论文摘要
The Kenyan textile industry witnessed a rapid growth when Kenyan gained its independence in 1963 and the number of integrated textile mills（consisting of spinning, weaving and dyeing units） increased from 6 to 52,with an installed capacity of 115 million square meters,along with 110 large scale garment manufacturing units in 2 decades.This made the sub-sector the second largest manufacturing industry after food processing. Despite the many problems which have plagued the Kenyan textile industry,especially in the late 1990’s the textile industry has maintained its position as one of the important industrial sectors in the Kenyan economy and also ranked among the major foreign exchange earners.The aim of this thesis was to improve the prediction efficiency of yarn quality properties prediction models with special reference to the Kenyan manufactured ring spun cotton yarn.In order to achieve the above stated aim this thesis has been divided into six chapters. Chapter one deals with literature review,short comings of the previous research work,the general outline and contributions of this thesis.The literature review covered the types of mathematical,empirical,statistical and artificial neural network（ANN） models used to study the cotton ring spun fiber-to-yarn process.The conclusion of the literature review indicated that in comparison with statistical models the ANN models give better performance.On the other hand the statistical model have an advantage over mathematical and empirical models during the study of the fiber to yarn process.Therefore it can be concluded that the ANN models perform better than statistical,mathematical and empirical models during the study of the complex multi-stage fiber-to-yarn process.In view of the above mentioned conclusions,yarn quality properties prediction models were designed using ANN models.From the literature review undertaken in chapter one three shortcomings of the previous research work were identified.These short comings included network generalization, network scale and efficiency of the network training algorithm.Network generalization（R-value）,which is defined as the ability of the network to respond to unseen data is critical for the application of the network in solving day to day problems in the industry.This area has received little attention especially for the fiber-to-yarn process.The second area of immense importance to the performance of the ANN models is the scale of the model.Most researchers use many input factors for the purpose of obtaining a low network error.While the use of more input factors may lead to a more comprehensive study of the problem,there is however the danger of increasing the scale.More input factors normally give a lower network error but leads to difficulties in training due to the increased scale.On the other hand fewer input factors give less training problems due to the reduced scale but the network error may be slightly higher.Therefore there is a need for a study of the scale reduction methods for yarn quality properties prediction models.The third short coming which this thesis will address is the improvement of the network training algorithms.While there are many training algorithms which can be used to train ANN models,there is still room for the improvement of the performance of the training algorithms so that better performing models can be designed.Better performing algorithms will lead to improved model performance.In chapter two the basic principles of multi-layer feedforward network（MLFN） and the multi-layer Perceptron（MLP）) were covered.The yarn quality prediction MLP models were designed based on Cybenko theorem which gives the architecture of the MLP to be one hidden layer and the output is one.The transfer functions used in the hidden and output layers were sigmoid and purelin transfer functions respectively.The network target error e was set to 0.00land the number of neurons in the hidden layer was increased gradually until e was attained.The input factors were selected based on literature review and advice from experts in the textile industry.A total of 19 input factors namely,13 HVI fiber properties,4 machine parameters（spindle speed,ringframe draft,ring diameter and traveler weight） and 2 yarn parameters（twist and count） were used.The 13 HVI fiber properties were fiber length,fiber length uniformity,fiber micronare,fiber maturity,Spinning Consistency Index （SCI）,Short fiber Index（SFI）,fiber strength,fiber elongation,fiber trash weight,fiber reflectance,fiber trash area,fiber grade and fiber yellowness.The samples of cotton lint an d manufactured yarn were collected from three factories in Kenya.The results obtained after training the yarn prediction MLP models showed that yarn strength could be predicted with mse value of 0.0009 and R-value of 0.869 using 12 neurons in the hidden layer.The elongation and unevenness models showed similar mse values like the strength model but needed 15 neurons in the hidden layer.The R-values for the elongation and unevenness models were 0.842 and 0.888 respectively.The results were acceptable but the network scale seemed to be large.The network generalization（which is indicated by R-value） had to be improved.The improvement of the designed yarn quality prediction models was done using two main strategies.The first strategy involved reducing the number of input factors by using the Principal Component Analysis（PCA） method.This has been discussed in chapter three.The second strategy involved the improvement of the network training algorithm.In chapter four the network training algorithm was optimized by using Differential Evolution（DE） algorithm,while in chapter five a non-BP network training algorithm called Extreme Learning Machine（ELM） was used.As mentioned above the third chapter of this thesis attempted to reduce the scale of the yarn quality prediction models designed in chapter two by compacting the 19 input factors to 14 using the PCA method.The 14 input factors were referred to as carefully selected input factors and used to re-train the yarn quality properties prediction models.The re-trained strength model had 10 neurons in the hidden layer while the elongation and unevenness models had 11 neurons in their hidden layers.The strength prediction model showed an improved R-value of 0.917 and mse value of 0.00083.The elongation prediction model showed an improved R-value of 0.894 and mse value of 0.00089.The unevenness prediction model had an improved R-value of 0.915 and mse value of 0.00087.Thus the input data reduction method caused a reduction of the models scale and an improvement of network performance and generalization.To further understand the effect of each input factor on a given yarn quality property,the input skeletonization（subtraction） method was used to study the impact of input factors on yarn quality properties.The influence of input factors on yarn strength listed in order of importance was as follows;fiber strength,fiber length,fiber micronaire,fiber length uniformity,yarn twist,spindle speed,fiber elongation, fiber maturity,SCI,yarn count,ring diameter,SFI,fiber yellowness and fiber trash grade. For yarn elongation the influence of input factors（given in order of decreasing importance） was;fiber elongation,fiber length,fiber length uniformity,fiber micronaire,yarn twist, spindle speed,fiber maturity,fiber strength,SCI,SFI,fiber trash grade,yarn count,fiber yellowness and ring diameter.The impact of input factors on yarn unevenness（listed in order of decreasing importance） was as follows;fiber length,fiber micronaire,fiber length uniformity,SFI,fiber trash grade,spindle speed,yarn twist,SCI,fiber strength,yarn count, fiber maturity,ring diameter,fiber yellowness and fiber elongation.The models obtained in chapter three were improved by using a hybrid algorithm combining DE and LM algorithm and referred to as DELM algorithm.In chapter four the DELM algorithm was designed such that by increasing the number of generations in the DE algorithm the performance and generalization（R-value） of the model was improved.The weights and biases in chapters two and three were initialized by random selection.When using the DELM algorithm the DE algorithm optimized the initial values of weights and biases by using mutation,crossover and selection strategies.The optimized weights and biases were then past on to the LM algorithm which was used for training.The carefully selected 14 input factors were used to train the yarn quality properties prediction models using the DELM algorithm.The strength prediction model showed an improved R-value of 0.959 and a reduced network scale of 7 neurons in the hidden layer with 3 generations of the DE algorithm.Similarly the elongation prediction model showed an improved R-value of 0.94 and a reduced network scale of 6 neurons in the hidden layer with 6 generations of the DE algorithm.The unevenness prediction model showed an improved R-value of 0.939 and a reduced network scale of 6 neurons in the hidden layer with 7 generations of the DE algorithm.The results obtained using the DELM models were much better compared to the previous results.Thus the DELM algorithm was able to produce more compact yarn quality prediction models which had better performance and generalization.Due to their reduced scale the DELM yarn prediction models will need less computational resources when compared to the LM yarn quality properties prediction models.In chapter five,Extreme Learning Machines（ELM） a non-BP network training algorithm was used to train yarn quality properties prediction models with the carefully selected 14 input factors.The ELM network training algorithm is a non-BP algorithm which randomly selects the input layer to hidden layer weights and biases and then analytically calculates the hidden layer to output layer weights and biases.The main difference between the ELM and the BP algorithms is that the ELM algorithm does not use the iterative procedure to determine the final weights and biases and hence the ELM training speed is very high when compared to the LM algorithm.The ELM strength,elongation and unevenness prediction models needed 41,43 and 58 neurons to achieve mse values of 0.000949,0.00097 and 0.000906 respectively.The training time was however very low;0.0311,0.0317 and 0.0313 seconds.This was over 80 times faster than the time needed to train yarn quality properties prediction models trained using the LM algorithms.In order to reduce the scale of the ELM yarn quality properties prediction models and to approximate the global prediction value,a hybrid algorithm combining DE and ELM and referred to as DE-ELM was designed.The optimized DE-ELM strength prediction model showed a reduced scale of 2 neurons in the hidden layer,an improved R-value of 0.992,an improved mse value of 0.00039 and a training time of 0.7188 seconds.The optimized DE-ELM elongation prediction model also reported a reduced scale of 2 neurons in the hidden layer,an improved R-value of 0.977,an improved mse value of 0.00005 and a training time of 0.6875 seconds.Similarly the optimized DE-ELM unevenness prediction model showed a reduced scale of 2 neurons in the hidden layer,an improved R-value of 0.984,an improved mse value of 0.00033 and training time of 0.6563 seconds.Thus the DE-ELM models showed reduced network scale, improved mse values and improved R-values.The final chapter of this thesis（chapter six） summarized the results obtained and briefly discussed the limitations of this research work.From the results obtained in chapters two, three,four and five,it is clear that the network scale kept on reducing,the network performance kept on improving and the network generalization kept on improving as the network improvement strategies were implemented.Therefore the following conclusions can be made;（i） LM algorithm combined with early stopping method can be used to predict yarn quality properties giving good network performance and generalization,（ii） Principal Component Analysis（PCA） can be used to reduce the number of inputs factors thus compacting the scale of the network used for yarn quality properties prediction,while at the same time improving its performance （iii） Hybrid algorithms made up of local search algorithms like LM and global search algorithms like DE can be used to reduce the scale of the network and improve the performance and generalization of yarn quality prediction models（iv） Extreme Learning Machines（ELM） a non-BP network training algorithm can be used to predict yarn quality properties.ELM has extremely high training speeds but needs more neurons in the hidden layer when compared to the LM algorithm（v） DE-ELM hybrid algorithm made up of DE and ELM algorithms gives improved network scale,performance and generalization when compared to LM algorithm during the prediction of yarn quality properties.While the results obtained in this research work are acceptable,there are however some limitations which were noticed in the course of the research work.The first limitation of this research concerns the network architecture.In the present work the network parameters （weights and biases） were optimized with DE algorithm,however the number of neurons in the hidden layer was designed based on experiments.The experimental approach is good, but limits the efficiency of the design process and needs more time.Network architecture can be optimized using genetic algorithm（GA） to produce more compact and efficient models.Due to the time limitations it was not implemented in our present work but this area of study holds good prospects for future research work.The second limitation of this research work deals with the use of other types of ANN models.There are many types of artificial neural networks such as;Multi-Layer Feedforward Neural Network（MLFN）, Recurrent Neural Network（RNN）,Self-organizing neural network,Hopfield neural network,Fuzzy Neural Network（FNN）,Pulse Coupled Neural Network（PCNN） etc.This research work concentrated on MLFN models.Commonly used MLFN models include the Multi-layer Perceptron（MLP）,RBF network,wavelet network and recurrent network.In this research work the yarn quality prediction models were designed using the MLP model. We wish to recommend that future research work should be done to model the fiber-to-yarn process using other types of MLFN.This kind of research will increase the scope of models available for the modeling of the fiber-to-yarn process. ??In summary this research work has resulted in the better understanding of the Kenyan manufactured cotton ring spun yarn,through the models designed to predict yarn quality properties.Such an understanding is critical for the improvement of the performance of the Kenyan textile industry.