Quality Analysis And Prediction Of Rotor Yarn Based On Small Sample Data

In the development of differentiated yarns,there is a real need to produce a wide range of small batch rotor spinning yarns,and product qualities are important for stable production and additional value.However,these products often suffer from a limited size of single batch data,which makes the selection of quality prediction models difficult and there are no general guidelines for this situation.This paper investigates the applicability of linear regression,random forest and neural network models for quality prediction of small sample rotor spinning yarns,so as to recommend a more reasonable quality prediction model for different types and sizes of small sample rotor spinning yarn data.The study samples were obtained by designing a rotor spinning yarn experiment to investigate the effects of raw cotton and rotor process on yarn quality,and then using linear regression,random forest and neural network models to predict yarn quality and compare their differences in prediction results.The main research methods and conclusions of this paper are as follows:Firstly,the design of spinning solutions for rotor yarn samples and the selection of key components for rotor machines.In the beginning,23 spinning schemes were proposed for 7cotton raw materials mainly by using quadratic general rotary design and uniform design as the main experimental design methods.Then,according to the production scenario of spinning 21~Spure cotton yarn on Schlafhorst BD6 rotor machine,the orthogonal test was designed and the test results were analyzed by fuzzy evaluation method.The best combination of rotor machine components was OS21 combing roller,R6KS5 navel and G32D rotor cup.At the end,in order to obtain quality indicator data for yarn quality analysis and prediction,the cottons and yarns were tested.Secondly,the influence of the raw cotton and the rotor spinning process on the yarn quality and the percentage of the combined influence were investigated.The single-factor analysis of the test data led to the conclusion that raw cotton had a significant influence on the yarn breaking tenacity,breaking elongation,unevenness and nep;a suitable combination of the three process parameters(rotor speed,carding roller speed and twist factor of yarn)helped to improve yarn quality.Based on this,the combined influence of spinning process and raw cotton on yarn quality was analyzed as 62.1:37.9 using assessing the feature importance based on the random forest model.Thirdly,prediction of yarn quality based on machine learning models and comparative analysis of the results of different models.The coefficient of determination of the three models was 0.567,0.499 and 0.197 for linear regression,random forest and neural network respectively,in other words,linear regression and random forest were at the same level of predictive ability and higher than neural network.Fourth,the effect of data characteristics on the prediction performance of the models was explored.Using Spearman correlation analysis to remove some of the input features and make predictions,the prediction result shows that both linear regression and random forest’s accuracy did not change significantly after the input parameters were reduced,meanwhile neural network’s accuracy improved.For the training size of 40-100 train data,the accuracy of linear regression did not change significantly as the sample size increased,random forest improved slightly,neural network improved significantly.Finally,a comprehensive conclusion was drawn for the prediction of small sample yarns’quality:the linear regression prediction method is recommended when the input parameters are small or the sample size is very small;the random forest prediction method is recommended when the input parameters are large,the input parameters are non-linear variables or the sample size reaches hundreds.