Research And Application Of Yarn Quality Prediction Methods

Yarn quality prediction is a major and difficult research area in the spinning field and has been attracting the attention of many scholars.Yarn quality prediction refers to the prediction of final yarn quality by raw cotton parameters and process parameters,thus reducing the number of trial spinning and cost.Due to the many parameters in the yarn quality prediction model,which is a complex nonlinear problem,the prediction accuracy is often low when using the traditional algorithm for yarn quality prediction model,which is difficult to meet the actual production needs.Therefore,to address the problems of the traditional algorithm,new improvement schemes are proposed and successfully used in yarn quality prediction cases.The details of the research are as follows.(1)For the problem of low accuracy of yarn strength prediction,an Expert Weighted Neural Network(EWNN)based on Particle Swarm Optimization(PSO)is proposed for yarn strength prediction algorithm.The algorithm is prepared by multiple experts giving upper and lower bounds on the parameter weights,and the PSO algorithm is used to optimize the average absolute percentage error of the test set to obtain the optimal parameter weights.Then,this weight is substituted into the Expert Weighted Neural Network for training,and thus the trained Expert Weighted Neural Network is obtained.The experimental results show that the average absolute percentage error of the EWNN optimized by PSO algorithm is reduced by 1.50 compared with the unoptimized EWNN.(2)For the problem of low prediction accuracy of yarn strength Coefficient of Variation(CV),a Bayesian strength Coefficient of Variation(CV)prediction algorithm based on Principal Component Analysis(PCA)and Automatic Correlation Determination(ARD)is proposed and named as P-ARD algorithm.The algorithm processes the data with multiple covariance by PCA,changes the parameters with covariance into linearly unrelated parameters,and reduces the complexity of the model by selecting the principal components.Finally,based on the selected principal components,the ARD-based Bayesian linear regression method is used to predict the yarn strength CV.The experimental results show that the P-ARD algorithm reduces the error by 0.02 in the worst case and 1.83 in the best case compared with the conventional ARD-based Bayesian linear regression method.(3)For the problem of low accuracy of yarn unevenness prediction,a yarn unevenness prediction algorithm based on Broad Multi-layer Neural Network(BMNN)is proposed.The algorithm uses feature extraction and feature enhancement to extend the features of the original data.The extracted features are then fed into the neural network and the relationship between these features and yarn unevenness is trained by an error back propagation algorithm.Experimental results show that the complexity of the BMNN algorithm is comparable to that of the four-layer neural network,while in terms of yarn unevenness prediction accuracy,the BMNN algorithm solves the problem of large fluctuations in the prediction of the traditional broad learning system algorithm and provides higher accuracy than the four-layer neural network prediction.