6 min) was constructed for the commercial ravioli product and can size used in this study, and is shown in Fig. 18.6. In order to validate the safety assurance of this on-line control strategy, a number of heat penetration experiments were carried out in which process deviations were deliberately perpetrated by manual shut-off of the steam supply to the retort, causing the retort temperature and pressure to fall to a lower level for several minutes, after which the steam supply valve was reopened and the deviation quickly recovered. As soon as the complete anatomy of the deviation was known upon

© Woodhead Publishing Limited, 2013

430 Robotics and automation in the food industry 8

120

7 6

100

5 80

4 60

Fo (min)

Temperature (°C)

Commercial process 140

3 40

2

20

Tcp TRT Fo

0

1 0

0

10

20

30 40 Time (min)

50

60

Fig. 18.5 Profiles of retort (TRT) and internal product cold spot temperatures (Tcp) over time (scale on left), along with profile of accumulated lethality over time (scale on right) from a heat penetration test with ravioli in cans (0.075 m diameter, 0.113 m height) experiencing perpetrated process deviation immediately followed by temporary high retort temperature correction (calculated on-line). Equivalent process 180 160

Time (min)

140 120 100 80 60 40 110 112 114 116 118 120 122 124 126 128 Temperature (°C)

Fig. 18.6 Iso-lethality curve showing equivalent combinations of process time and retort temperature that achieve the same process lethality (Fo = 6 min) for ravioli packed in cylindrical cans (0.075 m diameter, 0.113 m height).

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 431 Table 18.1 Process conditions and parameters used for heat penetration test with ravioli packed in cylindrical cans (0.075 m diameter, 0.113 m height) producing results shown in Fig. 18.2 Process parameters (units)

Value chosen for heat penetration test

Reference temperature (°C) Scheduled process time (min) Low temperature at deviation (°C) Initial internal product temperature (°C) Time duration of deviation, Δ (min) Target process lethality, Fo (min) Microbial temperature factor, z (°C) Scheduled retort temperature (°C)

121.1 61 122 19 5.5 8 10 125

recovery, Equation [18.2] was used to calculate the high temperature equivalent process time (tHi), from which to obtain the higher retort temperature (TRTH) needed to accomplish the correction, using the iso-lethality curve in Fig. 18.6 The retort controller set point was immediately adjusted upward to the correction temperature (TRTH), and brought back down to the originally scheduled retort temperature after an elapsed time equal to the duration of the initial perpetrated deviation. The process was then allowed to proceed normally for the duration of the remaining originally scheduled process time. During each test, retort and internal product cold spot temperatures were continually measured and recorded, and accumulated lethality (Fo) was calculated as a function of cold spot temperature over time using the General Method (assuming a z-value of 10°C). Results from a typical test run are presented in Fig. 18.5, with the process conditions and parameters used for the test listed in Table 18.1. Both temperature and lethality are shown as functions of time on Fig. 18.5, with the temperature scale shown along the left side vertical axis, and the lethality (Fo) scale shown along the right side vertical axis. In the case of this test, the target value for process lethality (Fo) was 8 min and the normal scheduled retort temperature was intended to be 125°C for a scheduled process time of 61 min. The deviation was deliberately initiated after approximately 45 min into the process, and held for 5.5 min, during which time the retort temperature fell to 122°C. Upon recovery from the deviation, the retort temperature was elevated to approximately 128°C (determined from the calculation procedure described above) for five more minutes, and returned to the originally scheduled 125°C for the remainder of the scheduled 61 min process time. The measured retort temperature profile (TRT) can be seen in Fig. 18.5, clearly revealing the profile of the deviation immediately followed by the high temperature correction process and return to normal, with the cooling cycle beginning right on schedule at the originally appointed process time of 61 min. The measured internal product cold spot temperature curve (Tcp) in Fig. 18.5 reflects the expected erratic response to the combined deviation and correction perturbations experienced by the dynamic retort temperature. Most importantly, in spite of the

© Woodhead Publishing Limited, 2013

432 Robotics and automation in the food industry erratic profile of the internal product cold spot temperature, the final accumulated lethality (Fo), calculated as a function of this profile by the General Method, still reached the target value of 8 min specified for the process.

18.3 Validation of computer-based control systems In a review by Teixeira and Tucker, (1997), the state of the art in retort control strategies at that time was aimed primarily at new and better ways to achieve and maintain constant the retort temperatures specified for a scheduled process time, in the hopes of minimizing the chance occurrence of unexpected process deviations. Focus was placed on replacing traditional hard-wired relay-logic systems with new electronic programmable controllers capable of communicating directly with new computer-based control systems. Since then, the food processing industry has begun to rely increasingly on computer-based process control systems that rely heavily upon sophisticated software containing intricate programming codes developed by software engineers. The complexity and sophistication of these systems bring with them the potential for hidden errors and pitfalls ‘bugs,’ which could lead to serious problems if commissioned for use prior to proper validation. Government regulatory agencies responsible for assuring food safety for the consuming public have quickly come to recognize to need for such validation, and have been addressing this issue aggressively by working closely with industry in developing and proposing methods and procedures for carrying out documented validation of new computer-based control systems. Leading scientists from the US Food and Drug Administration (FDA) have often been invited to speak on this topic at major conferences, and are an excellent source for information on the approach to validation (Larkin, 2004). Another useful source of information is the work of McGrath et al. (1998), who described appropriate software and hardware tools that provide a powerful yet flexible platform from which to implement a process control strategy for the food processing industry. At the heart of validating any process control system is the need to know precisely what needs to be controlled. This extends far beyond the simplistic view of controlling equipment operating parameters. For example, it does little good to assure the retort temperature was held constant at the level specified for the scheduled process, if something went wrong during product preparation that caused the product to heat differently from how it was expected to heat. Identifying all the factors that must be controlled in order for a scheduled process to be assured effective is an important part of the approach to Hazard Analysis and Critical Control Points (HACCP) studies. These, in turn, are an important component of process and control system validation. Approaches to validation of thermal process control involving verification by the HACCP system were described well by Leaper and Richardson (1999). In that work they placed due emphasis on new approaches aimed at assuring safety without undue overprocessing, which would have a negative impact in

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 433 the marketplace because of compromised product quality and manufacturing efficiency. Chen and Ramaswamy (2003) conducted an extensive analysis of critical control points affecting deviations in thermal processing using artificial neural networks. They conducted an analysis of different critical process and product variables with respect to their importance on accumulated lethality, process time, cooling time, and total cycle time under various process conditions; along with the combined effects of deviations happening to multiple variables at the same time. These are only examples of the growing attention being paid to the need for process validation and its documentation in connection with the adoption of new computer-based control systems in the food process industry, and the food canning industry in particular. The importance of process validation cannot be overemphasized, and for this reason, it was chosen as a topic to be addressed in this chapter.

18.4 Industrial automation of batch retorts Many of the most recent advances made in the design of industrial batch retorts has come about in response to the increasing popularity of flexible retort pouches and retortable semi-rigid microwavable plastic dinner trays and lunch bowls. These flexible and semi-rigid containers lack the strength of traditional metal cans and glass jars to withstand the large pressure differences experienced across the container during normal retort operations. To safely process these types of flexible packages, careful control of overriding air pressure is needed during retort processing, and pure saturated steam, alone, cannot be used as the heat exchange medium. Instead, new retorts designed to be used with pressure-controlled steamair mixtures, water spray, or water cascade have been recently developed for this purpose (Blattner, 2004). Examples of some of these new retort designs are given in Fig. 18.7. A close-up view of some of the specially designed racking configurations used to hold flexible retortable packages in place during retorting is shown in Fig. 18.8. Perhaps the most significant advances made in the food canning industry to date have been in the area of automated materials handling systems for loading and unloading batch retorts. Traditionally, the loading and unloading of batch retorts has been the most labor-intensive component in food canning factories. Unprocessed sealed containers would be manually stacked into baskets, crates or carts. Then, the baskets or crates would be loaded into empty vertical retorts with the aid of a chain hoist, or wheeled carts would be loaded into horizontal retorts with the aid of track rails. In recent years leading manufacturers of retort systems have been hard at work designing and offering a host of new automated materials handling systems to automate this retort loading and unloading operation. Most of the new automated systems available to date are based on the use of either automated guided vehicles (Heyliger, 2004), or orthogonal direction shuttle

© Woodhead Publishing Limited, 2013

434 Robotics and automation in the food industry

Fig. 18.7 New retort systems (rotating and still-cook) with specially designed racking configurations for processing flexible and semi-rigid packages. (Source: Courtesy of ALLPAX, Covington, LA.)

systems (Blattner, 2004; Heyliger, 2004). Both types of systems are designed for use with horizontal retorts. The automated guided vehicles (AGV) work like robots. They carry the loaded crates of unprocessed product from the loading station to any designated retort on the cook room floor that is ready to be loaded. They also carry the loaded crates of finished processed product from the unloaded retort to the unloading station for discharge as out-going product exiting the cook room to the case packing operations. These robotic AGV’s are designed to integrate with the loading station in such a way that sealed product containers arriving

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 435

Fig. 18.8 Rack designs for flexible and semi-rigid retortable packaging systems. (Source: Courtesy, ALLPAX, Covington, LA.)

on a conveyor automatically stack into the crate carried by the AGV, which later inserts the entire crate into the designated retort. Unloading at the unloading station for finished product discharge is likewise accomplished in a similar automated way, but in reverse. The AGVs are guided by an underground wire tracking system buried beneath the cook room floor. This leaves the cook room floor space open and free of any rail tracks or guide rails that would otherwise impede the safe movement of factory workers in their normal work flow operations. A panoramic

© Woodhead Publishing Limited, 2013

436 Robotics and automation in the food industry

Fig. 18.9 Automated batch retort system with use of AGVs in large cook room operation. (Source: Courtesy, FMC Food Tech., Madera, CA.)

Fig. 18.10 Automated guided vehicle for batch retort loading/unloading. (Source: Courtesy, FMC Food Tech., Madera, CA.)

view of a large cook room operation using an automated batch retort system with AGV is shown in Fig. 18.9 (Heyliger, 2004), and a close-up view of an automated guided vehicle in the process of loading or unloading a horizontal retort is shown in Fig. 18.10. An alternative to the AGV system is the shuttle system offered by several retort manufacturers. Unlike the AGV system, the shuttle system relies upon a set of tracks or rails that are fixed in place on the cook room floor. These rails span the length of the cook room along the row of horizontal retorts, allowing a shuttle carrying loaded crates to slide along these rails until it has aligned itself in front

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 437

Fig. 18.11 Automated shuttle-based batch retort control system. (Source: Courtesy of ALLPAX, Covington, LA.)

of the designated retort waiting to be loaded. In a similar fashion, when a retort is ready for unloading, an empty shuttle slides along these rails until it has aligned itself with that retort to receive the loaded crates of processed product. Then the shuttle slides along the rails to the far end of the cook room where unloading of processed product takes place for discharge out of the cook room. Normally, the unprocessed product loading station and the processed product unloading stations are located at opposite ends of the cook room (Fig. 18.11). Figures 18.12 and 18.13 illustrate the shuttle systems offered by ALLPAX and FMC, respectively.

18.5 Advances in research and future trends If the food canning industry is to continue to remain competitive in an ever-expanding global market, technological advances will be needed to increase productivity, achieve better product quality with enhanced safety assurance, and all at lower and lower cost; this means that advances in automation and intelligent on-line control will inevitably continue at a rapid pace. New developments that are likely to occur soonest will be the application of computer-based retort control systems for on-line correction of process deviations. These developments

© Woodhead Publishing Limited, 2013

438 Robotics and automation in the food industry

Fig. 18.12 Automated shuttle batch retort system. (Source: Courtesy of ALLPAX, Covington, LA.)

Fig. 18.13 FMC shuttle system for automated batch retort loading/unloading. (Source: Courtesy, FMC Food Tech., Madera, CA.)

© Woodhead Publishing Limited, 2013

Automatic control of batch thermal processing of canned foods 439 are likely to occur first in relatively small canneries with labor-intensive batch retort operations located in developing countries. These are the companies with the greatest need and that are most receptive to adopting new technology appropriate to their level of processing operations. Very simple systems are now under development intended for this market. These systems will consist only of a microcontroller equipped with software containing the mathematical heat transfer models described in this chapter. These lap top computers will communicate with commercially available data acquisition modules (data loggers), and serve as traditional data acquisition systems, but with a twist. The data logger will be continuously reading the retort temperature, and sending this information to the lap top computer. From these data and with the mathematical heat transfer model, the computer will be continuously calculating the increasing internal cold spot temperature and associated accomplished lethality, and comparing this accomplished lethality with the specified target value for the process. When calculated accomplished lethality reaches the specified target value, the lap top computer will signal the operator to take the necessary action to end heating, and commence cooling. These systems will involve no computer-controlled actuators to automatically shut off or turn on valves. Instead the operator will be expected to be watching the monitor as the process is under way, and operate the retort as always (‘opto-digital’ control). The operator will simply wait until the computer indicates when to turn off the steam, rather than doing so when indicated by a stop watch. In the area of materials handling automation, the industry may witness a move to more and more sophisticated robotics that may ultimately replace the shuttle systems and AGVs that have become the state of the art today.

18.6 References Alonso, A., Banga, J. and Perez-Martin, R. (1993). A new strategy for the control of pressure during the cooling stage of the sterilization process in steam retorts. Part I. A preliminary study. Food and Bioproducts Processing, Trans IchemE, 71(c), 197–205. Blattner, M. F. (2004). Advances in automated retort control, and today’s new packaging. Presentation at IFT Symposium, 2004 IFT Meeting, Las Vegas, NV. Chen, C. R. and Ramaswamy, H. S. (2003). Analysis of critical control points in deviant thermal processes using artificial neural networks. Journal of Food Engineering, 57, 225–235. Datta, A. K., Teixeira, A. A. and Manson, J. E. (1986). Computer-based retort control logic for on-line correction of process deviations. Journal of Food Science, 51(2), 480–483, 507. Heyliger, T. L. (2004). Advances in retort control for batch and continuous systems. Presentation at IFT Symposium, 2004 IFT Meeting, Las Vegas, NV. Leaper, S. and Richardson, P. (1999). Validation of thermal process control for the assurance of food safety. Food Control, 10, 281–283. Larkin, (2002). Personal Communication. Branch Chief, National Center for Food Safety and Technology, Food and Drug Administration (FDA/NCFST), Chicago, IL, USA.

© Woodhead Publishing Limited, 2013

440 Robotics and automation in the food industry Larkin, J. W. (2004). Validation of software-driven control systems. Symposium presentation at 2004 Annual IFT Meeting, July 17–21 2004, Las Vegas, NV. McGrath, M. J., O’Connor, J. F. and Cummins, S. (1998). Implementing a process control strategy for the food processing industry. Journal of Food Engineering, 35, 313–321. Saguy, I. and Karel, M. (1979). Optimal retort temperature profile for optimizing thiamine retention in conduction-type heating canned foods. Journal of Food Science, 44, 1485–1490. Simpson, R., Figueroa, I. and Teixeira, A. (2007a). Simple, practical, and efficient on-line correction of process deviations: Batch retorts operations. Food Control, 18(5), 458–465. Simpson, R., Teixeira, A. and Almonacid, S. (2007b). Advances with intelligent on-line retort control and automation in thermal processing of canned foods. Food Control, 18(7), 821–833. Simpson, R., Figueroa, I. and Teixeira, A. (2007c). Preliminary validation of on line correction of process deviations without extending process time in batch retorting: Any canned foods. Food Control, 18(8), 983–987. Simpson, R., Figueroa, I. and Teixeira, A. (2006). Optimum on-line correction of process deviations in batch retorts through simulation. Food Control, 17(8), 665–675. Teixeira, A. A. and Manson, J. E. (1982). Computer control of batch retort operations with on-line correction of process deviations. Food Technology, 36(4), 85–90. Teixeira, A. A. and Tucker, G. S. (1997). On-line retorts control in thermal sterilization of canned foods. Food Control, 8 (1), 13–20. Teixeira, A., Dixon, J., Zahradnik, J. and Zinsmeiter, G. (1969). Computer optimization of nutrient retention in the thermal processing of conduction-heated foods. Food Technology, 23(6), 845–850. Teixeira, A. A., Zinsmeister, G.E. and Zahradnik, J.W. (1975). Computer simulation of variable retort control and container geometry as a possible means of improving thiamine retention in thermally-processed foods. Journal of Food Science, 40(3), 656–659.

© Woodhead Publishing Limited, 2013

19 Automation for a sustainable food industry: computer aided analysis and control engineering methods A. I. Papadopoulos, Centre for Research and Technology – Hellas, Greece and P. Seferlis, Aristotle University of Thessaloniki, Greece and Centre for Research and Technology – Hellas, Greece

DOI: 10.1533/9780857095763.2.441 Abstract: This chapter discusses automation in the form of computer-aided analysis and control engineering methods that facilitate sustainable development in the food industry. The concepts of sustainability and automation are analyzed in the context of food manufacturing processes. A number of automated tools, methods and technologies that facilitate automated design and control in the broader chemical process industry are reviewed and discussed with respect to their ability to incorporate sustainability considerations. A review of the scientific developments associated with automated design and control in food manufacturing is also performed to evaluate the assimilation of the presented tools, methods and technologies. It is concluded that food manufacturing processes provide significant opportunities for further integration of automated tools, methods and technologies that advocate sustainable development. Key words: sustainability, automated design, food processing, optimization, process synthesis, automatic control, computer-aided molecular design, process integration, life cycle assessment.

19.1 Introduction Traditionally the development of the food industry has been guided by decisions that revolved around economic growth with the aim of achieving maximum financial returns. This model of development has been fueled by the predominant notion of a linear economy (Garcia-Serna et al., 2007), where intensified methods are used to exploit natural resources and to manufacture edible products, while all the unfit for consumption wastes are simply disposed of out of sight. In the

© Woodhead Publishing Limited, 2013

442 Robotics and automation in the food industry past three decades it has become increasingly clear that economic growth in the food industry cannot be sustained based on such a notion because it inherently involves two major flaws: it is assumed that raw materials can be replenished at any desired rate by nature, and that the ecosystems can absorb inexhaustible amounts of waste. Obviously, such assumptions are far from reality as the capacity of nature to provide resources and absorb pollution is finite. Nowadays, this is a measurable fact as numerous organizations monitor and register the environmental impact of human activities worldwide. In this respect, the modern food industry is addressing the challenge of a new growth model, which is based on the constraint of sustainability through less intense utilization of natural resources and increased recycling of wastes. Practically, this means ‘getting more from less’. Improved, low cost products and services are expected through efficient processes with decreased environmental impact. An important question is therefore raised in the context of such developments: how can the food industry remain competitive and viable in the future, when significant new restrictions are imposed on the management of the two major ingredients, namely natural resources and wastes, which enabled its mono-dimensional economic growth in the past? The need to enable a swift and smooth adaptation of the food industry to the model of sustainable development requires a shift towards highly efficient manufacturing processes that accomplish prudent utilization of natural resources (e.g., energy carriers, water, and raw materials) and zero waste policies. Automation involves a collection of key methods and technologies that enable the levels of efficiency required to realize sustainable food manufacturing. The proliferation of computers and modern electronics has allowed the widespread utilization of automation that transformed industrial operations in the past thirty years. Various industrial sectors have recently adopted automation both in terms of computer-aided analysis methods (e.g., process integration, life cycle product and process assessment, eco-efficiency, etc.) and control engineering methods (e.g., supervisory control and data acquisition, distributed control systems, etc.) with the aim of endorsing a firm regulation of raw materials and energy utilization, emissions and environmental impact. The food industry is required to follow similar practices and support the utilization of automation to contribute to sustainability. The processes involved in food manufacturing require high energy consumption for freezing, chilling, heating and sterilizing, which mostly originates from fossil fuels and is rarely recovered for further utilization. Water is also used extensively for cooling and sanitization purposes, as well as an ingredient, hence efficient recycling and purification are required to avoid undesired misuse. Packaging can be a major source of pollution on top of other solid or liquid wastes that need to be treated. The adoption of automation in food manufacturing can have a significant impact in addressing these challenges. Section 19.1 presents major challenges that need to be addressed by the food industry in the transition from economic growth to sustainable development. Section 19.2 provides formal definitions of sustainability and identifies the links that associate the food industry with sustainable development. Section 19.3 argues that automation can have a significant impact in the context of sustainable food

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 443 industries, as it provides the key methods and technologies that enable the levels of efficiency required to realize sustainable food manufacturing. Section 19.4 presents a collection of essential automated tools utilized for the design and automatic control of sustainable food manufacturing processes. The considered tools represent generic frameworks that enable automated design and control in the process industry and involve process and product optimization, process integration, realtime process optimization and control, and life cycle assessment. Case studies of applications in the food industry are used to illustrate their benefits. Section 19.5 presents advanced tools for automated and sustainable design and operation of food manufacturing systems, namely process and product synthesis, integration and dynamic operability and computer-aided molecular design. These tools have been developed in recent years based on the tools presented in Section 19.4 and have found several applications in the food industry. Their evolution into generic and systematic methodologies is described in the context of chemical process examples as their ability to capture economic, environmental and societal dimensions in increased detail shows great potential to address highly challenging problems in the food industry. Section 19.6 presents technologies in the form of commercially available software packages that implement several of the considered tools and methods. Section 19.7 presents the conclusions and identifies future trends.

19.2 Definition of sustainability and links with the food industry In general, sustainability or sustainable development refers to the realization of economic growth, which simultaneously takes into account environmental protection and social viability. Sustainable development involves the exploitation of natural resources with a rate that is lower than or equal to the rate that they can be replenished in the future in order to avoid environmental degradation and enable high living standards. Among numerous definitions of the terms ‘sustainable development’ and ‘sustainability’ that are available in published literature, two of them stand out as being easily comprehensible but also of inclusive context. The first one is by Brundtland (1987) and reads (as quoted in Garcia-Serna et al., 2007): ‘Sustainable development is development which meets the needs of the present without compromising the ability of the future generations to meet their own needs.’ The second is by Liverman et al. (1988): ‘Sustainability involves the indefinite survival of human species (with a quality of life beyond mere biological survival) through the maintenance of basic life support systems (air, waste, land, and biota) and the existence of infrastructure and institutions which distribute and protect the components of these systems.’ Within the context of these definitions it appears that sustainability involves interplay between economy and environment, but also involves the impact of this interplay on society. These three dimensions of sustainability can not be treated in isolation. As noted in Sustainable Measures (2010) the economy exists because of the interactions between humans that are organized within societies. In turn, societies exist within the environment

© Woodhead Publishing Limited, 2013

444 Robotics and automation in the food industry that provides all the necessary raw materials to support them (i.e., food, water, air, energy). In this respect sustainable development should ensure that all three dimensions of sustainability evolve simultaneously in a way that no dimension dominates over another. This can be achieved only by accounting for their interactions in all human activities. Taking into account the definition of sustainability, the growth model of the food industry, discussed in Introduction, can now be addressed within a broader context. In the past, economic growth was the sole concern of the food industry. A major positive effect of this growth is the continuous creation of new jobs, which generally enables reduction of poverty and improvement of personal living standards. However, the lack of attention to the environmental impacts has resulted in increased production of untreated wastes, which pollute the environment. Environmental pollution, considered in conjunction with the contribution of all other industrial sectors that follow a similar growth model, has significant implications, such as deterioration of the air quality and the food and water reserves, while resulting in poor nutrition, health problems and low living quality, among other things. Evidently, such implications obstruct the progress and evolution of societies, when they are pursued solely for economic growth. The food industry is a typical example of the strong links that exist between economy, environment and society. A report published by the UK’s Department for Environment Food and Rural Affairs (DEFRA, 2006) highlights several significant links between the food and beverage industry (beyond the farm gate) to the three dimensions of sustainability. In particular, the food industry is: • a major energy user, as it accounts for approximately 14% of energy consumption by UK businesses, • a major contributor to carbon emissions, as it produces approximately 7 million tones of carbon per year, • a major water consumer, as it consumes approximately 10% of the water utilized by the entire UK industry, • a major source of waste generation, as it accounts for about 10% of the industrial and commercial waste stream in the UK, which is largely associated with packaging in the case of the food industry, • a significant contributor to impacts associated with transportation, as it accounts for 25% of all heavy goods vehicle mileage in the UK, • a significant contributor to the provision and dissemination of healthy food choices to consumers, • a major employer, as it is associated with 12.5% of the UK’s workforce. While the numbers may vary for different countries, it is clear that the food industry is extensively affecting significant and sensitive sectors of enviro-socio-economic activities. In this respect, it has the potential to become a major contributor to sustainable development. From an environmental standpoint it should operate within natural limits by minimizing the consumption of resources (energy, water, raw materials) and maximizing the recycling of wastes. Its contribution to society

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 445 should focus on the production of safe, healthy and nutritional products within a hygienic and humane working environment that adds to the welfare of the employees. Based on such principles economic growth will be sought concurrently with societal development and environmental preservation and not at their expense.

19.3 Automation and sustainability in food manufacturing The term ‘food industry’ encompasses a broad range of activities that revolve around the production of food and may involve the entire chain from the cultivation of raw materials (agriculture) to the retailers and the food service sector. Although the contribution to sustainability of all such stakeholders can be significant, in practice it is the food manufacturing processes that carry the weight of social, environmental and economic impacts on the entire chain. In this respect, the firm regulation of energy consumption, water utilization, waste and emissions generation, as well as of raw material and product prices in food manufacturing, is expected to be a major measure of sustainable development. In terms of food manufacturing, firm regulation translates to the development and operation of efficient processes that: 1. Enable cautious management and utilization of resources. 2. Achieve significant reduction and recycling of wastes. 3. Maintain the production of high quality and widely affordable products. Automation is the key to achieving high efficiency within the context of sustainable development in food manufacturing. It involves appropriate technologies and methods that enable the targeted interventions required in food manufacturing processes to approach the previously presented goals. Automatic tools and procedures are used to replace human operators in tasks that require high accuracy, reliable repetition, increased safety and hygienic precautions. This results to numerous significant benefits, which are summarized as follows: 1. Productivity is increased as operations are performed at a maximum possible rate. 2. Operating economics are tightly regulated as all costs are accounted for with high accuracy. 3. Product quality is improved substantially as the error margins are practically eliminated. 4. Humans remain protected from hazardous conditions as only mechanical equipment is used to perform unsafe or unhygienic tasks. 5. The required workforce is turned from one that performs hard physical or monotonous work to one that is fitted with the specialized skills required to manage the associated automation equipment. Such benefits mostly concern process operations, which were considered as the main target for the application of automation technologies in the past. With the recent wide proliferation of computing and information technologies it soon

© Woodhead Publishing Limited, 2013

446 Robotics and automation in the food industry became clear that the extension of automatic technologies to the field of process design and development can also generate significant benefits. This is because modern requirements such as sustainability, intensification, flexibility and globalization, to name a few, have added complex constraints to the design of industrial operations that overcome the human capabilities to perform calculations. The design of products and processes imposes significant challenges that require the utilization of systematic computer-aided methodologies and tools. A typical example of such requirements lies with the rather complex task of identification and minimization of the environmental and economical impacts caused by the energy requirements of a process. It involves a vast number of decisions in the form of choices, constraints and objectives, which can only be addressed through the use of systematic and automated methods such as process integration or process optimization, in order to identify the optimum trade-offs among the economic, environmental and operating performance of the process. Similar design complexities are involved in the requirement for safer or less hazardous chemicals, which can be effectively addressed by computer-aided molecular design methods. These are only a few examples of the numerous available automated tools, methods and technologies discussed in the following section that can be utilized to replace human intervention in sustainable product and process design.

19.4 Tools for automated sustainable design and operation in food engineering This section presents an overview of major automated tools that can be used in food manufacturing within the context of sustainability. They include a set of generic frameworks that are broadly utilized in the process industry with specific application case studies in the food industry. Several of these tools support the development of sustainability metrics and indicators, whereas others facilitate their incorporation during process/product design and control.

19.4.1 Process and product optimization Product and process optimization represents a set of technologies enabling the systemic development and evaluation of complex decision making problems. Scope and description The design of products and processes involves numerous decisions that need to be considered prior to their development in order to meet desired specifications that are often associated with physical/chemical/mechanical characteristics of products and processes or with constraints representing economic, environmental and operating impacts. Such constraints are often imposed by legislation specifying and enforcing acceptable limits to pollution or quality in foods. Process operations commonly used in the food industry, such as thermal processing, drying, contact cooking and microwave heating, are typical cases where optimization-based

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 447 methods can generally be used to minimize energy utilization or processing time below a particular temperature while meeting product quality or safety specifications (Banga et al., 2003). For example, in thermal processing the goal is to identify the optimum temporal temperature profile that maximizes the nutrient retention in a pre-packaged food item while satisfying micro-biological lethality constraints (Banga et al., 2003). Conventional approaches address the decision making involved by combining trial and error with engineering know-how in efforts that incorporate intense human intervention. Although this often enables the identification of product or process features that are feasible within the particular specifications, optimality is not guaranteed. Product and process optimization represents a set of technologies enabling the systemic evaluation of complex decision making problems. This section presents an inclusive overview of such technologies in the context of food engineering. Optimization technologies provide the automation required to enable a systematic identification and evaluation of process and product features, which further lead to optimum performance. They employ robust formulations of the addressed design problem, which involve systematic and strict mathematical mechanisms, hence prohibiting the admittance of subjective interpretations in the design solutions. This is achieved by utilizing mathematical models to enable a realistic representation of the real-world behavior of processes and products. Such models are adapted to optimization algorithms that evaluate the performance of the investigated process or product system based on appropriate performance indicators. Figure 19.1 summarizes major algorithmic features involved in the optimization sequence and indicates the points where sustainability considerations can be introduced into the sequence. Clearly, the optimization algorithm is only part of the optimization sequence. Important ingredients to enable the formulation and solution of a design optimization problem involve the following: (a) the development of a model, (b) the identification and selection of decision parameters, (c) the determination of a performance indicator and constraints and (d) the selection of an optimization algorithm. Model development Models are formal mathematical transcriptions of the characteristics that determine the behavior of a system. In general, models are used to capture the intake and output of mass, energy and/or information in a system as well as the physical, chemical, mechanical, electrical and other phenomena that determine the way that mass, energy and/or information are processed by the system. The purpose for which the models are utilized determines the assumptions that are considered during their development and their ability to represent the considered process or product system sufficiently. • Simplified models capture major system characteristics by transforming engineering know-how, heuristics and empirical or approximate observations into mathematical relations. Such models are usually represented by simple algebraic equations.

© Woodhead Publishing Limited, 2013

448 Robotics and automation in the food industry

Set initial state

Use system model to simulate state Calculate the employed performance indicator and constraints •Economic, environmental, operating, safety indicators and constraints

•Process and product models •Physical, chemical, environmental, safety characteristics

Evaluate performance

Generate new state Optimization algorithm

Fig. 19.1 Major algorithmic features involved in optimization.

• Rigorous models reflect the behavior of real-life systems with high detail as they use complex and analytical mathematical representations in the form of non-linear algebraic or differential equations. Figure 19.1 indicates that models are used in iterative simulations within the optimization sequence. Numerous simulations are often required to identify the optimum solution hence the type of utilized model (simplified or rigorous) affects the computational effort involved in optimization. • Simplified models enable a fast screening of the desired or available design and operating options for a considered system, but the obtained solutions are expected to require further refinement prior to utilization in practical applications. • Rigorous models facilitate the identification of realistic design features at the expense of increased computational effort. This trade-off between computational efficiency and immediate applicability of the obtained design solutions should always be considered during the development of a model. In the food industry there is generally a need to use rigorous models in equipment design due to the very strict safety and quality constraints that require a highly detailed emulation of the underlying phenomena. On the other hand, simplified equipment models are often used to address food manufacturing problems involving complex planning and scheduling operations. In these cases the chemical or physical performance characteristics of equipment are often pre-determined as the investigation of inventory and stock sizes as well as temporal delivery specifications are more important. Georgiadis et al. (1998) presented research work developing highly rigorous plate heat exchanger models to address milk fouling costing the fluid milk industry approximately $140 millions per year (at

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 449 1998 prices) due to the need for very frequent cleaning. The use of such models allowed them to identify appropriate heat exchanger arrangements that could have a beneficial impact in milk fouling mitigation. Kopanos et al. (2010) studied the optimum scheduling and lot-sizing of yogurt production plants. Production management in such plants is rather challenging: building inventories is not effective as yogurt is perishable and changes in customer orders are often made just before dispatch. The authors of this work developed comprehensive models consisting of practical constraints and capturing a large number of interactions among production lines operating in parallel and involving multipurpose equipment. Such models enabled them to identify scheduling configurations that resulted in up to 20% reduction in the inventory costs. Decision parameters The mode of behavior observed for a process system or a product signifies the system state at a particular time instance. In the optimization sequence of Fig. 19.1, system states are represented through the values of the parameters utilized in the model equations. In this respect, the different states that are potentially obtainable by a process system or a product are emulated by deliberately assigning different values to the model parameters in each optimization iteration. As a result, important process and product operating parameters (e.g., stream temperature, pressure, flowrate, and so forth) are varied during optimization in order to influence the design decisions and to identify solutions for optimum operating performance. Such parameters are often combined with discrete or continuous design characteristics of the investigated systems (e.g., size, type, number, shape of vessels, and so forth) to enable the identification of both structural and operating features that lead to optimum performance. The choice of the decision parameters and the variation range employed during optimization for each parameter generates a trade-off between computational performance and quality of the optimum solution. The consideration of large numbers of decision parameters that represent a broad optimization search space is desirable in order to identify a truly optimum solution, but this often requires increased computational effort due to the numerous combinations that need to be examined. Fewer decision parameters obviously reduce the computational effort involved, but the exploration of a limited search space results in small improvements in terms of the obtained design features. Performance indicators and constraints Performance indicators, known in optimization as objective functions, are used to evaluate the performance of a selected set of decision options. Objective functions are either minimized or maximized and involve mathematical terms that are affected by the decision parameters. Cost is often the objective function of choice for most systems; however, environmental impacts, energy utilization and physical or chemical properties of products are also quite common, as they are directly linked to driving forces that affect the performance of a system. In many cases it is necessary to simultaneously evaluate more than one performance indicator. This is of particular interest, when the employed objective functions represent

© Woodhead Publishing Limited, 2013

450 Robotics and automation in the food industry conflicting choices. For example, the thermal processing problem might have conflicting objectives; when there is a need to maximize the retention of nutrients and other quality characteristics on the surface of a food, the processing time might also need to be minimized. In another case, when there is a need to minimize the micro-biological load, the nutrient retention might also need to be maximized (Banga et al., 2003). In such cases, modifications to the optimization algorithms are employed that result in multi-objective optimization algorithms in order to evaluate the trade-offs generated among the conflicting objective functions. Constraints are used in optimization in order to emulate the upper and lower limits often imposed in systems design due to performance requirements, environmental regulations and resources availability. In this respect, constraints are useful as they specify whether a solution obtained through optimization is feasible or infeasible. In the case of identification of design solutions that are infeasible, several features of the design problem must be re-defined in order to meet the employed constraints. This is particularly useful, as constraints can be used in order to identify design limits or performance targets of the considered system. Optimization algorithms Optimization algorithms incorporate mathematical mechanisms that enable the evaluation of the system performance in the current state, and subsequently lead the optimization search towards a new promising direction as a result of this evaluation. Deterministic and stochastic optimization algorithms represent the two major algorithmic categories. Deterministic algorithms make use mostly of topological or geometrical methods to address optimization problems and generally involve linear programming (LP), non-linear programming (NLP), mixed integer linear programming (MILP) and mixed integer non-linear programming (MINLP) approaches. In the design of process systems and products the NLP and MINLP approaches are the most relevant, as non-linear models are required to enable a detailed mathematical representation of their complex functionalities, operating and design features. Such approaches enable the identification of strict mathematical optima that are useful in order to identify design solutions that minimize important performance indicators such as cost, environmental impacts and so forth. However, they also involve shortcomings that are associated with non-convexities and discontinuities often observed in process and product design. In such cases deterministic algorithms are likely to converge to locally optimal solutions, they can not be easily used during the initialization of the design problem, while they make use of derivative transformations that increase the computational effort. On the other hand, stochastic optimization algorithms avoid such complications due to the fact that they mimic natural phenomena using simple mechanisms. They are called stochastic because they use a random probabilistic search method to explore the design features that optimize an objective function. There are numerous stochastic optimization algorithms available such as Simulated Annealing (SA), Genetic Algorithms (GA) and Ant Colony Algorithms, to name a few. Such algorithms have been inspired by physical phenomena (e.g., annealing

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 451 of metals) or processes observed in nature (e.g., natural selection) and feature several advantages. Although they do not guarantee global optimality, they result in final solutions very close to the global optimum. They also achieve significant savings in computational effort by employing mechanisms to enable a selective search of a much smaller fraction of the solution space for problems involving large numbers of variables. Remarks In this context, the previously discussed sustainability dimensions can be used in the course of optimization, provided that appropriate indicators or constraints are available for their mathematical representation in the optimization formulation. The available optimization algorithms involve mechanisms in the form of a systematic set of actions performed to address all possible outcomes generated from the interplay among the employed decision variables, process and product models and performance measures. In this respect, optimization algorithms are all available in the form of automated, computerized procedures. As a result, process and product optimization can be utilized off-line to enable optimum process and product design, or it can be incorporated online to enable optimum control. Case studies The problems of nutrient retention and thermal processing are reported as representative examples of the important merits resulting from the implementation of process optimization in the food industry. Nutrients retention during food processing Banga and Singh (1994) investigated the retention of ascorbic acid during air drying of potato disks formulated as an optimization problem consisting of all the previously reported ingredients: • The employed model was based on a thin slab of cellulose and involved rigorous mass and heat transfer equations to emulate the moisture transfer within the solid, reaction kinetics of degradation of the ascorbic acid and enzymes as well as equations to calculate the energy transmitted to the solid and the energy incorporated into the air. • The considered decision parameters involved the relative humidity of air and the dry bulb temperature subject to desired upper and lower limits. • The authors investigated various cases using as performance measures the nutrient or enzyme retention, the process time and the energy efficiency, while upper and lower limits were imposed as constraints in all these quantities including the average moisture content. • The optimization algorithm chosen was the Integrated Controlled Random Search for Dynamic Systems stochastic algorithm, which is based on a computationally efficient NLP algorithm. The performed optimization resulted in an optimal variable air temperature profile suggesting that multiple drying chambers operating under optimum policies

© Woodhead Publishing Limited, 2013

452 Robotics and automation in the food industry could lead to better products with up to 40% energy savings compared with the traditional constant air temperature profile process. Thermal processing and sterilization Balsa-Canto et al. (2002) used optimization for the design of food thermal sterilization processes. They studied two types of problems. The first, coined as the classical thermal process design, involves the calculation of the processing time at a pre-specified constant heating temperature to ensure a certain amount of minimum microbial lethality at a final time at the end of a cooling period. The second, coined as the thermal sterilization problem, requires the identification of a heating temperature profile that enables a pre-specified microbial lethality while maximizing a performance target such as the final nutritional quality of the product. From an optimization perspective both problems are challenging because there is a need to utilize heat transfer models emulating geometrical characteristics of pre-packaged conduction-heated foods often solved in three dimensions. To address this challenge the authors develop reduced-order models which enable a very detailed representation of the physical characteristics of the real-world problem (i.e., heat flow though the food container) but result in a reported speed-up of 15–20 times compared to using conventional models. This is very important from a practical perspective. For example, in case of an unexpected temperature drop during operation, an online optimization system receiving real-time information can recompute the optimum temperature profile in order to maintain microbial lethality or food nutrient quality within acceptable limits. These computations need to take place as fast as possible to minimize product losses.

19.4.2 Process integration Process integration is an approach that considers the unit operations included in a process flowsheet in a holistic manner with the aim of identifying interactions that lead to efficient exploitation of resources and minimization of the associated costs. Scope and description The Pinch technology (Linnhoff and Flower, 1978) is the tool mostly associated with process integration. This technology enables the identification of thermodynamically attainable energy targets and of structural changes required to achieve them in a process flowsheet. It has been extensively used as a method to recover and utilize wasted energy in process systems. Implementation of process integration methods in the food industry may have significant impacts as it involves numerous energy demanding operations. For example, Kemp (2005) reports that drying accounts for up to 20% of industrial energy demand in the UK and Europe, while it has often a thermal efficiency of below 50%. This indicates the very high energy intensity of drying operations but also implies that large amounts of potentially recoverable energy are currently wasted. Krokida and Bisharat (2004)

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 453 Heating load

Hot composite curve Temperature

ΔTmin Pinch

Cold composite curve Cooling load Enthalpy

Fig. 19.2 Composite curves and pinch in a temperature–enthalpy diagram. ΔTmin denotes the minimum design temperature of two streams in a heat exchanger.

calculated that 40% of the wasted exhaust heat can be recovered from a convective dryer by appropriate placement of a heat exchanger and a heat pump. The Pinch method The implementation of this method requires information about the temperatures and flowrates of all hot and cold streams in a process. Such information is plotted in a temperature–enthalpy diagram (Fig. 19.2) to give the hot and cold composite curves. The hot and cold curves represent the total heat available and the total heat required, respectively. The difference between the curves at each point of the temperature axis represents the temperature difference available to be used for heat transfer. Take for example the case of dryers used in the food industry. Major hot streams involve the outlet exhaust gas, including both sensible and latent heat, and the hot solids with a sensible cooling load. Cold streams on the other hand involve the incoming drying air that needs to be pre-heated to the dryer inlet temperature in convective dryers, or the heat supply to the jacket in contact dryers including the incoming solids that need to be pre-heated (Kemp, 2005). Temperature and heat load information regarding such streams result in the dryer composite curves. If the hot composite curve is moved to the right without touching the cold curve, then all the heating or cooling requirements can be satisfied from within the process. This is because the two curves overlap and there is sufficient temperature difference to enable heat exchange. In practice, there is one point of minimum temperature difference between the two curves, namely the pinch point, where the hot curve can not be moved further to the right. This point determines the non-overlapping area between the two curves. This area represents the heating and cooling loads that must be provided from external resources. In the case of a dryer, the range over which the hot and cold streams overlap is where heat can be recovered from the

© Woodhead Publishing Limited, 2013

454 Robotics and automation in the food industry hot exhaust stream to the cold inlet air (Kemp, 2005). Figure 19.2 shows that above the pinch the cold curve extends beyond the hot curve hence external heating is required. Below the pinch, the hot curve extends beyond the cold curve hence external cooling is required. Heat may be transferred between the streams above or below the pinch, but not across it. This is because in a cross-pinch heat transfer the heat deficit above the pinch and the heat excess below the pinch will both increase, hence requiring additional external heating and cooling. Extensions The Pinch method has been extended to address other issues related to energy recovery, other than energy targeting. The composite curves provide information regarding the overall energy targets, but they do not specify what utilities to use and how to place them. For example, often external heating is provided by both high pressure and low pressure steam. The goal is to minimize the use of the expensive utilities. This is done by the grand composite curve, which indicates the energy deficit and surplus of the considered process as a function of temperature and enables a quick screening of feasible options for the cold and hot utility in order to satisfy the process requirements (Townsend and Linnhoff, 1983). Using the grand composite curves of a drying process Kemp (2005) identified the temperature and heat load characteristics at which different utilities such as steam, solar heating or gas turbine combined heat and power (CHP) systems could be fitted in dryers. Furthermore, it is also possible to set targets for the minimum number of heat exchange units required in a process, hence linking capital costs (e.g. heat exchange surface) with operating costs (e.g. energy consumption). This can be combined with modifications in the operation of the considered process with the aim to further reduce the energy requirements. The plus–minus principle involves a set of simple rules that enable the identification of beneficial modifications. All these extensions can be used during the analysis stage in order to explore numerous options for process improvement. The results of the analysis represent targeted improvements that need to be implemented in a subsequent stage, which involves the design of a heat exchange flowsheet that considers such improvements. The Pinch method provides a set of rules and guidelines that enable the design of heat exchange flowsheets. Alternatively, process synthesis and optimization methods can be utilized in order to identify structural process modifications of optimum performance. Remarks Although in cases of energy integration Pinch analysis utilizes temperature as the main driving force, it has also found applications in the integration of water-using activities within a plant with the aim of reducing water consumption and waste water generation (Wang and Smith, 1995); more recently, applications for hydrogen generation and consumption savings have been reported (Hallale et al., 2003). Hydrogen in particular is expected to affect the food industry in several ways. First of all it is an environmentally benign fuel intended for extensive industrial use in the not so distant future. Furthermore, applications directly involving the wastes from the food sector include the transformation of used cooking oils into biodiesel

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 455 through treatment with pure hydrogen (Bezergianni et al., 2011). As such, the Pinch technology is an approach that inherently aims to improve sustainability in process systems. Its wide success is due to its simple use, the relatively few data required for application, and its intuitive graphical representation. Such benefits are accompanied by significant savings of up to 35%, 40% and 20% in energy, water, and hydrogen consumption, respectively (Natural Resources Canada, 2003). As a result, the technology has been incorporated in numerous computer-aided software packages to date. Case studies The problems of heat recovery in the milk industry and water recovery in citrus plants are reported as representative examples of the important merits resulting from the implementation of process integration in the food industry. Heat recovery from milk powder spray dryer exhausts Atkins et al. (2011) investigated the thermal integration of milk powder spray dryers using Pinch related analysis to match several different heat sinks to the dryer exhaust stream. The investigated plant consists of the following major processes: • A multi-effect evaporator train is used to concentrate milk at 52% using mechanical vapor recompression and thermal vapor recompression. • The concentrated milk is heated to the dryer feed temperature and sprayed in the drying chamber. • The so called cow water removed in the evaporators is used to partially pre-heat the incoming evaporator stream. • The dryer air is taken from within the facility and heated to 200°C at the dryer inlet. • A set of three fluidized beds, two cyclones and a bag house are situated after the dryer and the exhaust air from the dryer and the fluidized beds are combined before they exit the bag house at 75°C. • Site hot water is also used around the site for cleaning-in-place operations. The cold streams of the process involve the raw milk, the milk concentrate, the inlet air in the dryer and the fluidized beds and the site hot water. In these streams the temperature needs to be raised from a minimum of 10°C (raw milk) to a maximum of 200°C (dryer inlet air). The hot streams involve the cow water, the thermal vapor recompression, and the dryer air exhaust, where the temperature drops from a maximum of 75°C (air exhaust) to a minimum of 20°C (cow water, air exhaust). Using the temperature information together with the mass flowrates and thermal capacities of the streams the authors develop the hot–cold composite curves and the grand composite curve to identify hot and cold utility targets of approximately 25 and 15 kW, respectively, as well as to investigate efficient ways for heat recovery. The authors consider the matching dryer exhaust stream with the following sinks:

© Woodhead Publishing Limited, 2013

456 Robotics and automation in the food industry • The dryer inlet air: In this case it is found that a 12.8% decrease in the hot utility requirement is possible, while this value can improve further by using multi-pass heat exchangers. Furthermore, it is shown that the heat transfer will increase by using smaller diameter tubes. • The site hot water: This stream presents the advantage that it has a lower supply temperature than the inlet air stream, enabling a larger temperature difference, hence more heat is recoverable. Furthermore, plate heat exchangers can be used in this case, which are considerably less expensive than fin tube heat exchangers used in the case of the air inlet stream. The hot utility savings are reported to be in the range of 15%. • Heating of the raw milk: In this case the cow water otherwise used for raw milk heating can now be used to pre-heat first the inlet air and then the site hot water. Despite the high amount of heat recovered in the exhaust the load on the raw milk heater is increased, hence this case is practically similar to the case of the dryer inlet air. A more efficient variation of this case would be to use the cow water to first pre-heat the inlet air and then the raw milk. This option results in a 21% decrease in hot utility, while the cow water needs no further cooling hence avoiding the further use of a cooling tower. • Splitting of the exhaust stream: In this case two splits can be matched to the site hot water and the inlet air increasing heat recovery and hot utility savings to approximately 20%. Furthermore, this configuration is quite simple to implement, but the use of two heat exchangers would increase capital expenditure. Water recovery and re-use opportunities in a citrus plant: the water Pinch analysis Thevendiraraj et al. (2003) perform a water Pinch analysis in a citrus plant in order to identify opportunities for reductions in the utilized water and wastewater streams. As the plant produces concentrated juice and the raw fruit contains 90% water, a large portion of contaminated water is lost to the wastewater network. To address this issue the authors employ the water Pinch analysis to maximize water re-utilization and to identify regeneration opportunities. In this respect, all streams and process operations utilizing water are represented in the form of inlet and outlet contaminant concentrations to develop purity profiles for the entire process. Such concentrations are associated with equipment corrosion, fouling and process constraints in the mass transfer and flowrates of water. All streams are classified as water sources and sinks. Based on this information the composite curves can be constructed in a diagram indicating water quantity (flowrate) in the X-axis and water quality (purity) in the Y-axis (Fig. 19.3). Similarly to the heat pinch, overlapping of curves indicates opportunities for water re-utilization limited by the pinch point where the sink and source curves are in contact. Nonoverlapping areas in the sink and source composite indicate fresh water requirements and waste water generation, respectively. Reductions of the plant freshwater and waste water treatment are accomplished by mixing source streams while meeting requirements of the corresponding sources. This way the pinch point is

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 457

Purity of effluents

Pure water

Pinch

Impure effluents for treatment

Reutilization opportunity

Flowrate

Fig. 19.3 Composite curves in water Pinch analysis.

recalculated in an iterative procedure until the minimum freshwater requirement and waste water generation is identified. These targets can be identified for fixed process conditions, by increasing the concentration limits in selected sinks to identify trade-offs between contamination and water-re-utilization, as well as by regeneration of several waste water streams prior to re-utilization. The first two options usually involve piping modifications, while the last might involve capital costs for treatment facilities (Natural Resources Canada, 2003). Using the above analysis, Thevendiraraj et al. (2003) investigated several different configurations of the water utilization network in the citrus plant. They found that appropriate redistribution and utilization of the available water results in a 22% freshwater/wastewater reduction. This can be accomplished by fitting only five new pipes in the existing network, which implies lower plant capital expenditures. Furthermore, they found that an additional 8% reduction in the freshwater consumption is possible by regenerating the available wastewater. In this case the impact on capital expenditure is more significant as it requires a regeneration plant.

19.4.3 Real-time optimization and control Real-time (online) optimization is a powerful tool for the preservation of high economic benefits and improved food processing within a perpetually changing environment (Banga et al., 2003). Scope and description Raw material quality variation, equipment limitations due to maintenance work, utility limitations, equipment configuration changes, equipment performance variation, market demand fluctuations, and product specifications variations are some factors that affect the overall economic performance of the plant and its performance against sustainability indices associated with energy and water utilization or waste production. The frequency of variations may be hourly in utility

© Woodhead Publishing Limited, 2013

458 Robotics and automation in the food industry System objectives

Scheduling system

Model based optimization

Scheduling commands

Supervisory-optimization control system Controller set points

Regulatory controllers Control actions

Sensors

Actuators

Measurements

Valve positions

Unit operations

Fig. 19.4 Schematic of the automated control system based on real-time optimization.

limitations (e.g., steam and cooling water system availability), daily in equipment availability, feedstock quality, weekly in market properties (e.g., product demand or product prices) and product specifications (e.g., product grades based on scheduling plans) and monthly in equipment performance (e.g., heat exchanger fouling). Real-time optimization accounts for all possible changes that affect the plant and determines the optimal operating conditions that would maintain a high level of profitability and sustainability. The optimal operating points are directed to the regulatory feedback control system as setpoints to the controllers that subsequently aim to move the plant towards these targets. The plant sensors provide feedback to the supervisory-optimization control system regarding the current status of the plant. The supervisory control system then evaluates overall plant performance after the plant has reached a steady-state operation. Therefore, real-time optimization is viewed as part of the plant control system. The supervisory-optimization control system encompasses all the features described in the previous sections regarding model-based optimization and receives information from the production scheduling level of decisions regarding the production plan of the plant. The complete structure of the real-time optimization framework in food processing is shown in Fig. 19.4. Within the objectives of the supervisory-optimization control system the plant sustainability targets can be explicitly considered in the objective function of the optimization problem formulation. Furthermore, the current status of the performance of several subsystems (e.g., wastewater treatment plant, steam generation plant, individual thermal systems such as evaporation stations and pasteurization units) in a typical food processing plant are monitored through the transfer of online measurements to the supervisory-optimization control system. The holistic approach achieved with model-based optimization enables the optimal distribution of the available resources to the various plant subsystems so that the overall

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 459

Set point dynamics

+ e(t) –

Controller dynamics

d(t)

ysp(t)

m(t) Actuator (value) u(t) dynamics

ym(t)

Process disturbance dynamics

Process dynamics

++

y (t)

Sensor dynamics

Fig. 19.5 Schematic of the regulatory control system. Symbols denotes the signals for the controlled, y(t), measured, ym(t), manipulated, u(t), controller output, m(t), error, e(t), setpoint, ysp(t), and disturbance, d(t), variables.

resource utilization is kept low while satisfying product specifications and safety, environmental, and operational requirements. The regulatory control system, shown in structure in Fig. 19.5, aims to drive the plant to the optimal operating point set by the supervisory-optimization control layer in a smooth, robust and fast fashion under the presence of multiple and simultaneous disturbances that constantly affect the plant operation. Several examples of feedback control techniques and configurations are provided by Seferlis and Voutetakis (2008). Assuming that the regulatory control system is capable of achieving the desired operating point, due to the fact that the operating point lies within the feasibility region of the actual plant and the control system itself is able to compensate for disturbances effectively, the implementation of a real-time optimization scheme is an effective solution. Case studies The problems of control in milk processing and thermal sterilization systems as well as real-time optimization of extrusion cooking processes are reported as representative examples of the important merits resulting from the implementation of real-time optimization and control methods in the food industry. Control of a milk pasteurization process The control of milk pasteurization has been studied by Negiz et al. (1998a, 1988b) and Morison (2005). According to the process flow diagram shown in Fig. 19.6 milk feed enters the regeneration section to absorb heat from the hot pasteurized milk stream and then enters the pasteurization section where it reaches the pasteurization temperature for a specified time period with the aid of process steam. Pasteurization time equals the time spent in the holding tube. The pasteurization temperature needs to be controlled tightly usually within ±0.5°C to avoid the buildup of bacteria which accurs if the temperature is lower than the target level, or the destruction of the milk nutrients which happens if the temperature exceeds the desired level. The main manipulated variable for the control of the pasteurization temperature T3 is the

© Woodhead Publishing Limited, 2013

460 Robotics and automation in the food industry

Bypassed milk TwC

Pasteurized milk, T5

T4 Steam Regeneration section

Milk feed, F1, T1 T1

Holding tube

Pasteurization section

T3 T2

TwY

T3,p T3C

T1Y

Fig. 19.6 Conceptual control diagram of a milk pasteurization process. Symbol T denotes process stream temperature and F process stream flowrate. T1Y, T3C, TwC, and TwY indicate controller elements whereas T1 indicates a measurement element.

steam flow. The effect of changes in the steam flow rate on the pasteurization temperature is quite slow and may involve significant dead time due to transportation time delays for the steam and the milk streams (i.e., temperature responds to steam flow changes only after a time period that is characteristic of the process system). A single feedback loop would therefore be very slow in rejecting disturbances in the steam pressure, resulting in large deviations of the pasteurization temperature from the set point – a totally unacceptable control performance. A secondary control loop that regulates the heating water temperature enables the prompt response of the system to disturbances in the steam pressure before such a change is observed in the milk temperature. The primary temperature loop takes care of all other possible sources of disturbances (e.g., heat exchanger fouling, feed milk temperature and flow rate) and provides the set point for the heating water temperature for the much faster secondary control loop. Both controllers are PID and are tuned sequentially starting with the PID in the primary loop that is responsible for the overall control performance. This is a typical form of a cascade control scheme. The control system can be further reinforced against disturbances in the milk feed temperature. Such changes can be compensated by the control system before influencing the pasteurization temperature with a feedforward-cascade control scheme. A measurement of the milk feed temperature is utilized with the aid of a model describing the dynamic relation between the measured disturbance (e.g., feed milk temperature) and the controlled variable (e.g., pasteurization temperature) to calculate the response of the control valve to counterbalance the upcoming disturbance. The signal from the feedforward controller is then added to the feedback signal from the primary loop in the cascade control scheme. Real-time optimization of an extrusion cooking process Food extruders present a challenging control problem, as the product quality attributes such as degree of gelatinization and bulk density are usually measured

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 461 in the laboratory using off-line sensors. Such a procedure introduces significant delays in the response of a typical control system to disturbances in the feedstock, resulting in a degraded product quality. Pomerleau et al. (2003) implemented a real-time optimization scheme for a twin screw extruder consisting of two layers. The lower layer uses conventional PID (proportional-integral-derivative) controllers (Seferlis and Voutetakis, 2008) to maintain the melt temperature and the pressure at the die near their setpoints. The screw speed and the feed rate are the manipulated variables. The upper control layer involves a first principles model that utilizes online, easy and inexpensive measurements to calculate the optimal setpoint values for the controlled variables of the first layer. The measured quantities act as inferential variables for the product quality variables. The performance of the control scheme is further enhanced using a model update step that utilizes online process measurements to calculated accurate estimates of the model parameters. The proposed approach is reported to be 10–100 times faster than conventional optimization-based control. Optimal control in thermal sterilization of canned foods Chalabi et al. (1999) incorporated optimal control in the thermal sterilization of canned foods. A detailed model for heat diffusion in the canned food was developed and validated using experimental data leading to accurate estimates for the heat diffusivity parameter. The optimal control scheme identified optimal retort temperature profiles so that the required sterilization conditions are achieved without product nutrients degrading, while utilizing the minimum energy requirements and minimizing the batch time.

19.4.4 Life cycle assessment Life cycle assessment (LCA) is presented as a method of addressing the environmental impacts of food products and processes throughout their entire life. Scope and description LCA is a method that enables the determination and quantification of environmental burdens and impacts caused by products and processes, in order to reveal weaknesses and identify opportunities for improvements. The determination and quantification of burdens is accomplished by assessment of all the energy and material flows associated with a process or a product, including the wastes that are disposed of into the environment. In this respect, the full life cycle of products is examined from cradle to grave. This means that the impacts of the production stage are considered together with all the impacts involved prior to and following this stage, namely the production of raw materials, product manufacturing, transportation, distribution, use, re-use, maintenance, recycling and disposal. Typical indicators often utilized to represent environmental impacts in LCA involve global warming potential, ozone depletion potential, acidification, human toxicity, ecotoxicity, smog and eutrophication.

© Woodhead Publishing Limited, 2013

462 Robotics and automation in the food industry Methodology LCA is implemented in four main phases, namely goal and scope, life cycle inventory, life cycle impact assessment, and interpretation. P1. Goal and scope. The determination of the goals and scope of LCA is a very important phase because the specifications, assumptions and simplifications considered at this point will have effects in the final result of the assessment. Specifications involve the depth of the assessment, the required quality of associated data, the selection of parameters and indicators that provide an efficient representation of the considered impacts, the determination of the functional unit and boundaries of the considered system. The determination of the functional unit is one of the most important issues in this phase. The functional unit is a quantitative reference function to which all flows in the LCA are related. The system boundaries determine the processes that need to be accounted for during LCA. To give an example of this phase we refer to Hospido et al. (2006) reporting on the implementation of life cycle analysis in canned-tuna processing. The employed functional unit is 1 ton of raw frozen tuna entering the factory, while the considered system is quite extensive. It starts from the harbor receiving the frozen tuna, it includes the processing factory, and it extends to the final distribution to the market (wholesale and retail) as well as the household use of the tuna. The factory itself is subdivided into major subsystems including different processing steps such as reception, thawing and cutting, cooking, manual cleaning, liquid dosage and filling, sterilization, quality control and packaging, assembly shop for cans and wastewater treatment plant. P2. Life cycle inventory. In this phase a flow model of the considered system is developed, while the material and energy flows are represented as input−output data for all the processes included within the boundaries of the considered system. The subsequent association of the individual processes represents the relations between the system and the environment. The mass/energy balances represent the inventory of the entire system. The environmental burdens of the system are then calculated, using the functional unit as a reference. In the canned-tuna case study detailed inventories are developed for all the previously considered processing steps including material and energy flows entering and leaving the subsystems. For example, inputs to the sterilization step includes materials in the form of tuna, coagulants, flocculants and water expressed in weight and volume units, electricity for the sterilization and recycling system expressed as energy units, thermal energy for the sterilization of the cans, bags and the thawing room expressed as energy units, transportation flows of materials used in this stage expressed as weight-distance units. Outputs from this step include the processed tuna expressed as weight units and the waste-to-treatment by-products including left-over fats and wastewater expressed in weight and volume units, respectively. It is worth noting that the calculation of all these flows requires a very thorough and consistent data acquisition system, which is based on historical data registered in the day-today operation of the processing facilities as well as from external databases

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 463 regarding materials and procedures utilized during processing. This shows the extent of the analysis involving resources and therefore impacts far beyond the system itself. P3. Life cycle impact assessment. The impact assessment involves determination of the impacts based on the available data. This includes classification of the data into appropriate impact categories, characterization of the impacts by use of equivalency factors multiplied with the inventory parameters included in each category and summation of all the parameters to obtain the result in each category. In the case study, the classification considers categories such as eutrophication, stratospheric ozone depletion, global warming, acidification, photo-oxidant formation and abiotic resources depletion. The characterization stage results in the calculation of the impacts which show that the tinplate production and transportation are responsible for approximately 60% and 55% respectively of the global warming and acidification potential associated with the processing of canned tuna. P4. Interpretation. The interpretation phase involves the analysis of the results, which should facilitate the identification of significant impacts. The conclusions resulting from the analysis should be evaluated in terms of their completeness, sensitivity and consistency, considering the assumptions and simplifications taken into account during the goal and scope phase. Further recommendations regarding actions to reduce the identified impacts should be made on the basis of consistent conclusions. Based on the resulting impact the recommendation for the canned-tuna factory involves changes in the packaging policies in order to reduce the impacts associated with tin cans. One policy involves an increase in the use of recycled tin while a different policy involves the use of plastic bags instead of tin. In both cases reductions of up to 50% are reported for the global warming and acidification potentials compared to the current practice. Remarks LCA is a systematic method that has been included in numerous software packages (Garcia-Serna, 2007); hence, it is currently performed automatically. A downside to LCA is that it addresses environmental impacts, which represent only one dimension of sustainability. The idea of Eco-efficiency, originally expressed by the World Business Council for Sustainable Development, has been recently used to advocate the integration of environmental, economic and social considerations in the context of sustainability. Research in this area has been lead by BASF, a multinational producer of chemicals, who proposed a systematic methodology for the implementation of Eco-efficiency analysis (Saling et al., 2005). The approach utilizes LCA as a first step to develop an ecological fingerprint of a product or a process. This is followed by an assessment of the economic aspects of the alternative products or processes, which is implemented through an overall cost calculation taking into account the flow of material and energy and including all relevant secondary processes. Recently, the proposed

© Woodhead Publishing Limited, 2013

464 Robotics and automation in the food industry approach has been extended to account for the societal dimension of sustainability in an analytical tool called SEEbalance®, which measures and visualizes the impact of each sustainability dimension (Schmidt et al., 2004). Case studies LCA has been the tool most commonly utilized in food processes in order to evaluate environmental impacts. Mattsson and Olsson (2001) note that energy use and energy related emissions represent the highest contributions to the environmental impacts associated with food processing. Heating, chilling and freezing processes involve high consumption of energy, with a large part of the emissions being due to the use of fossil fuels. In this context, a notable example regarding the use of LCA in the food industry is the work of Andersson (Andersson, 2000; Andersson et al., 1998), who has utilized LCA in case studies of tomato ketchup production. The objective of that work was mainly to highlight the merits and shortcomings of this tool and also to generate information regarding the environmental impacts of food processing systems. The investigated system was quite broad as it included: • The tomatoes cultivation and paste generation phase taking place in Mediterranean countries, with inputs such as fertilizers as well as packaging for the produced paste. • The ketchup production phase taking place in Sweden, with inputs such as sugar, salt, spices, etc. • The ketchup packaging phase, including details about the production utilization and disposal of the bottles, caps and other employed materials. • The transportation phase included in most of the previous phases, as all raw materials are produced in different geographical locations. • The consumption phase involving transportation to/from wholesalers and retailers and storage in refrigerators. Within the context of all the above phases the authors developed life-cycle inventories regarding the incoming and outgoing material and energy flows in order to determine contributions of the overall system in impact categories such as primary energy utilization, global warming, acidification, eutrophication, photo-oxidant formation, human toxicity, ecotoxicity and radioactive waste. The packaging and food processing phases were identified as important impacts in several of the categories. However, it was the findings about the time of ketchup household storage that highlight the strength of the LCA methodology. For the case of a yearround storage it was found that the household energy use could be as high as the combined energy use of packaging and food processing together. This indicates that the simple yet systematic calculations involved in LCA studies can spot very important interactions within a system that are not otherwise obvious. 19.4.5 Applications combining multiple tools This section presents several applications in the food industry which simultaneously combine many of the previously presented tools. © Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 465 Soybean oil production Li et al. (2006) perform an LCA-based case study regarding the production of soybean oil. The study is performed to address the environmental details of the life cycle for soybean oil production and to identify the environmental impact of several stages involved in soybean oil production. The process is energy intensive mainly due to the need to dry the raw soybeans prior to further processing but also due to the use of solvent-based extraction of the soybean oil, where intense heating is required. The focus is on investigating energy efficient technologies involving supercritical CO2 extraction to replace the conventional hexane-based extraction, which is both environmentally harmful and flammable. The authors compare different CO2-based extraction methods and propose operating modifications to improve the environmental performance of the production process. This work highlights how different structural and operating process configurations (e.g., type of equipment, processing temperature or pressure) can be combined with LCA considerations to improve the environmental performance. However, the attempt to consider the environmental perspective of different processing options is addressed heuristically. Optimization-based scheduling for optimum food distribution systems A more systematic attempt to combine the decision making involved in the development of food processing systems and sustainability considerations has been proposed by Akkerman and co-workers. In particular, Akkerman and Donk (2008) focus on the development of a decision support method that addresses product losses in the food industry and their effects in process economics and sustainability. In particular, the proposed decision support tool enables the evaluation of different scenarios for planning decisions and production parameters. This work is further expanded by Akkerman, Wang and Grunow (2009) to enable a broader consideration of sustainability, and address decision making through an optimization-based approach. The proposed approach is presented through a case study on the production and distribution of prepared meals and addresses the environmental, social (health and safety) and economic dimensions of sustainability. The main focus is on linking economic decisions regarding distribution of food products with the aim to maintain high quality and avoid food losses. In this respect the goal is to determine the optimum storage and transportation temperatures of the meal products combined with production quantities and delivery paths to minimize the total costs subject to demand and quality constraints. Quality is modeled as a function of storage temperature and time and is combined with planning parameters such as transport lead times, production capacities, batch sizes, retailer demands, cooling, storage and waste disposal costs, etc. (Rong et al., 2011). The environmental and economic implications of planning decisions in the operation of the process are quantified and incorporated in an optimization model that consists of linear equations and is solved using MILP technology. Results indicate that such systematic and relatively easy to use tools may reduce planning-related food losses due to quality deterioration by 20%.

© Woodhead Publishing Limited, 2013

466 Robotics and automation in the food industry Scheduling and design of cheese manufacturing processes Stefanis et al. (1997) address the consideration of environmental impacts in the optimal design and scheduling of batch processes. Their approach is mostly focused on wastes, rather than the entire life cycle of products or processes; however, major elements of LCA are used in the proposed methodology, such as identification of system boundaries and environmental impact indicators obtained from inventory analysis. The presented applications involve the optimum design and scheduling of a cheese manufacturing process represented as a multipurpose plant. Such plants involve the production of several products through shared equipment. In this case the products are low and high fat cheese curd and the utilized equipment (e.g., reactors to produce curd from milk and draining vessels) can become contaminated both microbiologically and from fouling deposits. Cleaning is therefore required with cleaning-in-place stations to flush detergents. This is time consuming so the entire process needs to be designed and operated (i.e., schedule cleaning and production tasks) in a way that the economic performance is maximized and the environmental impact of the process and cleaning wastes is minimized. In this respect the design is rather challenging, and multi-objective optimization is used to simultaneously evaluate environmental impact and economic performance. The considered models and design constraints include unit capacity and batch size, storage, utilities, mass balances, production requirements, unit operations and a reverse osmosis design model. The optimization results identify both process structures (i.e., connectivity between employed equipment) and schedules (i.e., temporal utilization of each equipment for each product) with different trade-offs among economic and environmental performance. For example the authors find that plants of minimum cost result in high product yields and waste generation. On the other hand, low environmental impact can be achieved with low capital equipment expenditures but this is counter-balanced by increased energy and raw materials consumption.

19.5 Advanced tools and methods for sustainable food engineering with potential applications This section describes advanced tools for automated and sustainable processing as well as case studies in food engineering. These tools have found considerably fewer applications in the food industry compared with the tools described in Section 19.4, however, they show, evidence of great potential for the future. The final part of this section presents combined applications of these tools in the form of advanced methods. As such applications are currently scarce in the food industry, examples are retrieved from the broader chemical industry in order to illustrate the evolution and merits of these methods.

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 467 19.5.1 Process and product synthesis Process and product synthesis is viewed as an extension to optimization in order to address the issue of identifying optimum design schemes that offer significant improvements. Introduction and scope As previously noted, conventional optimization mostly aims at setting a formal and systematic mathematical framework to evaluate the combinations of decision parameters based on a set of performance measures. Clearly, the selection of the appropriate decision parameters to use in optimization is a crucial step. Practically, there are a vast number of system parameters that can be used to address the decision making involved in a design optimization problem, only limited by the imagination of the engineer who defines the problem, as noted in Johns (2001). Process and product synthesis includes a collection of tools that provide the means to identify the desired decision parameters, to justify their selection, and to organize them systematically, in order to facilitate the identification of high quality design solutions. In this respect, the formulation of a computationally impracticable optimization problem, due to disordered consideration of all available decision parameters, is avoided. The limited value of conventional optimization, due to arbitrary selection of few decision parameters, is significantly increased by the far-reaching design improvements that can be obtained by process and product synthesis. This is well illustrated in the application of optimization-based synthesis for the design of rice processing plants (Phongpipatpong and Douglas, 2003), described in the following sections. The authors develop efficient rice processing flowsheets and report 74% increase in profits compared to conventional systems. Process synthesis includes numerous types of tools, namely evolutionary synthesis, expert panel, thermodynamic pinch, implicit enumeration, superstructure optimization, graph-theoretic superstructure generation, artificial intelligence, design rationale, and hierarchical decomposition (Johns, 2001). Product synthesis is mostly realized through computer-aided molecular design, which is discussed later in this section. Most of the available process synthesis tools are automated and included in software packages, even though with simple capabilities. In particular, three tools have been identified as promising options for further utilization, namely implicit enumeration, superstructure optimization, and artificial intelligence (Tsoka et al., 2004). These tools address (a) the efficient representation of decision options, and (b) the solution of the resulting process synthesis problem. Description and case study This section details the implementation of process and product synthesis through case studies reported in the food industry. Representation of decision options All process flowsheets involve a set of interconnected tasks that process and transform raw materials into products. The design of flowsheets often involves selection of the tasks and their connections based on prior know-how with regards

© Woodhead Publishing Limited, 2013

468 Robotics and automation in the food industry 6 1

1

4

2

2 4

6 3

3

5

Process flowsheet

5

Representation as graph

Fig. 19.7 Representation of process flowsheets as graphs.

to the desired process. This limits the opportunities to capture novel process schemes of optimum performance. In process synthesis, flowsheets are represented by directed graphs as shown in Fig. 19.7. The associated published literature involves different adaptations of such graphs, called state-task networks (Kondilli et al., 1993), mixer-splitter models (Biegler et al., 1997), or p-graphs (Friedler et al., 1992). Such graphs represent generic systems that have the following characteristics: • Structure that is determined from the system parts and their composition. • Behavior that involves intake, processing and output of mass, energy or information. • Components that involve structural and functional relationships. • Limits that are determined from differentiation of the above characteristics from those of a different system. Clearly, the above characteristics are common to all process systems regardless of the processing tasks involved or the way that they are interconnected. In this respect, each process can be considered as a node (vertex) and each stream can be considered as an edge, without the need to consider mass balances or design equations. The desired connections between nodes are determined through an adjacency matrix, which represents connections through binary (0−1) variables. If real variables are used instead, the adjacency matrix represents both desired connections and flow compositions within the process streams. Multidimensional adjacency matrices can be used to model multicomponent flows or multiphase systems. Overall, the use of graphs in process synthesis results in the representation of: • all possible or desired connections between processes, • streams of any desired composition, • processes that are independent of their functionality in a particular system. As a result, different processing tasks can be assigned to each process by allocating appropriate sets of design equations (e.g., mass, heat transfer etc.) in each node of the graph. Figure 19.8 shows a case of a mixer–splitter model commonly utilized in superstructure optimization. Each cell (node) can become a different processing task, while mixers and splitters represent summation and redistribution of streams, respectively. As a result, the type, size and numbers of nodes can

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 469 1

2

n

Splitter Mixer Side feed Bypass Recycle Side product

Fig. 19.8 Superstructure representation using a mixer−splitter model.

become decision parameters, together with operating conditions in each node, mass flowrates and composition of streams, side feed and product positions, etc. The representation using graphs is common to implicit enumeration and superstructure optimization. On the other hand, artificial intelligence methods include the use of rules to generate process flowsheets. Rules can be considered as a systematic representation of heuristic knowledge that has been obtained through engineering know-how or analysis of already available optimization results. In this respect, artificial intelligence methods can be used in conjunction with implicit enumeration or superstructure representations to either enhance or reduce their representation capacity. Phongpipatpong and Douglas (2003) applied a superstructure-based process synthesis approach in the design of a rice processing system. The system consists of a series of drying, cooling and tempering processes that require increased energy. However, these processing stages are necessary to maintain high grain quality and result in a dry product facilitating the subsequent milling operation. The authors identify seven possible routes for the raw material (e.g., drying or drying-cooling, etc.) that needs to pass a maximum eight times through the available units prior to storing or milling. This results in over 5 million potential combinations that need to be screened in order to identify the optimum sequence of units. In this respect they develop a superstructure in the form of a graph connecting different tasks (i.e., drying, cooling, tempering) at different passes of the raw material, capturing all the considered combinations in a systematic way. Solution of process synthesis problem Numerous decision options considered in process synthesis are represented by integer variables. For example, the decision whether to consider or not a piece of equipment in a process flowsheet can be represented by a binary variable, while other decisions may involve real integer numbers. This is also the case in the rice processing case study, where the existence or non-existence of the drying, cooling or tempering unit in each pass is modeled using binary variables. Implicit enumeration algorithms enable the solution of such integer programming problems using a coordinated, heuristically structured search of the space of all feasible solutions (Edgar and Himmelblau, 1988). The main idea is to explore numerous combinations of the discrete variables in order to identify solutions to the discrete design problem that can be eliminated from further consideration. In this respect, the optimum solution to the integer optimization problem is approached through a

© Woodhead Publishing Limited, 2013

470 Robotics and automation in the food industry reduced enumeration of the potential solutions. A simple yet inclusive elaboration of the method implemented through a graphical representation of a flowsheet can be found in Johns (2001). In superstructure-based process synthesis the main goal is to obtain specific designs out of a superstructure that considers every feasible realization for the addressed process flowsheet. In this respect, the optimization targets the removal of excessive, and often complex, features from the superstructure in order to identify an optimum solution. This is in contrast to implicit enumeration, which is based on a depth-first search of the feasible solutions starting from simple configurations, without necessarily requiring the evaluation of more complex ones (Sargent, 1998). However, it is often the case that novelty is approached in process synthesis through not immediately noticeable complex configurations, which are subsequently refined to practical realizations. Integer variables are also used in superstructure optimization, while the solution of the resulting problem is approached by MI(N)LP or stochastic search algorithms. The rice processing case study is solved using an MINLP solver. The problem of identifying the optimum combination of units, number of passes and operating conditions is formulated as a typical optimization problem. Phongpipatpong and Douglas (2003) consider six objective functions including minimum production time, minimum number of processing units, maximum head rice yield (a quality indicator), minimum energy consumption and minimum operating cost. The considered constraints mostly involve the moisture content of the rice at different passes and temperature conditions. The authors compare their results with a rice processing system developed using a conventional approach and find that their solution gives a 74% increase in profits, at 22% decrease in energy consumption and a 2.4% improvement in the product quality (head rice yield). Process synthesis using artificial intelligence is based entirely on heuristics; hence, it does not involve mathematical optimization. In this respect, the previous two methods are expected to produce a better result in most cases. However, artificial intelligence methods are computationally faster than mathematical methods, as they present lower demands on robust models and numerical analysis. This advantage can be exploited in order to introduce the much required heuristic design knowledge in mathematical synthesis methods in order to improve the obtained results and reduce the required computational effort. Remarks Process synthesis methods are suitable for automated design using computing environments and will benefit greatly from future advances in computing and optimization technologies. These tools are able to handle the increased design complexity arising from sustainability requirements, while accomodating wellestablished heuristics that incorporate engineering know-how in process design. Although such tools are yet to be widely adopted, their application to industrial design projects has already enabled significant cost savings. Increasingly stricter environmental regulations and sustainability requirements are expected to drive their broader commercial utilization (Johns, 2001; Tsoka et al., 2004).

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 471 19.5.2 Integration and dynamic operability This section describes process dynamic controllability issues in the context of process integration and design for applications in the food industry. Introduction and scope Even though process integration enables resource utilization minimization and eventually the improvement of sustainability in the overall industrial processing system, the interaction among various subsystems increases significantly. Unless a careful design of the integrated system that explicitly considers the operability characteristics of the individual process subsystems is performed, the advantages of process integration usually reflecting steady-state economics may be diminished under real operating conditions. A high degree of process integration unavoidably leads to higher interaction among the various subsystems. Therefore, disturbances affecting one subsystem are likely to affect other subsystems through the interconnecting and interacting process streams. For example, heat integration will eliminate extensive use of steam or cooling water but variability in the temperature of a hot process stream that replaces steam will be transferred to the temperature of the cold process stream. Such a situation will limit the ability of the control system to compensate for the effects of disturbances on the process specifications and control objectives with significant impact on the product quality and the plant economics unless the overall process is designed in a way to anticipate the increased degree of interaction (Luyben, 2004). Thus the sustainability indicators will deteriorate under dynamic variations. The solution is the simultaneous integration of processes with the design of the control system for improved dynamic operability. Dynamic operability can be perceived as the ability of the process and the control system to successfully anticipate the effects of a disturbance. The simultaneous design evaluates the impact that design decisions (e.g., flowsheet alternatives such as by-pass streams, additional heaters or coolers, stream splitting and mixing and so forth) have on the control objectives. Similarly, the selection of a suitable control structure enables successful disturbance compensation while satisfying the safety, product quality and other operating specifications. Usually, process controllability is conceived as an inherent property of the process design and the control structure, independent of the control algorithm and its tuning (Seferlis and Grievink, 2004). Description and case study The implementation of integration and dynamic controllability methods in the soybean oil production industry is reported as a representative example. Expected challenges Soybean oil manufacturing provides an illustrative case study, where heat integration and dynamic operability may play an important role in the overall system performance. Extracted soybean oil is passed through a number of multi-effect evaporation stages and then is processed in a stripper for separation from the extraction solvents, as shown in Fig. 19.9 (Hammond et al., 2005). This is a typical sequence of energy intensive operations found in many food processing

© Woodhead Publishing Limited, 2013

472 Robotics and automation in the food industry Feed Steam

CW

H2O

H2O H2O

Extraction solvent

Steam

Soybean oil

Fig. 19.9 Base-case flowsheet for separation of soybean oil from extraction solvent. CW indicates a cooling water stream.

Feed

Steam make-up

H2O H2 O

Extraction solvents

Steam Soybean oil

Fig. 19.10 Potential imporvement of base-case flowsheet for energy reduction. CW indicates a cooling water stream.

flowsheets. The energy requirements can be reduced if the gas stream from the top of the distillation column is condensed in the first evaporator as shown in Fig. 19.10. In such a design, the need for cooling water and steam will be significantly reduced, while investment costs will be decreased due to the removal of the condenser. However, in the heat integrated design, changes in the water content of the feed stream will have a strong impact on the operation of the distillation column and the stream leaving from the top of the column. It is apparent that the system lacks heating power to compensate for the additional water content.As a remedy, an additional steam stream is attached to the second evaporator to provide the necessary heating power to the system. Another problem in the heat integrated flowsheet is the start-up procedure that will be substantially prolonged leading to

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 473 larger amounts of off-spec product. Therefore, the strong interaction between the evaporators and the distillation due to heat integration poses numerous challenges for the control system performance. Solution methods The solution to the aforementioned problems can be resolved by considering the integrated process design (with a high degree of integration for steady-state economic benefits) and control system design. These approaches can be categorized as successive and simultaneous. Successive methods commence with an analysis of the key static and dynamic characteristics introduced to the system by the design decisions, process integration and control structure selection and subsequently lead to an evaluation of alternative flowsheet configuration and control system structures using meaningful performance criteria. Operability regions (Subramanian and Georgakis, 2005) and non-linear sensitivity analysis (Seferlis and Grievink, 2004) examine the steady-state feasibility window for the process and control system under single and multiple disturbances, respectively. For instance, these methods can determine the water content changes that can be tolerated by the system so that the associated control structure can meet the product specifications. Frequency response analysis (Lewin, 1996) investigates the effect of disturbances at different frequencies (e.g., flow variations due to pump vibrations) on the steady-state control effort and compensates for them. Bifurcation analysis (Dimian and Bildea, 2004) provides useful information on the non-linear effects so that undesired system behavior (e.g., instability) can be avoided. On the other hand, simultaneous approaches involve the determination of the optimal design configuration and controller within a unified framework that involves a complete evaluation of the dynamic performance of the system (Schweiger and Floudas, 1997; Kookos and Perkins, 2004; Sakizlis et al., 2004). Successive methods build knowledge for the design gradually through small targeted steps, ensuring feasibility of the final design. Simultaneous methods on the other hand enable the investigation of all possibilities and combinations within a pre-defined search space but suffer when dealing with large scale problems due to high complexity.

19.5.3 Computer-aided molecular design Computer-aided molecular design (CAMD) is reported as an important product synthesis and optimization tool with potential applications and benefits for the food industry. Introduction and potential applications CAMD is a tool that has been widely utilized for the sustainable design of chemical products such as environmentally benign solvents (Buxton et al., 1999; Hostrup et al., 1999; Marcoulaki and Kokossis, 2000a; Papadopoulos and Linke 2006a, 2006b, 2009; Folic et al., 2008) refrigerants (Duvedi and Achenie, 1996; Marcoulaki and Kokossis, 2000b) and working fluids for Organic Rankine Cycle processes (Papadopoulos et al., 2010a, 2010b). Although the approaches followed

© Woodhead Publishing Limited, 2013

474 Robotics and automation in the food industry by these authors differ significantly, optimization- and synthesis-based tools have been employed on several occasions to enable the development of chemical products with desired actions and functionalities that can be conveniently linked with sustainability indicators. This tool can have significant implications in the food industry in the following ways: Design and selection of efficient solvents The previously reported work by Li et al. (2006) addresses the production of soybean oil through a solvent-based process. This existing process utilizes hexane as the solvent, which is harmful and flammable. These authors propose the use of supercritical CO2 extraction, which has better environmental performance but is generally less tested in the chemical industry compared to organic solvents. Instead, CAMD tools could be used to identify organic solvents with improved safety, environmental and economic properties compared to both hexane and supercritical CO2. Papadopoulos and Linke (2006a) have shown that solvents designed using CAMD tools may result in up to 50% reduction in the operating costs of industrial separations compared to conventionally selected solvents. Design and selection of efficient refrigerants and heat recovery working fluids Application of the CAMD tool in the optimum design of refrigerants is equally important for the food industry. Refrigerant fluids, being the most important components of refrigeration units, can enable significant reduction in energy consumption by improving the heat transfer between a low temperature source and a high temperature sink. Huan (2011) argues that as much as 70% of food products are maintained using refrigeration through the so called cold chain. Refrigeration accounts for approximately 50% of the total energy consumption in the food industry, while in the United Kingdom alone the food industry consumes about 4500 GWh/y of electricity (Huan, 2011). Clearly, the use of highly efficient refrigerants can result in significant benefits in terms of the utilized energy with immediate reductions in CO2 emissions as this energy is often provided by conventional, harmful energy supplies (e.g., fossil fuels). Furthermore, the use of CAMD can result in refrigerants with significantly improved environmental and safety properties. Most refrigerants currently involve chlorinated or fluorinated compounds, which have a detrimental effect on ozone depletion and global warming, and they are in a phasing out process following the Kyoto and Montreal protocols. However, these compounds are significantly safer (less flammable) than other fluids with better environmental properties. CAMD has already been used to design new refrigerants that are at least equally efficient and have a better environmental performance than the widely used refrigerant Freon-12 (Marcoulaki and Kokossis, 2000b). In a similar context (Papadopoulos et al., 2010a, 2010b) presented the first application of CAMD in the design and selection of refrigerants and other organic compounds to be used in Organic Rankine Cycle (ORC) systems. These systems

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 475 are very important for the production of electricity from recovered low temperature heat (>100°C) that is widely available in most industrial processes and is often wasted in the environment. Food industries involve operations of intense heating, therefore it is reasonable to expect that streams of low temperature exist that can be used for power generation. ORCs are relatively simple units, very similar to refrigerators with the exception of a turbine introduced between the vaporizer and the condenser. They are used to produce power from a few kW to several MW, and current applications involve mainly waste heat recovery in major chemical plants (e.g., refineries) as well as power generation from geothermal fields. Papadopoulos et al. (2010a, 2010b) showed that the use of CAMD tools for the design of ORC working fluids enabled an approximately 35% increase in the economic performance of such systems compared to conventional fluids, while maintaining beneficial environmental and safety characteristics. Description This section presents a brief description of CAMD development and implementation details. CAMD features The design of molecules using CAMD capitalizes on the development of computational methods in order to screen for molecules with superior performance, while avoiding time consuming and unfocused experimental studies. Molecules are represented through simple chemical (functional) groups of previously measured and registered contributions that enable the calculation of a wide range of chemical, physical, environmental, safety and other properties (Fig. 19.11). The random choice of available chemical groups results in the synthesis of a molecule that is tested with regards to its chemical feasibility using chemical rules. The desired molecular properties are subsequently calculated based on the contributions of the groups contained in the molecule. One or several of such properties are utilized as molecular performance measures in optimization. An optimization method is employed to evaluate the iterative generation of molecules and to guide the optimization search towards highly performing choices. As such, the behavior of already known molecules can be predicted or novel molecules can be synthesized, hence providing radical performance improvements when compared with conventional molecular options. Remarks The work of Tsoka et al. (2004) has identified CAMD as the tool with the highest potential contribution to the sustainable development of process industries. This is not surprising, considering that chemical materials are directly responsible for economic, environmental and societal impacts, while the industrial process systems are simply the means to produce them. In view of the high importance of chemicals, CAMD tools are useful because they systematically address the vast combinatorial complexity that exists in the design and selection of products with desirable characteristics. As noted in Tsoka et al. (2004), the combination of only

© Woodhead Publishing Limited, 2013

476 Robotics and automation in the food industry Molecular simulator CAMD

Molecular simulator i.e., –CH2–, OH, –CH3 –COOH etc. Chemical constraints e.g., valence, free bonds, etc. Molecule i.e., CH3–OH

Property calculation (group contributions) e.g., selectivity

Optimization algorithm Generation of new molecular structure

Evaluation of performance measure

Fig. 19.11 Algorithmic stages of optimization-based computer-aided molecular design.

20 carbon and 42 hydrogen atoms may exceed the number of chemicals known to mankind. Clearly, the automatic generation of molecules is one of the great advantages of CAMD, accompanied by the immediate evaluation of the molecular performance using indicators that are directly associated with sustainability.

19.5.4

Applications of advanced methods for automated sustainable design The previous analysis highlighted major computer-aided tools that can contribute directly or indirectly to different dimensions of sustainable process or product design. This section identifies methods that involve the systematic combination of such tools in order to consistently exploit their benefits and address more than one dimension simultaneously. The methods are described through applications reported in published literature for the broader chemical process industry, as applications in the food industry have yet to be reported. The benefits of using these methods are similar to the benefits described for each one of the individual tools presented in the previous sections. The added value is that by combining all these tools in the form of systematic methodologies sustainability considerations can be integrated within a broader context, capturing economic, environmental

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 477 and societal dimensions in more detail and hence resulting in improved design and operating solutions. Early attempts Kniel et al. (1996) proposed a method to incorporate LCA considerations with process optimization. The method is presented as a case study on a nitric acid production plant. LCA is used to quantify the environmental performance of a number of design alternatives aimed at waste reduction. Economic models are developed for each alternative and associated with environmental models obtained from LCA. The optimization is performed using multi-objective technology. The focus of the work is to link the environmental models, which are developed based on a full implementation of LCA, with process design decisions. The process models, decision parameters and objective functions employed are rather limited. The author identifies these shortcomings and also suggests that social factors be considered in the future, together with environmental impacts. Azapagic and Clift (1999) proposed the application of LCA together with process optimization in order to identify trade-offs between economic and environmental performance measures. The environmental measures employed as objective functions in optimization are identified during the goal and scope phase of LCA; hence, optimization is performed either at the life cycle inventory or at the impact assessment phases. The case study presented uses a linear model for the process system, the objective functions consist of environmental (17 burdens or 7 impacts) and 2 economic indicators, while the decision parameters represent material and energy flows. The problem is addressed using multi-objective optimization technology. Advanced process synthesis and optimization using LCA and environmental impacts Diwekar (2003) proposed a multilevel process design and optimization framework aiming to enable incorporation of green engineering principles as early as possible at all levels of engineering decision making. The framework employs process simulation models to assess the effects of variations imposed on decision and uncertain parameters using multi-objective optimization technology. The performance measures utilized as objective functions during optimization include indicators associated with the green engineering principles of Anastas and Warner (1998), which are essential guidelines providing design directions towards sustainable industrial processes. The work considers principles that can be quantified through operating, structural and environmental parameters represented in process models and utilized as design decisions. However, the author states that the framework does not explicitly address questions related to LCA, industrial ecology and socio-economic sustainability, as such issues involve considerations of a non-quantitative nature. Guillen-Gosalbez et al. (2008) propose the combined utilization of superstructure optimization and LCA for the systematic incorporation of environmental factors in the design of process flowsheets. The method employed addresses the synthesis of process flowsheets using structural and operational decision options.

© Woodhead Publishing Limited, 2013

478 Robotics and automation in the food industry The employed LCA considers 11 impacts based on Ecoindicator 99, which is an index representing damage to human health, ecosystem quality and resources. The burdens corresponding to these impacts are associated with the process based on operating parameters that represent flows of raw materials, by-products and consumed energy. Shortcut (simplified) process models are utilized in the case study involving the design of a toluene hydro-dealkylization process. Multi-objective technology is employed to address the optimization problem. Sustainability considerations in design Azapagic, Millington and Collett (2006) presented a multilevel design framework that employs LCA enhanced with sustainability indicators that cover economic and societal aspects in addition to environmental. The focus of this work is mainly on how to incorporate such indicators in process design stages from project initiation to final design. The approach is illustrated through a case study that addresses only preliminary design, but illustrates the relevance of sustainability criteria in process design, and how to assess design decisions in view of sustainability. Sugiyama et al. (2008) present a multistage process design framework that considers sustainability criteria. The framework targets the early design phase during the development of a process flowsheet that includes the identification of promising chemistries and involves conceptual design for the selection of the appropriate equipment and operating conditions. The four stages comprising the framework involve screening and filtering of promising routes, followed by multi-objective optimization and sensitivity analysis. The improvements targeted in each stage are represented by indicators addressing economic criteria, life-cycle environmental impacts and hazards that involve workers’ health and safety. In this respect, the framework considers sustainability indices that capture both environmental and societal impacts of the process design decisions, in addition to economic impacts.

19.6 Software technologies for automated sustainable design There are currently several commercially available software packages that incorporate tools reported in the previous sections. The Aspen Suite includes numerous tools that enable process modeling, optimization and control such as Aspen Plus®, Aspen Hysys® or Aspen OnLine® (Aspen Technology Inc., www.aspentech.com). These tools allow the development and simulation of process flowsheets using interconnected models of process systems, while performance indicators can also be utilized to enable optimization. Similar capabilities are included in other well known software packages such as the SuperPro Designer and Chemcad, which provide a user-friendly environment to perform process simulation and optimization by supporting numerous types of generic modules that represent chemical process equipment. Other software package, such as gPROMS (Process Systems Enterprise Ltd, www.psenterprise.com) or GAMS (GAMS Development Corporation, www.gams. com), enable the development of custom-made models using a programming-oriented

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 479 approach. Such packages mainly target the robust solution of more complex simulation and optimization problems using powerful solvers. Process synthesis software is focused mostly on provision of modules for generic representation of process networks and highly specialized optimization algorithms. PNS is one such software package that utilizes the P-graph synthesis approach (Friedler et al., 1992). It provides graphical interfaces to model process networks and also utilizes appropriate algorithms for their optimization. LogMIP and MIPSYN (Lam et al., 2010) are two software packages that target the efficient representation and optimization of complex process synthesis problems. These two packages integrate the GAMS environment for the development of models with advanced MINLP-based optimization algorithms. Such packages generally require expertise in process synthesis, but the provided optimization capabilities are significantly more advanced than environments such as the Aspen Suite. Process integration software packages that are based on the Pinch technologies are widely available. The most notable packages that incorporate user-friendly interfaces involve SPRINT, STAR, WORK and WATER (Lam et al., 2010), which have been developed by a group at the University of Manchester that pioneered the Pinch technology. A comprehensive review of these and other similar packages is available from Lam et al. (2010). Software packages that incorporate CAMD methods are more limited. One such package that includes a graphical interface is the ProCAMD, developed by a group in the Technical University of Denmark. Optimization-based CAMD software has also been developed as a result of research work by Marcoulaki and Kokossis (2000a) and further extended by Papadopoulos and Linke (2005). This software enables synthesis of novel or conventional molecular structures regardless of the application in which they will be utilized, and incorporates advanced synthesis and optimization methods. Software packages that enable LCA are also widely available. Garcia-Serna et al. (2007) provide a comprehensive review of such packages. Despite the existence of all these computer-aided packages, the more advanced methods reported in Section 19.5, are systematic methodologies resulting from research work. Such methodologies provide significantly more advanced capabilities than the individual packages; however, they are rather difficult to incorporate in a commercial computer-aided product based on the currently available software technologies. The mounting pressure for sustainability in all industrial sectors and the promising developments in terms of methods and algorithms provide the opportunity for a broader application of such technologies in the future.

19.7 Conclusions and future trends The presented work has highlighted a number of important automated tools, methods and software packages utilized in the process industry that can be used to address sustainability issues in the food processing industry. Tools such as process optimization, synthesis, integration and CAMD can generally be utilized to systematically identify processing or product design options that can improve the

© Woodhead Publishing Limited, 2013

480 Robotics and automation in the food industry economic, operating or environmental performance of a process system or a product. These tools have been mostly utilized to target process or product economics, while the associated environmental implications are often covered through indicators that focus on particular production stages or impacts. Process integration for dynamic operability is seldom considered in the design of food processing plants despite the potential benefits in both economical and sustainability terms. The incorporation of real-time optimization in food processing will eventually enhance the ability of the plant to respond to multiple changes at different layers (e.g., operational, market, environmental) in the most effective and profitable way. On the other hand, a tool such as LCA is used to identify opportunities for environmental improvements in the entire life cycle of a process or a product and provides the indicators required to quantify environmental impacts. In this respect, LCA has been used in combination with several tools in order to enable a more inclusive representation of the environmental implications involved in process or product development. Recently presented research efforts are also attempting to incorporate the social dimension of sustainability in methods that already address process economics and environmental impacts. These methods mostly involve the use of process optimization and process synthesis, while process integration and CAMD are mostly limited to addressing economic performance and targeted environmental impacts. Despite the huge scope for utilization of these tools and methods in the food processing industry, the applications are still limited. The reported works mostly involve the use of LCA, the combined use of optimization and LCA, or the use of process integration methods. Clearly, there is a need to utilize more advanced tools and methods in the food processing industry in order to enhance sustainability. The availability of numerous advanced computer-aided software tools is expected to meet the needs for a sustainable food processing industry in the near future.

19.8 Sources of further information and advice Lam et al. (2010) have recently presented a comprehensive work regarding software packages for process integration, modeling and optimization that address energy saving and pollution reduction. The work is essentially a review of process integration and retrofit analysis tools, general mathematical suites with optimization libraries, flowsheeting packages and process optimization tools. It further provides an analysis on how such tools can be utilized for the design of sustainable industrial process systems. The book by Biegler et al. (1997) is also one of the most important works on the design, synthesis and optimization of chemical processes. It starts by discussing methods for preliminary analysis and assessment of process systems such as mass and energy balances and costing of equipment, and addresses the issue of utilizing rigorous unit operation models in process design and flowsheeting. Subsequently, the book proceeds to more advanced issues regarding process syn-

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 481 thesis and integration methods, while it also discusses all the major MI(N)LP techniques employed in synthesis and optimization of chemical processes. The work by Seferlis and Georgiadis (2004) is a comprehensive collection of recent advances in integrated design and control that addresses process characterization, methods for simultaneous design and control and plantwide control system design. The work of Azapagic et al. (2004) is also an interesting collection that presents the issue of sustainability applied to case studies in the chemical process industry. The book covers issues of wastewater treatment and management, process design in the context of sustainability applied in polymer production and renewable energy systems for power generation, to name a few. A similarly comprehensive collection that covers sustainability issues such as water and energy management with particular emphasis in the food industry is the work by Klemes et al. (2008). The book addresses issues that range from legislation to technologies utilized in water and energy management and is very broad in scope as it can be utilized as a reference point for the application of energy and water integration methods in any other industrial sector.

19.9 References Akkerman R. and van Donk D.P. (2008), ‘Development and application of a decision support tool for reduction of product losses in the food processing industry’, Journal of Cleaner Production, 16, 335–342. Akkerman R., Wang Y. and Grunow M. (2009), ‘MILP approaches to sustainable production and distribution of meal elements’, International Conference on Computers and Industrial Engineering, CIE 2009, art. no. 5223667, 973–978. Anastas P.T. and Warner J.C. (1998), Green Chemistry: Theory and Practice, Oxford University Press, New York. Andersson K. (2000), ‘LCA of food products and production systems’, International Journal of LCA, 5(4), 239–248. Andersson K., Ohlsson T. and Olsson P. (1998), ‘Screening life cycle assessment (LCA) of tomato ketchup: A case study’, Journal of Cleaner Production, 6, 277–288. Atkins M.J., Walmsley M.R.W. and Neale J.R. (2011), ‘Integrating heat recovery from milk powder spray dryer exhausts in the dairy industry’, Applied Thermal Engineering, 31, 2101–2106. Azapagic A. and Clift R. (1999), ‘The application of life cycle assessment to process optimization’, Computers and Chemical Engineering, 23, 1509–1526. Azapagic A., Millington A. and Collett A. (2006), ‘A methodology for integrating sustainability considerations into process design’, Chemical Engineering Research and Design, 84(A6), 439–452. Azapagic A., Perdan S. and Clift R. (2004), Sustainable Development in Practice, West Sussex, U.K., John Wiley and Sons Ltd. Balsa-Canto E., Banga J.R. and Alonso A.A. (2002), ‘A novel, efficient and reliable method for thermal process design and optimization. Part II: applications’, Journal of Food Engineering, 52, 235–247. Banga J.R., Balsa-Canto E., Moles C.G. and Alonso A.A. (2003), ‘Improving food processing using modern optimization methods’, Trends in Food Science and Technology, 14, 131–144.

© Woodhead Publishing Limited, 2013

482 Robotics and automation in the food industry Banga J.R. and Singh R.P. (1994), ‘Optimization of air drying foods’, Journal of Food Engineering, 23, 189–211. Bezergianni S., Dimitriadis A., Kalogianni A. and Knudsen K.G. (2011), ‘Toward hydrotreating of waste cooking oil for biodiesel production. Effect of pressure, H2/oil ratio, and liquid hourly space velocity’, Industrial and Engineering Chemistry Research, 50(7), 3874–3879. Biegler L.T., Grossmann I. E. and Westerberg A.W. (1997), Systematic Methods of Chemical Process Design, New Jersey, U.S.A., Prentice Hall International Series in the Physical and Chemical Engineering Sciences, Prentice Hall PTR. Brundtland G.H. (1987), For World Commission on Environment and Development, Our Common Future, Oxford, UK, Oxford University Press. Buxton A., Livingston A.G. and Pistikopoulos E. (1999), ‘Optimal design of solvent blends for environmental impact minimization’, AIChE Journal, 45(4), 817–843. Chalabi Z.S., van Willigenburg L.G. and van Straten G. (1999), ‘Robust optimal receding horizon control of the thermal sterilization of canned foods’, Journal of Food Engineering, 40, 207–218. Department for Environment Food and Rural Affairs (DEFRA) (2006), Food Industry Sustainability Strategy, Technical report, Available from: http://www.defra.gov.uk Dimian A.C. and Bildea C.S. (2004), ‘Component inventory control in recycle systems’, In Seferlis P. and Georgiadis M.C., eds., The Integration of Process Design and Control, Computer Aided Chemical Engineering, 17, 401–429. Diwekar U.M. (2003), ‘Greener by design’, Environmental Science and Technology, 37, 5432–5444. Duvedi A.P. and Achenie L.E.K. (1996), ‘Designing environmentally safe refrigerants using mathematical programming’, Chemical Engineering Science, 51(15), 3727–3739. Edgar T.E. and Himmelblau D.M. (1988), Optimization of Chemical Processes, McGraw Hill Chemical Engineering Series, McGraw Hill Book Company. Folic M., Gani R., Jiménez-González C. and Constable D.J.C. (2008), ‘Systematic selection of green solvents for organic reacting systems’, Chinese Journal of Chemical Engineering, 16(3), 376–383. Friedler F., Tarján K., Huang Y.W. and Fan L.T. (1992), ‘Graph-theoretic approach to process synthesis: Axioms and theorems’, Chemical Engineering Science, 47(8), 1973–1988. Garcia-Serna J., Perez-Barrigon L. and Cocero M.J. (2007), ‘New trends for design towards sustainability in chemical engineering: Green engineering’, Chemical Engineering Journal, 133, 7–30. Georgiadis M.C., Rotstein G.E. and Macchieto S. (1998), ‘Modelling and simulation of complex plate heat exchanger arrangements under milk fouling’, Computers and Chemical Engineering, 22, S331–S338. Guillen-Gosalbez G., Caballero J.A. and Jimenez L. (2008), ‘Application of life cycle assessment to the structural optimization of process flowsheets’, Industrial and Engineering Chemistry Research, 47, 777–789. Hallale N., Moore I. and Vauk D. (2003), ‘Hydrogen optimization at minimal investment’, Petroleum Technology Quarterly, 2, 83–90. Hammond E.G., Johnson L.A., Su C., Wang T. and Whit P.J. (2005), Bailey’s Industrial Oil and Fat Products, Sixth Edition,. Fereidoon Shahidi. Ed, John Wiley & Sons, Inc. Hospido A., Vasquez M.E., Cuevas A., Feijoo G. and Moreira M.T. (2006), ‘Environmental assessment of canned tuna manufacture with a life-cycle perspective’, Resources, Conservation and Recycling, 47, 56–72. Hostrup M., Harper P.M. and Gani R. (1999), ‘Design of environmentally benign processes: Integration of solvent design and separation process synthesis’, Computers and Chemical Engineering, 23(10), 1395–1414.

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 483 Huan Z. (2011), ‘Energy saving opportunities for food refrigeration’, Presentation from Department of Mechanical Engineering, Vaal Technical University, Available from: http://www.solansystems.com/articles/Energy%20saving%20opportunities.pdf Johns W.R. (2001), ‘Process synthesis: Poised for a wider role’, Chemical Engineering Progress, 4(2001), 59–65. Kemp I.C. (2005), ‘Reducing dryer energy use by process integration and pinch analysis’, Drying Technology, 23(9–11), 2089–2104. Klemes J., Smith R. and Kim J.K. (2008), Handbook of Water and Energy Management in Food Processing, Woodhead-Publishing Ltd. Kniel G.E., Delmarco K. and Petrie J.G. (1996), ‘Life cycle assessment applied to process design: Environmental and economic analysis and optimization of a nitric acid plant’, Environmental Progress, 15(4), 221–228. Kondilli E., Pantelides C.C. and Sargent R.W.H. (1993), ‘A general algorithm for short-term scheduling of batch operations- I. MILP formulation’, Computers and Chemical Engineering, 17(2), 211–227. Kookos I.K. and Perkins J.D. (2004), ‘The back-off approach to simultaneous design and control’, In Seferlis P. and Georgiadis M.C., eds., The Integration of Process Design and Control, Computer Aided Chemical Engineering, 17, 216–238. Kopanos G.M., Puigjaner L. and Georgiadis M.C. (2010), ‘Optimal production scheduling and lot-sizing in dairy plants: The yogurt production line’, Industrial and Engineering Chemistry Research, 49, 701–718. Krokida M.K. and Bisharat G.I. (2004), ‘Heat recovery from dryer exhaust air ’, Drying Technology, 22(7), 1661–1674. Lam H.L., Klemes J.J., Kravanja Z. and Varbanov P.S. (2010), ‘Software tools overview: Process integration, modelling and optimization for energy saving and pollution reduction’, Asia-Pacific Journal of Chemical Engineering, DOI: 10.1002/apj.469. Lewin D.R. (1996), ‘A simple tool for disturbance resiliency diagnosis and feedforward control design’, Computers and Chemical Engineering, 20(1), 13–25. Li Y., Griffing E., Higgins M. and Overcash M. (2006), ‘Life cycle assessment of soybean oil production’, Journal of Food Process Engineering, 29, 429–445. Linnhoff B. and Flower J.R. (1978), ‘Synthesis of heat exchanger networks: I. Systematic generation of energy optimal networks’, AIChE Journal, 24, 633–642. Liverman D.M., Hanson M.E., Brown B.J. and Merideth Jr R.W. (1988), ‘Global sustainability: Toward measurement’, Environmental Management, 12(2), 133–143. Luyben M.L. (2004), ‘Design of industrial processes for dynamic operability’, In Seferlis P. and Georgiadis M.C., eds., The Integration of Process Design and Control, Computer Aided Chemical Engineering, 17, 352–374. Marcoulaki E.C. and Kokossis A.C. (2000a), ‘On the development of novel chemicals using a systematic synthesis approach. Part I. Optimization framework’, Chemical Engineering Science, 55(13), 2529–2546. Marcoulaki E.C. and Kokossis A.C. (2000b), ‘On the development of novel chemicals using a systematic optimization approach. Part II. Solvent Design’, Chemical Engineering Science, 55(13), 2547–2561. Mattsson B. and Olsson P. (2001), ‘Environmental audits and life cycle assessment’, In Dillon M. and Griffith C., eds., Auditing in the Food Industry, 174–193, Woodhead Publishing Limited. Morison K.R. (2005), ‘Steady-state control of plate pasteurizers’, Food Control, 16, 23–30. Natural Resources Canada (2003), Pinch analysis: For the efficient use of energy, water and hydrogen, Technical Report, Available from: http://canmetenergy.nrcan. gc.ca/sites/canmetenergy.nrcan.gc.ca/files/files/pubs/2009-052_PM-FAC_404DEPLOI_e.pdf

© Woodhead Publishing Limited, 2013

484 Robotics and automation in the food industry Negiz A., Ramanauskas P., Cinar A., Sclesser J.E. and Armstrong D.J. (1998a), ‘Modeling, monitoring and control strategies for high temperature short time pasteurization systems – 2. Lethality-based control’, Food Control, 9, 17–28. Negiz A., Ramanauskas P., Cinar A., Sclesser J.E. and Armstrong D.J. (1998b), ‘Modeling, monitoring and control strategies for high temperature short time pasteurization systems – 3. Statistical monitoring of product lethality and process sensor reliability’, Food Control, 9, 29–47. Papadopoulos A.I. and Linke P. (2006a), ‘Efficient integration of optimal solvent and process design using molecular clustering’, Chemical Engineering Science, 61(19), 6316–6336. Papadopoulos A.I. and Linke P. (2006b), ‘Multiobjective molecular design for integrated process-solvent systems synthesis’, AIChE Journal, 52(3), 1057–1069. Papadopoulos A.I. and Linke P. (2009), ‘Integrated solvent and process selection for separation and reactive separation systems’, Chemical Engineering and Processing: Process Intensification, 48(5), 1047–1060. Papadopoulos A.I., Stijepovic M. and Linke P. (2010a), ‘On the systematic design and selection of optimal working fluids for Organic Rankine Cycles’, Applied Thermal Engineering, 30(6–7), 760–769. Papadopoulos A.I., Stijepovic M., Linke P., Seferlis P. and Voutetakis S. (2010b), ‘Power generation from low enthalpy geothermal fields by design and selection of efficient working fluids for organic Rankine cycles’, Chemical Engineering Transactions, 21, 61–66. Phongpipatpong M. and Douglas P.L. (2003), ‘Synthesis of rice processing plants. II. MINLP optimization’, Drying Technology, 21(9), 1611–1629. Pomerleau D., Desbiens A. and Barton G.W. (2003), ‘Real time optimization of an extrusion cooking process using a first principles model’, IEEE Conference on Control Applications, 1, 712–717. Rong A., Akkerman R. and Grunow M. (2011), ‘An optimization approach for managing fresh food quality throughout the supply chain’, International Journal of Production Economics, 131, 421–429. Sakizlis V., Perkins J.D. and Pistikopoulos E.N. (2004), ‘Simultaneous process and control design using mixed integer dynamic optimization and parametric programming’, In Seferlis P. and Georgiadis M.C., eds., The Integration of Process Design and Control, Computer Aided Chemical Engineering, Elsevier, 17, 187–215. Saling P., Maisch R., Silvain M. and Konig N. (2005), ‘Assessing the environmental-hazard potential for life-cycle assessment, eco-efficiency and SEEBalance’, International Journal of LCA, 10(5), 364–371. Sargent R.W.H. (1998), ‘A functional approach to process synthesis and its application to distillation systems’, Computers and Chemical Engineering, 22, 31–45. Schmidt I., Meurer M., Saling P., Kicherer A., Reuter W. and Gensch C.-O. (2004), ‘SEEbalance ®: Managing sustainability of products and processes with the socio-eco-efficiency analysis by BASF’, Greener Management International, 45, 79–94. Schweiger C.A. and Floudas C.A. (1997), ‘Integration of design and control: Optimization with dynamic models’, In Hager W.W. and Pardalos P.M., eds., ‘Optimal Control: Theory, Algorithms and Applications’, Kluwer Academic Pub., 388. Seferlis P. and Georgiadis M.C. The Integration of Process Design and Control , Computer Aided Chemical Engineering,17, 1–9. Seferlis P. and Grievink J. (2004), ‘Process design and control structure evaluation and screening using nonlinear sensitivity analysis’, In Seferlis P. and Georgiadis M.C., eds., The Integration of Process Design and Control, Computer Aided Chemical Engineering, Elsevier, 17, 326–351.

© Woodhead Publishing Limited, 2013

Automation for a sustainable food industry 485 Seferlis P. and Voutetakis S.S. (2008), ‘Measurement and process control for water and energy use in food industry’, In Klemes J., Smith R. and Kim J.K., eds., Handbook of Water and Energy Management in Food Processing, Woodhead-Publishing Ltd., 387–418. Stefanis S.K., Livingston A.G. and Pistikopoulos E.N. (1997), ‘Environmental impact considerations in the optimal design and scheduling of batch processes’, Computers and Chemical Engineering, 21(10), 1073–1094. Subramanian S. and Georgakis C. (2005), ‘Methodology for the steady-state operability analysis of plantwide systems’, Industrial and Engineering Chemistry Research, 44(20), 7770–7786. Sugiyama H., Fischer U., Hungerbulher K. and Masahiko H. (2008), ‘Decision framework for chemical engineering process design including different stages of environmental, health and safety assessment’, AICHE Journal, 54(4), 1037–1053. Sustainable Measures (2010), Introduction to Sustainable Development, Technical report, Available from: http://www.sustainablemeasures.com/node/42, 2010. Thevendiraraj S., Klemeš J., Paz D., Aso G. and Cardenas G.J. (2003), ‘Water and wastewater minimisation study of a citrus plant’, Resources Conservation and Recycling, 37(3), 227–250. Townsend D.W. and Linnhoff B. (1983), ‘Heat and power networks in process design, Part I: Criteria for placement of heat Engines and heat pumps in process networks’, AIChE Journal, 29(5), 742–748. Tsoka C., Johns W.R., Linke P. and Kokossis A. (2004), ‘Towards sustainability and green chemical engineering: Tools and technology requirements’, Green Chemistry, 6, 401–406. Wang Y.P. and Smith R. (1995), ‘Wastewater minimization with flowrate constraints’, Chemical Engineering Research and Design, 73(A8), 889–904.

© Woodhead Publishing Limited, 2013

Index

adapted systemic analysis, 205–7 methodology for expert knowledge handling, 206 type of formats for the variables of interest, 207 adaptive-network-based fuzzy inference system (ANFIS), 16 advanced process controller, 7–16 model-based control, 8–10 neural network-based control, 10–15 neuro-fuzzy control, 15–16 process modelling, 7–8 Advanced Robotic Technology for Efficient Pork Production (ARTEPP), 317 agriculture wireless sensor networks (WSN), 171–95 applications, 184–94 development, 172–84 future trends, 195 air, 295–7 continuous chilling/freezing tunnel, 296 continuous spiral chiller/freezer, 297 schematic of impingement chilling/ freezing system, 296 Air Cycle, 301 Airgas, 299 almonds, 279–80 examples of good almonds, hulls, in-shell and hard shell, 280 analogue/digital (A/D) data acquisition system, 422 animal behaviour monitoring, 189–90 animal environment monitoring, 190–1

animal health monitoring, 190 Ant Colony Algorithms, 450–1 area camera, 75–6 ARTEPP pork primalisation robot, 317–19 artificial intelligence, 470 artificial neural network (ANN), 5, 10–11, 13, 247 Aspen Hysys, 478 Aspen OnLine, 478 Aspen Plus, 478 AUTO-FOM, 316 automated carcass weighing systems, 316 automated chill rooms, 316 automated guided vehicles (AGV), 350, 434 automated inspection, 315 optical sensors and online spectroscopy for food products quality and safety, 111–27 applications in the food industry, 117–26 future trends, 126 optical sensing and spectroscopic techniques, 112–17 automated sheep boning system, 322 automated sustainable design tools applications combining multiple tools, 464–6 optimisation-based scheduling for optimum food distribution systems, 465 scheduling and design of cheese manufacturing processes, 465–6

© Woodhead Publishing Limited, 2013

488 Index automated sustainable design tools (cont.) soybean oil production, 464–5 food engineering operations, 446–66 applications combining multiple tools, 464–6 life cycle assessment (LCA), 461–4 real-time optimisation and control, 457–61 process and product optimisation, 446–52 case studies, 451–2 scope and description, 446–51 process integration, 452–7 scope and description, 446–51 decision parameters, 449 major algorithmic features involved in optimisation, 448 model development, 447–9 optimisation and algorithms, 450–1 performance indicators and constraints, 449–50 remarks, 451 automatic control batch thermal processing of canned foods, 420–39 computer-based control systems validation, 432–3 industrial automation of batch retorts, 433–7 on-line control strategies, 421–32 research advances and future trends, 437, 439 food chilling and freezing, 288–302, advances in research and future trends, 301 automation in food cold storage systems, 299–301 automation in refrigerated food retail display, 290–2 automation in refrigerated food transport systems, 292–4 automation of refrigeration and freezing operations in food catering, 292 best estimate of top ten food refrigeration processes in UK, 290 sequence of events within a typical cold-chain, 289 automatic horizontal plate freezers, 297–8 automatic process control food industry, 3–18 future trends, 16–18 adaptive-network-based fuzzy inference system, 16

Internet enabled systems, 17 soft computing-based systems, 17–18 wireless sensors and sensor networks, 16 process control methods, 5–16 advanced process controller, 7–16 proportional-integral-derivative (PID) controller, 5–6 process control systems and structure, 4 automation, 25–6, 201 batch processing on a bakery assembly line, 26 carcass production processes after primary chilling, 316–24 boning, 320–3 primal cutting, 316–20 slicing and portioning, 323 trimming, 323–4 carcass production processes before primary chilling, 309–16 automated chill rooms, 316 automated inspection and grading, 315 evisceration and dressing, 313–15 hair or hide removal, 311–13 lairage, 309–10 shackling, 311 splitting, 315 stunning, sticking and killing, 310–11 computer aided analysis and control engineering methods in food industry, 441–80 advanced tools and methods with potential applications, 466–78 automated sustainable design tools and food engineering operations, 446–66 food manufacturing, 445–6 future trends, 479–80 software technologies, 478–9 sustainability definition and links with food industry, 443–5 current technology and future trends in robotics in poultry industry, 329–52 bulk packing and poultry meat shipping, 347–51 characteristics and associated challenge, 329–31 future trends, 351–2 live hanging and first poultry processing, 331–4 second poultry processing, 334–47 delta robots performing pick-and-place operations, 27

© Woodhead Publishing Limited, 2013

Index 489 high-speed processing on a high-volume pancake line, 26 robotics and packaging in confectionery industry, 401–16 confectionery market and its business requirements, 402–6 future trends, 414–16 reconfigurable mechanism technology, 407–8 reconfigurable system case study for carton folding, 408 robotics for bulk sorting in food industry, 267–86 current applications, 275–85 future trends, 286 operation principles, 268–72 recent advances in technology, 273–5 requirements, 272–3 robotics in fresh produce industry, 385–99 information flow for food traceability and farming guidance, 396–8 machine vision system, 386–9 vegetable reprocessing and grading systems, 389–96 robotics in meat processing, 304–26 future trends, 324–5 robotics in seafood processing, 354–82 fish slaughtering, filtering, portioning and unit operation application, 366–73 future trends, 377–82 other unit operations in fish processing, 373–7 technologies, 359–66 SCADA for process control in food industry, 130–41 food processing, 136–9 future trends, 140–1 history, 133 implementation, 139–40 overview, 130–2 standards and applications, 133–5 sensor for food process control, 36–73 applications, 67–72 device integration, 60–7 food instrumentation considerations, 37–44 future trends, 72–3 measurement methods, 44–60 trends in capital costs of robots vs. European labour costs, 27 Baader Food Processing Machinery, 368 bacterial contamination, 40, 146

batch frying, 10 batch thermal processing automatic control of canned foods, 420–39 computer-based control systems validation, 432–3 industrial automation of batch retorts, 433–7 on-line control strategies, 421–32 research advances and future trends, 437, 439 beef evisceration, 314 beef primalisation, 319–20 Bernoulli gripper, 159, 160–1 gripping force, 162 working principle, 161 bioconversion, 227 fed-batch reactor, 226–61 design of PID controllers, 239–46 dynamical model, 229–30 population balance modelling in food processes, 231–3 real-time optimisation, 246–60 tuning of observer-based estimators, 233–9 boning, 320–3 beef foreleg deboning research system, 321 boundary tracking, 81–2 breast deboning, 339–44 bird on bone, 340 intelligent cutting work cell, 341 robot for washdown environment, 343 tray packing robot, 342 browning global appearance (BGA), 214 bulk packing robotics and automation in poultry meat shipping, 347–51 ABB Flexpicker, 350 key points on bird for cutting, 349 poultry deboning line, 348 bulk solid flowmeters, 54 bulk sorting automation and robotics in food industry, 267–86 requirements, 272–3 current applications, 275–85 fruits and vegetables, 281–5 grain, 276–9 nut sector, 279–81 future trends, 286 food safety, 286 food security, 286 operation principles, 268–72

© Woodhead Publishing Limited, 2013

490 Index bulk sorting (cont.) ejection system, 272 image processor, 271–2 optical systems, 271 overview, 268–9 schematic diagram of main components of machine, 270 recent advances in technology, 273–5 data processing, 274–5 new materials, 273–4 sensors, 274 CAFE, 231 canned foods automatic control of batch thermal processing, 420–39 computer-based control systems validation, 432–3 on-line control strategies, 421–32 research advances and future trends, 437, 439 industrial automation of batch retorts, 433–7 automated shuttle batch retort system, 438 batch retort systems with use of AGVs in large cook room operation, 436 FMC shuttle system for automated batch retort loading/unloading, 438 guided vehicle for batch retort loading/unloading, 436 new retort systems, 434 rack designs for flexible and semi-rigid retortable packaging systems, 435 shuttle-based batch retort control system, 437 capacitance level transmitters, 55 Cartesian robot, 317 carton folding case study of reconfigurable system, 408, 410–14 demonstrator system testing, 413–14 motion control for demonstrator system, 411–12 prototype for folding carton trays, 410–11 reconfigurable demonstrator system, 408, 410 cereals insect location, 99, 101–3 bar and end detector application, 101–2

complete contaminant detection system, 103 non-insect contaminants location, 95–9 location of ergot amongst wheat grains, 100 morphological approach, 96–9 morphology to recover rat dropping shapes, 97–8 thresholding approach, 99 charge coupled devices (CCDs), 387 cheese mass, 220 chemical contamination, 39 chilled cabinets temperature, 291 chilling automatic control for freezing, 288–302, advances in research and future trends, 301 automation in food cold storage systems, 299–301 automation in refrigerated food retail display, 290–2 automation in refrigerated food transport systems, 292–4 automation of refrigeration and freezing operations in food catering, 292 automation in freezing systems, 294–9, 295–9 air, 295–7 direct contact, 297–8 high-pressure freezing, 299 immersion/spray, 298–9 vacuum cooling, 299 chocolate cakelet inspection, 92–4 CIMPLICITY, 139 Citect, 140 closed-loop system, 4 Coanda effect gripper, 158–9 colour cameras, 387 come-up-time (CUT), 429 Common Format Files (CFF) Commonwealth Scientific and Industrial Research Organisation (CSIRO), 310 computer-aided analysis control engineering methods and, automation for sustainable food industry, 441–80 advanced tools and methods with potential applications, 466–78 automated sustainable design tools and food engineering operations, 446–66 food manufacturing, 445–6

© Woodhead Publishing Limited, 2013

Index 491 future trends, 479–80 software technologies, 478–9 sustainability definition and links with food industry, 443–5 computer-aided molecular design (CAMD), 473–6 description, 475–6 algorithmic stages of optimisation, 476–8 CAMD features, 475 remarks, 475 introduction and potential applications, 473–5 design and selection of efficient refrigerants and heat recovery fluids, 474–5 design and selection of efficient solvents, 474 computerised tomography (CT), 388–9 conductance level switches, 55 cone loading, 338–9 gripper, 338 robot placing front-halve on cone, 339 confectionery industry business requirements and commercial viability, 403–6 conventional packaging machine, 405 future trends, 414–16 environmental impacts reduction, 415–16 reconfiguration, 414–15 market and its business requirements, 402–6 business requirements and commercial viability, 403–6 market size, 402 packaging usage, 402–3 market size, 402 UK consumption of packaged confectionery, 403 packaging usage, 402–3 UK market for confectionery by product sector, 404 variety of confectionery packs, 403 robotics, automation and packaging, 401–16 reconfigurable mechanism technology, 407–8 reconfigurable system case study for carton folding, 408, 410–14 connected-components analysis, 81 contact refrigeration methods, 297 continuous frying, 10 control engineering methods

computer-aided analysis and automation for sustainable food industry, 441–80 advanced tools and methods with potential applications, 466–78 automated sustainable design tools and food engineering operations, 446–66 food manufacturing, 445–6 future trends, 479–80 software technologies, 478–9 sustainability definition and links with food industry, 443–5 control system, 205–10, 364–5 building steps using fuzzy logic, 206 controlled atmosphere, 294 convolution mask, 77–9 coriolis flowmeters, 52–3, 59 corner enhancement mask, 80 cream biscuit inspection, 94–5 appearance, 94 Hough transform for locating biscuit centre, 95 cryogenic freezing systems, 299 crystallisation, 231 Danish Carcass Classification Centre, 316 Danish Meat Research Institute (DMRI), 308 dark, firm, dry (DRD) meat, 309 data-driven approach, 216–17 data management, 183–4 data processing, 274–5 clump of beans silhouette and result of applying object separation, 275 de-boning, 370–1 de-fleecing sheep, 312–13 de-hairing pork, 311 de-heading, 369–70 fish processing line from Baader including filleting and trimming, 370 de-hiding beef, 311–12 decision-aid system, 208–9 decision support system (DSS), 398 deep penetrating needle gripper, 155–6 long needle gripper, 156 modular long needle gripper, 156 degrees-of-freedom (DOF), 394 delta robots, 25 density measurement, 59 Department for Environmental Food and Rural Affairs (DEFRA), 444 device database files (GSD), 65

© Woodhead Publishing Limited, 2013

492 Index differential pressure flowmeters, 54 digital graphical replay (DGR), 139 digital signal processing, 104 digital signal processor, 271 discolouring, 146 distributed control systems (DCS), 65 dosing, 69–70 flowmeter-based dosing system, 70 dryer inlet air, 455–6 Dust-Ex, 42 dynamic matrix control (DMC), 10 dynamic model, 229–30 Haldane model, 230 dynamic operability description and case study, 471–3 base-case flowsheets for separation of soybean oil from extraction solvent, 472 expected challenges, 471–2 potential improvement of base-case flowsheet for energy reduction, 472 solution methods, 473 integration, 470–3 description and case study, 471–3 introduction and scope, 471 edge enhancement mask, 80 efficient object location, 86–8 Hough transform for biscuit location, 87 Hough transform principle, 87 eggplant grading, 393 colour images, 394 inspection line of eggplant grading system, 393 monochrome, 394 ejection system, 272 electrical flowmeters, 52–4 coriolis flowmeters installed on a filling machine, 53 electromagnetic flowmeters installed in a brewery, 52 electrical pressure transmitters, 46 operating conditions for pressure-measuring devices, 47 electrical temperature device, 49–50 temperature-measuring assembly components, 50 electromagnetic flowmeters, 52 Electronic Device Description Language (EDDL), 60 electronic device descriptions (EDD), 65 enclosed production cells, 380–1 enclosing gripper, 147, 148–54, 152

hygienic performance, 152–4 mechanism, 153 standard 2- and 3-finger gripper, 149 end-of-line packaging, 23 energy harvesting, 182–3 energy management, 71–2 system sample, 71 Enterprise Resource Planning (ERP), 365 environment monitoring, 184–5 Ethernet/IP, 66 fieldbus system device integration, 66 evisceration dressing, 313–15 beef evisceration, 314 pork evisceration, 313–14 sheep evisceration, 315 exhaust stream splitting, 456 expert knowledge-driven approach, 217 extraneous vegetable matter (EVM), 95, 282 Fanuc M710 robot, 320 fed-batch reactor bioconversion methods for food processes control, 226–61 basic dynamic model, 229–30 PID controllers design, 239–46 population balance modelling in food processes, 231–3 real-time optimisation, 246–60 tuning of observer-based estimators, 233–9 feed-forward network, 11–12 feedback recurrent network, 13, 15 fibre optic sensor, 116–17, 120–1 Field Device Tool (FDT), 60 Field Programmable Gate Array (FPGA), 271 filleting, 369–70 filling, 69–70 first processing, 333–4 fish preparation, 368–9 flow measurement, 51–4 areas of application of electrical flowmeters, 51 Food and Drug Administration, 432 food catering, 292 food cold storage systems, 299–301 food industry, 117–26, 211–19 automatic process control, 3–18 future trends, 16–18 process control methods, 5–16 process control systems and structure, 4

© Woodhead Publishing Limited, 2013

Index 493 automation and robotics for bulk sorting, 267–86 current applications, 275–85 future trends, 286 operation principles, 268–72 recent advances in technology, 273–5 requirements, 272–3 image sensing technology applications, 121–6 illustration, 125 machine vision, 75–107 application advances, 103–4 applications and case studies, 92–103 food inspection applications hardware, 104–5 future trends, 105–6 principles and methods, 77–92 optic fibre sensing techniques applications, 120–1 illustration, 124 robotics, 21–35 automation in food sector, 25–6 current manufacturing procedures, 23–5 food sector robot specification, 26–32 future trends, 32–4 overview, 21–3 SCADA and automated process control in food industry, 130–41 food processing, 136–9 future trends, 140–1 history, 133 implementation, 139–40 overview, 130–2 standards and applications, 133–5 spectroscopic techniques applications, 118–20 illustration, 124 wireless sensor networks (WSNs), 171–95 applications, 184–94 development, 172–84 future trends, 195 food industry robots gripper, 143–69 food process automation, 144–6 future trends, 169 penetrating (needle), 154–7 physics, 147–8 pinching and enclosing, 148–54 suction, 157–63 surface effect (freeze), 163–7 technology selection, 167–9 food inspection, 104–5

food instrumentation, 37–44 instrument housing, 41–4 process connections, 41 regulatory agencies, 37–8 wetted parts, 38–40 food material, 145 food process control, 216–19 feedback control, 219 fuzzy symbolic approach applied to biscuit baking, 218 SCADA and automation in food industry, 130–41 food processing, 136–9 future trends, 140–1 history, 133 implementation, 139–40 overview, 130–2 standards and applications, 133–5 sensor automation, 36–73 applications, 67–72 device integration, 60–7 food instrumentation considerations, 37–44 future trends, 72–3 measurement methods, 44–60 food processing advanced control methods and fed-batch reactor bioconversion, 226–61 basic dynamic model, 229–30 PID controllers design, 239–46 population balance modelling in food processes, 231–3 real-time optimisation, 246–60 tuning of observer-based estimators, 233–9 intelligent quality control system based on fuzzy logic, 200–22 applications in food industry, 211–19 principles, 203–11 research and future trends, 220–2 SCADA, 136–9 cost vs. benefit, 136–8 installation and maintenance costs estimates, 137 security and operation, 138–9 food products optical sensor and online spectroscopy for food quality and safety inspection, 111–27 applications in the food industry, 117–26 future trends, 126 optical sensing and spectroscopic techniques, 112–17

© Woodhead Publishing Limited, 2013

494 Index food quality safety, optical sensors and online spectroscopy automated inspection of food products, 111–27 applications in the food industry, 117–26 future trends, 126 optical sensing and spectroscopic techniques, 112–17 symbolic variables indirect measurement, 211–16 BGA diagnosis and evaluation comparison, 215 crusting degree of a sausage using the fuzzy meaning concept, 213 expert measurements to predict global browning appearance during heating, 216 length of black areas (LB) meaning, 214 sensory indicator percentage of spotted area (V2), 215 Food Refrigeration and Process Engineering Research Centre (FRPERC), 294 food safety, 286, 306, 345 food security, 286 food traceability system, 192–4 obstacles and issues, 193–4 foreign body, 39 FOUNDATION fieldbus, 65–6 fieldbus system device integration, 65 Fourier transform infrared (FTIR) spectroscopy, 114, 118 freeze gripper, 163–7, 364 design, 164 Peltier element gripping attachment times and temperatures, 164 hygienic performance, 167 mechanical release, 164–5 illustration, 165 freezing, 374, 376 automatic control for food chilling, 288–302, advances in research and future trends, 301 automation in food cold storage systems, 299–301 automation in refrigerated food retail display, 290–2 automation in refrigerated food transport systems, 292–4 automation of refrigeration and freezing operations in food catering, 292

fresh produce industry information flow for food traceability and farming guidance, 396–8 data flow from grading robot, 397 data management by agricultural cooperative association, 398 various information accumulated in a database, 396 robotics and automation, 385–99 machine vision system, 386–9 vegetable reprocessing and grading systems, 389–96 fruit grading robot, 393–6 grading of eggplant, 393 grading of oranges, 391–3 peeling and grading of leeks, 390–1 fruits, 284–5 bulk to bulk, 283 end of line, 283–4 optical sorting machine designed for bulk to bulk sorting, 284 vegetables with foreign material as seen by visible camera and InGaAs camera, 285 fresh product, 282 reject example from sorting fresh peas, 283 frozen product, 282–3 grading robot, 393–6 deciduous fruit, 394 top views of apples, peaches and pears, 394 fully robotic solution, 335 fungus contamination, 146 Fututech system, 312 fuzzy description, 208 fuzzy function, 208, 221, 222 fuzzy k-nearest neighbour method, 207 fuzzy logic, 5 applications in food industry, 211–19 cheese mass loss during ripening modelling, 220 cheese mass loss probabilities during ripening, 221 state diagnosis and food process control, 216–19 symbolic variables indirect measurement of food product quality, 211–16 intelligent control system principles, 203–11 control system, 205–10 fuzzy membership function, 204

© Woodhead Publishing Limited, 2013

Index 495 imprecise knowledge representation and propagation, 210–11 parametric representation of a membership function, 205 intelligent quality control system in food processing, 200–22 research and future trends, 220–2 fuzzy meaning, 208, 212–13, 214–15, 217, 222 fuzzy set, 202, 203–4, 207 GAMS, 478 general dynamic model, 233 General Method, 422 general packet radio service (GPRS), 294 genetic algorithms (GA), 450–1 geographical information system (GIS), 398 gPROMS, 478 grain, 276–9 defect categories examples found in wheat, 279 layout for three pass sort of rice, 277 sample of input to rice, accept from rice and reject from rice sort, 278 six optical sorting machines installation in a rice mill, 276 Graphical User Interface (GUI), 272 greenhouse management, 191–2 gripper, 28–30 complex origami shape in thin cardboard using a dexterous gripper design, 30 cylindrical gripper for acquisition and precision placing of thin pasta slices, 30 food industry robots, 143–69 food process automation, 144–6 hygienic requirements, 146 non-uniform shapes, 145–6 soft materials, 145 uneven surfaces, 145 future trends, 169 penetrating (needle), 154–7 physics, 147–8 picking and placing of thin cucumber and tomato slices on a sandwich, 29 pinching and enclosing, 148–54 suction, 157–63 surface effect (freeze), 163–7 technology selection, 167–9 qualitative performance parameters, 168

gripping tools, 359–64 characteristics and behaviour, 360–1 gripper developments, 361–4 compact needle gripper, prototype design, 363 needle gripper with crossing needles, prototype design, 363 pick and place gripper device, 362 three compact freeze gripper units, 364 growth model, 444 gutting, 368–9 hair removal, 311–13 de-fleecing sheep, 312–13 de-hairing pork, 311 de-hiding beef, 311–12 Haldane kinetics model, 230, 240, 248 Hamiltonian system, 241 hard automation solution, 25 hard-wired electro-mechanical system, 25 hardware platforms, 173, 177 HART, 61 device integration into a control and asset-management system, 64 Hazard Analysis and Critical Control Points (HACCP), 432 heat transfer rates, 295–6 high-pressure freezing, 299 hydrostatic level transmitters, 55, 57 hyperspectral imaging, 122–3, 125 image modality, 90–2 image-processing, 77–81 basic mask properties, 79 thresholding and edge detection, 78 image processor, 271–2 imaging sensor, 117, 121–6 immersion/spray systems, 298–9 automatic immersion freezing system for whole fish, 298 indium gallium arsenide (InGaAs), 274 Industrial Research Ltd (IRL), 311 infrared (IR) spectroscopy, 113–14, 118 infrared (IR) wavelengths, 271 Institut national de la recherche agronomique (INRA), 319 instrument housing, 41–4 degree of protection of enclosures, 43 ingress protection categories, 43 protection for gas and dust (D) atmospheres, 44 Integrated Controlled Random Search for Dynamic Systems, 451

© Woodhead Publishing Limited, 2013

496 Index intelligent quality control system applications in food industry, 211–19 cheese mass loss during ripening modelling, 220 state diagnosis and food process control, 216–19 symbolic variables indirect measurement of food product quality, 211–16 fuzzy logic in food processing, 200–22 principles, 203–11 control system, 205–10 fuzzy membership function, 204 imprecise knowledge representation and propagation, 210–11 parametric representation of a membership function, 205 research and future trends, 220–2 internal model control (IMC), 9–10 International Organization for Standardization (ISO), 293 internet enabled system, 17 interpreter device type manager (iDTM), 66 inventory management, 68–9 system sample, 69 IP rating, 41 ISA SP100.11a, 66, 67 Kalman observer, 235 killing see stunning knowledge representation possibility distributions, 210–11 heat transfer coefficient parameters in cheese ripening model, 212 probability distributions, 210 frequencies and probability law, 211 KUKA robot, 312, 318 KUKA 15SL, 335 lairage, 309–10 LaSalle-Yoshizawa theorem, 253 leeks peeling and grading, 390–1 acquired images of leek by colour and monochrome cameras, 391 preprocessing system, 391 root image before cutting at plate between stalk and roots, 390 leg boning machine, 322 legislation, 306 level measurement, 55–9 guided radar measurement in a flour silo, 58

hydrostatic level measurement on a beer-fermentation tank, 57 properties of level transmitters used in the food industry, 56 life cycle assessment (LCA), 461–4 case studies, 464 scope and definition, 461–3 methodology, 461–3 remarks, 463 light emitting diodes (LED), 271, 391 line-scan camera, 76 line-scan system, 123 linear programming (LP), 450 live hang, 331–3 live transfer, 332 LogMIP, 478–9 long pins, 147 Luenberger observer, 235 Lyapunov transformation, 238 machine vision system, 386–9 advances in application, 103–4 applications and case studies, 92–103 colour and fluorescent image of rotten orange fruit, 388 CT image and actual damaged apple, 389 food industry, 75–107 food inspection applications hardware, 104–5 future trends, 105–6 orange fruits images by colour camera and x-ray camera, 389 principles and methods, 77–92 basic image-processing techniques, 77–81 efficient object location, 86–8 image modalities, 90–2 morphology, 83–6 shape analysis, 81–3 statistical pattern recognition, 89–90 texture analysis, 88–9 sensor-related technology, 365–6 magnetic refrigeration, 301 manual intensive manufacturing, 24 master terminal unit (MTU), 131 mathematical knowledge formalisation, 207–8 fuzzy meaning describing the moisture content of a cheese, 208 Meat Industry Research Institute of New Zealand (MIRINZ), 308 meat processing barriers, 306–8

© Woodhead Publishing Limited, 2013

Index 497 commercial and organisational challenges, 306–7 technical challenges, 307–8 robotics and automation, 304–26 carcass production processes after primary chilling, 316–24 carcass production processes before primary chilling, 309–16 current status, 308–9 drivers, 305–6 future trends, 324–5 scope, 304–5 mechanical automation, 291–2 mechanical finger gripper, 364 microwave barriers, 54 milk pasteurisation process, 459–60 conceptual control diagram, 460 MIPSYN, 478–9 mixed integer linear programming (MILP), 450 mixed integer non-linear programming (MINLP), 450 mobile racking, 300 MODBUS serial, 61 field device integration, 64 MODBUS TCP, 61 field device integration, 64 model-based control, 8–10 IMC control loop, 9 PDC control loop, 11 moderate underpressure high-flow gripper, 158–9 Coanda effect gripper, 160 obtainable underpressure for a Coanda gripper, 160 monitoring and targeting (M&T), 72 monitoring issues observer-based estimators tuning, 233–9 experimental results on baker’s yeast production process, 240 monochrome camera, 387 morphology, 83–6 erosion and dilation, 84 scratch detection, 85 multi-finger gripper, 151–2 multi-input–multi-output (MIMO), 11 near-infrared (NIR) spectroscopy, 115–16, 118–19 nearest neighbour method, 89 needle gripper hygienic performance, 157 lifting performance, 156–7

NEMA rating, 41 NEMA 4X enclosure, 41 network architecture, 181–2 single-hop and multi-hop communication, 181 wireless network topology, 182 neural network-based control, 10–15 architecture of a feed-forward ANN, 12 architecture of a feedback recurrent ANN, 14 continuous snack-food frying process, 12 IMC control operation based on the one-step-ahead prediction, 13 IMC controller tuning colour and moisture response, 14 neuro-fuzzy control, 15–16 noise suppression mask, 79 non-contact grippers, 361–2 non-linear programming (NLP), 450 nut, 279–81 almonds, 279–80 walnuts, 281 observer-based estimators, 233–9 on-line control strategies, 421–32 deviations without extending process time and computer simulation software, 427–32 experimental validation: case study, 427–32 iso-lethality curve showing process time and retort temperatures, 430 process conditions and parameters used for heat penetration test, 431 TRT and internal product cold spot temperatures over time, 430 mathematically modelled foods, 421–5 dynamic process simulation during ramp-down portion, 424 dynamic process simulation during ramp-up portion, 424 on-line correction of temperature process without compromising quality, 423–5 on-line correction without change in process time, 422–3 pure conduction simulation, 422 retort temperature with minimum extended process time at recovered temperature, 425–7 proportional correction strategy development, 425–7 pure conduction simulation, 426

© Woodhead Publishing Limited, 2013

498 Index online spectroscopy applications in the food industry, 117–26 automated quality and safety inspection of food products, 111–27 future trends, 126 spectroscopic techniques, 112–17 electromagnetic spectrum, 113 open-loop system, 4 operating systems, 179–81 operation summary for WSN, 180 optical sensor applications in the food industry, 117–26 automated quality and safety inspection of food products, 111–27 future trends, 126 optical sensing, 112–17 optical sorting, 267–8 current applications, 275–85 fruits and vegetables, 281–5 grain, 276–9 nut sector, 279–81 operation principles machine designed for sorting dry products, 269 machine designed for sorting wet products, 270 optical systems, 271 optimisation-based CAMD, 479 optimisation technologies, 447 oranges grading, 391–3 schematic diagram of camera and lighting setup for citrus fruit grading, 392 Organic Rankine Cycle processes, 473 packaging automation and robotics in confectionery industry, 401–16 confectionery market and its business requirements, 402–6 future trends, 414–16 reconfigurable mechanism technology, 407–8 reconfigurable system case study for carton folding, 408 pale, soft, exudative (PSE) meat, 309 pallet loading process, 349 palletising, 23 penetrating gripper, 154–7 persistence of excitation (PE), 247 pick-and-place operations, 373–4 I-Cut 300 portioning machine from Marel and cut of white fish, 374

new processing line for sardines or shrimps from Cabinplant, 375 process line from Marel where fish fillets are portioned, freezed and packed, 376 solution from Marel of pick-and-place operation of salmon portions, 374 Pinch technology, 452–3 composite curves and pinch in temperature–enthalpy diagram, 454 pinching gripper, 147, 148–54 hygienic performance, 152–4 standard 2- and 3-finger gripper, 149 pinning gripper, 147 pixel–pixel operation, 77 Pontryagin Maximum Principle, 240 population balance, 231 pork evisceration, 313–14 developmental system, 314 pork primalisation, 317 portioning, 323 operations, 371–3 trimming and portioning operations at flow line, 372 possibility theory, 202 poultry industry current technology and future trends in robotics and automation, 329–52 bulk packing and poultry meat shipping, 347–51 characteristics and associated challenge, 329–31 future trends, 351–2 live hanging and first poultry processing, 331–4 second poultry processing, 334–47 Praxair, 299 pre-rigor processing, 377, 379 precision agriculture, 185–8 precision irrigation, 186–8 spatial data collection, 186 WSN for soil property monitoring, 187 precision livestock farming, 189–91 predictive control (PDC), 10 preservation, 227–8 pressure measurement, 45–6 pressure transmitter on a keg-filling line, 45 primal cutting, 316–20 beef primalisation, 319–20 case study: ARTEPP pork primalisation robot, 317–19 ARTEPP automation vs human cutting performance, 318

© Woodhead Publishing Limited, 2013

Index 499 production ARTEPP system, 319 pork primalisation, 317 sheep primalisation automation, 320 ProCAMD, 479 process connection, 41 sanitary couplings for the food industry, 42 process control, 208–9 process integration, 452–7 case studies, 455–7 composite curves in water Pinch analysis, 457 heat recovery from milk powder spray dryer exhausts, 455–6 water recovery and re-use opportunities in citrus plant, 456–7 scope and description, 452–5 extensions, 453–4 remarks, 454–5 the Pinch method, 453 software, 479 process synthesis software, 478–9 product consistency, 305 production quality, 305 PROFIBUS, 65 fieldbus system device integration, 67 programmable logic controllers (PLC), 130 proportional correction strategy, 426 proportional-integral-derivative (PID), 3, 4, 5–6, 239–46, 294 dissolved oxygen concentration PI control in industrial fed-batch bioreactor, 244 feedback closed-loop control scheme, 5 optimal control of fed-batch bioreactor, 243 output response of control system under step input, 6 PI control fed-batch bioreactor: constant gain, 245 PI control fed-batch bioreactor: proportional gain with an exponential term, 246 quality control inspection, 376–7 SensorX machine from Marel displaying x-ray images of fillets, 378 quantum well infrared photodetectors (QWIP), 114

radar level transmitters, 58–9 radiofrequency identification (RFID), 32, 173, 178–9, 294 Raman spectroscopy, 114–15, 118 RANSAC, 87 raw milk heating, 456 real-time optimisation, 246–60, 457–61 adaptive extremum-seeking control of fed-batch bioprocesses, 247–60 design controller, 250–3 estimation equation for gaseous outflow rate y, 248–50 stability and convergence analysis, 253 case studies, 459–61 milk pasteurisation process control, 459–60 optimal control in thermal sterilisation of canned foods, 461 real-time optimisation of extrusion cooking process, 460–1 dither signal design, 253–5 scope and definition, 457–9 schematic of automated control system, 458 schematic of regulatory control system, 459 simulation results, 255–60 convergence properties loop system illustration, 257 proposed scheme performance and controller, 258, 259 robustness properties illustration, 259, 260 receiver operating characteristic (ROC), 90 reconfigurable demonstrator system, 410 folding trajectories of faces and gussets of tray carton, 410 modules arrangement for tray folding, 411 motion control, 411–12 experiment flowchart, 412 timing diagram from control simulation of fingers and folders, 412 testing, 413–14 demonstrator in an industrial environment, 413 folding cartons by demonstrator in an industrial environment, 414 reconfigurable mechanism technology, 407–8 design and creation stages for confectionery packaging, 407–8

© Woodhead Publishing Limited, 2013

500 Index reconfigurable mechanism technology (cont.) design flowcharts, 409 modularity, 408 reconfigurability, 408 requirements, 408 work analysis, 407–8 refrigerated containers, 293 refrigerated food retail display, 290–2 refrigerated food transport systems, 292–4 regulatory agency, 37–8 regulatory and standardisation authorities, 38 regulatory control system, 459 rehang, 334–7 non-robotic rehang system first step, 336 second step, 337 robotic shackle loader, 335 reject ratio, 273 release gripper, 147 resistance temperature detectors (RTD), 49 retort temperature profile (TRT), 431 reversible heat flow gripper, 165–7 Peltier element freeze gripper holding a piece of salmon fillet, 167 Peltier element freeze gripper principle, 166 Robot Leonardo2000, 350 robotics automation and packaging in confectionery industry, 401–16 confectionery market and its business requirements, 402–6 future trends, 414–16 reconfigurable mechanism technology, 407–8 reconfigurable system case study for carton folding, 408 automation for bulk sorting in food industry, 267–86 current applications, 275–85 future trends, 286 operation principles, 268–72 recent advances in technology, 273–5 requirements, 272–3 automation in food sector, 25–6 automation in fresh produce industry, 385–99 information flow for food traceability and farming guidance, 396–8 machine vision system, 386–9 vegetable reprocessing and grading systems, 389–96

automation in meat processing, 304–26 carcass production processes after primary chilling, 316–24 carcass production processes before primary chilling, 309–16 future trends, 324–5 automation in seafood processing, 354–82 fish slaughtering, filtering, portioning and unit operation application, 366–73 future trends, 377–82 other unit operations in fish processing, 373–7 technologies, 359–66 current manufacturing procedures, 23–5 current technology and future trends in automation in poultry industry, 329–52 bulk packing and poultry meat shipping, 347–51 characteristics and associated challenge, 329–31 future trends, 351–2 live hanging and first poultry processing, 331–4 second poultry processing, 334–47 food industry, 21–35 food sector robot specification, 26–32 gripper technology, 28–30 sensor technology, 31–2 future trends, 32–4 design specifications, 33–4 Food Factory in a Pipe concept, 33 overview, 21–3 European retail market share, 22 market, 21–3 salting, 369 Scott Automation, 323 seafood processing current status and challenges, 355–7 robots shipments during 2003–2008, 356 future trends, 377–82 enclosed production cells, 380–1 fully automated processing lines and integrated control systems, 381–2 pre-rigor processing, 377–9 reconfigurable production lines to meet changing market trends, 379–80 super chilling, 379 sustainable manufacturing, 382

© Woodhead Publishing Limited, 2013

Index 501 traceability, 382 manual handling benefits with automated handling, 357–9 requirements for robotic handling vs manual handling, 359 other unit operations in fish processing, 373–7 chilling operations, 374, 376 freezing, 374, 376 further unit operations, 377 pick-and-place operations, 373–4 quality control and inspection, 376–7 robotics and automation, 354–82 fish slaughtering, filtering, portioning and unit operation application, 366–73 key drivers for enhanced automation, 354–5 technologies, 359–66 second poultry processing robotics and automation, 334–47 breast deboning, 339–44 cone loading, 338–9 rehang after chiller, 334–7 tray packing, 344–7 SEEbalance, 463 sensor, 31–2, 274 applications, 67–72 dosing and filling, 69–70 energy management, 71–2 inventory management, 68–9 automated food process control, 36–73 device integration, 60–7 4–20 mA/HART, 61 attributes of various control and fieldbus standards, 62–3 EtherNet/IP, 66 FOUNDATION fieldbus, 65–6 MODBUS serial/MODBUS TCP, 61, 64–5 PROFIBUS, 65 Wireless, 66–7 food instrumentation considerations, 37–44 future trends, 72–3 GRAIL food-grade prototype robot, 31 measurement methods, 44–60 analysis, 59–60 density, 59 flow, 51–4 level, 55–9 pH measurement in a brewery, 60 pressure, 45–6 temperature, 46, 48–51

sensor-based fixed-automation, 336 shackling, 311 shape analysis, 81–3 algorithm for connected-components analysis, 82 centroidal-profile method, 83 skeleton concept, 83 sheep evisceration, 315 sheep primalisation automation, 320 short needle skin penetrating gripper, 154–5 illustration, 155 short pins, 147 shuttle system, 436 Simulated Annealing (SA), 450–1 singulation system, 332–3 site hot water, 456 slaughtering operations, 366–7 robotised line for gill arch cut delivered by seaside, 367 slicing, 323 small-to-medium sized enterprises (SME), 22 sorption-adsorption systems, 301 sorting operations, 367–8 splitting, 315 SPRINT, 479 staff recruitment, 306 STAR, 479 state diagnosis, 216–19 feedback control, 219 fuzzy symbolic approach applied to biscuit baking, 218 statistical pattern recognition, 89–90 receiver operating characteristic (ROC), 91 sticking see stunning Stirling Cycle, 301 stirred tank reactors (STR), 235 strong underpressure gripper, 158 single suction cup and control system, 159 stunning, 310–11 suction gripper, 157–63 hygienic performance, 162–3 characteristics, 163 lifting performance, 162 underpressure range and lifting capacity, 163 super chilling, 379 supervisory control and data acquisition (SCADA) automated process control in food industry, 130–41 food processing, 136–9

© Woodhead Publishing Limited, 2013

502 Index supervisory control and data acquisition (SCADA) (cont.) future trends, 140–1 history, 133 implementation, 139–40 attributes of VTSCADA, CIMPLICITY, and CITECT, 141 overview, 130–2 benefits and drawback, 131–2 needs, 131 notional network for baking process, 132 standards and applications, 133–5 computer production, 135 installations in industry, 134 thermal power plant, 134 water treatment installation, 134–5 surface effect gripper, 163–7 sustainable food industry advanced tools and methods with potential applications, 466–78, 476–8 computer-aided molecular design (CAMD), 473–6 early attempts, 476–7 integration and dynamic operability, 470–3 process and product synthesis, 466–70 process synthesis and LCA optimisation and environmental impacts, 477 sustainability considerations in design, 478 automation, computer aided analysis and control engineering methods, 441–80 advanced tools and methods with potential applications, 466–78 food manufacturing, 445–6 future trends, 479–80 software technologies, 478–9 sustainability definition and links with food industry, 443–5 sustainable design tools and food engineering operations, 446–66 process and product synthesis, 466–70 description and case study, 467–70 introduction and scope, 467 remarks, 470 representation of decision options, 467–9 representation of process flowsheets as graphs, 468

solution of process synthesis problem, 469–70 superstructure representation using mixer-splitter model, 469 sustainable manufacturing, 382 temperature measurement, 46, 48–51 CIP pipe, 48 principle, 48 texture analysis, 88–9 busyness concept detection, 89 thermal mass flowmeters, 54 thermal power plant, 134 thermal processing, 452 Thermoacoustic, 301 thermocouple, 49 Thermoelectric, 301 time of flight (TOF), 339 Torry IQF Continuous Steel Band Freezer, 298 toxic contamination, 146 traceability, 382 translational velocity, 333 tray packing, 344–7 case loading robot from CAMotion, 346 estimated price index of industrial robots in US, 346 Trigeneration, 301 trimming, 323–4 operations, 370–1 Baader 988 S (Salmon trimming machine), 372 trio FDS 105 HS, 373 Trio Food Processing Machinery, 370 trunk boning machine, 322 tuned hydrostatic pressure transmitters, 59 two-finger gripper, 148–51 forces acting on an object in different acceleration cases, 150 two-finger servo gripper, 151 ultra-high molecular weight (UHMW), 283 ultrasonic flowmeters, 53–4 ultrasonic transmitters, 57–8 ultraviolet (UV) light, 388 ultraviolet (UV) spectroscopy, 116, 120 ultraviolet (UV) wavelengths, 271 vacuum cooling, 299 vacuum gripper, 148, 361 validation tool, 209–10 vegetable reprocessing grading systems, 389–96

© Woodhead Publishing Limited, 2013

Index 503 fruit grading robot, 393–6 grading of eggplant, 393 grading of oranges, 391–3 peeling and grading of leeks, 390–1 very-large-scale-integration (VLSI), 104 visual tag system (VTS), 140 vortex flowmeters, 53 VTScada, 140 walnuts, 281 WATER, 479 water-jet based trimming machines, 323 trimmed lamb chops, 324 uniform fat trimming concept, 324 water treatment, 134–5 wetted parts, 38–40 various stainless steels suitable for the food industry, 39

window–pixel operation, 77 wireless data logger platforms, 177–8 wireless sensor network (WSN), 16 agriculture and food industry, 171–95 applications, 184–94 machine and process control, 188–9 development, 172–84 commercial WSN hardware platforms, 174–6 future trends, 195 WirelessHART, 66–7 wireless sensor network integration, 68 WORK, 479 worker safety, 305–6 X-rays, 388

© Woodhead Publishing Limited, 2013