APPLICATION OF ARTIFICIAL NEURAL NETWORKS FOR PREDICTING YARN PROPERTIES AND PROCESS PARAMETERS by ANIRBAN GUHA DEPARTMENT OF TEXTILE TECHNOLOGY Submitted in fulfillment of the requirements of the degree of DOCTOR OF PHILOSOPHY to the INDIAN INSTITUTE OF TECHNOLOGY, DELHI February, 2002
Tq b77 :. 0
CERTIFICATE This is to certify that the thesis titled "Application of Artificial Neural Networks for Predicting Yarn Properties and Process Parameters", being submitted by Mr. Anirban Guha to the Indian Institute of Technology, Delhi, for the award of the degree of Doctor of Philosophy is a record of bonafide research work carried out by him. Mr. Anirban Guha has worked under our guidance and supervision and fulfilled the requirements for the submission of the thesis. The results contained in this thesis have not been submitted, in part or in full, to any other university or institute for the award of any degree or diploma. Dr. R. ChtopacftIyay Associate Professor Department of Textile Technology, lit Delhi, New Delhi - 110016, India o,ji&fv / ayadeva Associate Professor Department of Electrical Engineering, IIT Delhi, New Delhi - 110016, India
ACKNOWLEDGEMENT I would like to acknowledge the constant guidance and support given by my supervisors, Dr. R. Chatopadhyay and Dr. Jayadeva during the course of my work. Without their continuous encouragement and help at all stages of my work, this thesis would not have seen the light of the day. I also thank Prof. K. R. Salhotra, Prof. R. B. Chavan and Prof. B. L. Deopura who, as head of the Textile Department at various stages of my research, allowed me to use all facilities available in the department. It was necessary for me at some stages to work at odd hours. I am grateful to them for giving me permission to work according to my own schedule. I also thank all the professors of the Textile Department for giving me support and encouragement during the course of my work. During the initial period of my work, it was necessary to obtain data from an industry. I am grateful to Dr. P. R. Roy for allowing me to obtain the necessary data from Arvind Mills. I also thank Dr. J. J. Shroff, Mr. Vijay Chhabra, Mr. Subramaniam, Mr. Kamal K. Saha and Mr. Patel of Arvind Mills for helping me to get the necessary data from the mill. During the latter stages of my work, it was necessary to spin a significant quantity of yarn. I am grateful to Mr. M. P. Mukherjee for allowing me to obtain the raw material needed for this purpose from his industry. I am also grateful to Prof. Prabir Roy, Principal of The Institute of Jute Technology, Calcutta, for allowing me to use the excellent laboratory facilities available in his institute where I spun the yarn and conducted some tests on them. I am
grateful to Prof. Sunil Sett and Dr. Ashish Mukherjee for helping me in every stage of spinning the yarns and testing them in IJT. I thank the staff members of all the laboratories of the Textile Department for extending a helping hand whenever needed. In the same breath, I thank the staff members of the Spinnng and Testing laboratories of IJT. Keeping my composure for all these years would not have been possible without the aid of friends. I am grateful to all my friends in lit Delhi, specially those in the Shivalik Hostel, for helping me to keep a high morale for these four and half years. I would like to specially mention Mr. Anindya Ghosh for his help during the study of yarn structure and Mr. Alok Kanti Deb for his help during the study of Support Vector Machines. In the end, I would like to thank all the fuzzy sets of people with whom I have been in contact during the years of my research, iii
ABSTRACT Prediction of yarn properties from fibre properties and process parameters using artificial neural networks formed the prime focus of this thesis. The performance of neural network, mechanistic models and statistical tools for predicting ring yarn strength was judged - both on data reported in a paper as well as generated in the laboratory. Neural network outperformed the other techniques in both cases. The success of the neural network encouraged the use of this technique for predicting a range of properties of ring and rotor yarns spun in the industry and ring yarns spun in the laboratory. Half of the errors were less than 5% and about one out of ten result was very poor - more than 20%. The ability of a trained network to discern the relative importance of the input units has been investigated where the inputs were fibre properties. Skeletonization, an approach reported in literature failed in this task. A new approach proposed in this thesis - sensitivity analysis - has been found to be successful. Determination of process parameters from yarn properties, Le the reverse of what was being attempted so far, was next attempted. It was found that neural networks can indeed be used for this 'reverse engineering' provided that the yarn property combinations are feasible (i.e. practically achievable). The feasibility of a yarn property combination could be examined with the aid of principal component analysis. In the final part of the study, the possibility of improving the performance of neural networks was explored. It has been shown that improvement in the performance of ANNs is possible by orthogonalising the input data. When correlation between inputs is high, reduction of the least important orthogonalised components can bring about a further improvement in the network's performance. iv
CONTENTS Certificate Acknowledgements Abstract Contents List of figures List of tables Page No. ii iv ix xi Chapter 1 Introduction 1.1 Motivation for studying yarn property prediction 1 1.2 Previous attempts at yarn property prediction 1 1.3 Neural networks and textile engineering 2 1.4 Objective 3 Chapter 2 Literature Survey 2.1 Previous attempts at yarn property prediction 5 2.2 Structure of human brain 17 2.3 Artificial neurons 18 2.3.1 Backpropagation algorithm 24 2.4 Application of ANN in various textile fields 27 2.4.1 Application to fibres 27 2.4.1.1 Cotton cultivation 27 2.4.1.2 Fibre identification 28 2.4.2 Application to yarns 29 2.4.2.1 Spinnability 29 2.4.2.2 Yarn property prediction 30 2.4.2,2.1 Tensile properties 30 2.4.2.2.2 Other properties 33 2.4.3 Application in fabrics 34 2.4.3.1 Fabric manufacture 34 2.4.3.2 Fabric property prediction 35 2.4.3.3 Fabric classification 35
Page No. 2.4.3.4 Sewing and garments 39 2.4.4 Application to chemical processing 42 2.4.5 Application to man made textiles 43 2.4.6 Texturing 44 Chapter 3 A Comparison of Mechanistic, Statistical and Neural Network Models 3.1 Introduction 46 3.2 Deciding the area of investigation 47 3.3 A brief description of Frydrych's model 48 3.4 Constructing a statistical model 51 3.5 Constructing a neural network model 52 3.5.1 Optimising various network parameters 53 3.6 Appraisal of models 58 3.6.1 Cotton yarn data (from Frydrych's article) 58 3.6.2 Polyester yarn data 59 3.7 Conclusion 63 Chapter 4 Prediction of Various Yarn Properties 4.1 Introduction 64 4.2 Experimental 65 4.2.1 Ring yarn from industry 65 4.2.2 Rotor yarn from industry 68 4.2.3 Ring yarn (laboratory spun) 76 4.3 Results 80 4.3.1 Data obtained from industry 80 4.3.2 Data generated in laboratory 82 4.4 Conclusion 83 vi
Chapter 5 Investigation on Identifying Relative Importance of Fibre Properties on Yarn Properties Page No. 5.1 Introduction 85 5.2 Network used 86 5.3 Skeletonization method 86 5.3.1 Theory 86 5.3.2 Experimental 87 5.3.3 Results 88 5.4 Sensitivity analysis 91 5.4.1 Theory 91 5.4.2 Experimental 94 5.4.3 Results 94 5.5 Conclusion 99 Chapter 6 Prediction of Process Parameters from Yarn Properties 6.1 Introduction 100 6.2 Prediction of process parameters from actual yarn properties 6.2.1 Procedure 101 6.2.2 Results and discussion 103 6.3 Prediction from random combination of yarn properties 6.3.1 Procedure 104 100 103 6.3.2 Results and discussion 106 6.4 Analysis of partial failure of network 108 6.5 Identification of data cluster using principal component analysis 109 6.5.1 Theory 109 6.5.2 Application 112 6.6 Conclusion 118 vii
Chapter 7 Possibilities of Improving the Performance of Neural Networks Page No. 7.1 Introduction 119 7.2 Data used for the study 119 7.3 Study on ring yam data 119 7.3.1 Application of principal component analysis 124 7.4 Study on rotor yarn data 127 7.5 Conclusion 135 Chapter 8 Conclusion 136 Chapter 9 Suggestions for further work 139 References 141 Bio-data viii