Research On Yarn Quality Prediction Model Based On Deep Neural Network

In the whole textile industry chain,yarn production is one of the key links in the textile industry chain,and its quality has a great influence on the quality of finished products such as textile and clothing.For a long time,the textile industry has been hoping for a yarn quality prediction technology,which can accurately predict the final yarn quality index based on the known conditions such as raw materials and production technology,which is also called virtual spinning technology.In fact,the essence of this technology is to accurately reveal the relationship between raw material performance index,production process and yarn quality index through prediction model.Because the yarn production process is complex and there are many factors affecting the yarn quality,it is difficult to accurately express the complex relationship between them through simple mathematical models.Textile science and technology workers have done a lot of research work in this field.From physical models,statistical methods to neural networks,there has been a lot of progress,but overall,the prediction accuracy and applicability are not enough.With the improvement of computing power,deep neural network(i.e.deep learning)has been paid attention to for its excellent performance.Deep neural networks can solve the problem that shallow networks cannot accurately express nonlinear complex relationships due to the limitation of network layers.Under this background,this paper uses the deep neural network to establish the relevant prediction model,in order to be able to more accurately express the relationship between the yarn quality index and the factors such as the raw materials and the production process of the product,so as to better predict the yarn quality index,thus having potential application value.The main research work of this paper has three aspects:(1)From the angle of the influence of processing time sequence on yarn quality and the angle of optimizing eigenvalue through convolution and pooling layer in convolution neural network,three models are designed respectively: CNN-BP,CNN-GRNN and CNN-LSTM.CNN stands for convolution and pooling operation of convolution neural network,which is mainly designed from the angle of optimizing eigenvalue.LSTM is designed from the angle of considering the influence of processing sequence on yarn quality.BP and GRNN are designed without considering the influence of processing timing on yarn quality,where GRNN is designed from the perspective of small sample data sets.(2)The parameter determination and prediction effect of the three models are studied respectively.(3)Verify the yarn quality prediction of several spinning methods.The full text consists of eight chapters,which will be introduced respectively as follows:The first chapter is an introduction,which mainly includes four aspects: the first is the research background and significance;The second is the research content and objectives;The third is the research thinking and technical route;The fourth is the overall framework of the paper.Three cases of deep neural network model design with or without spinning timing are mainly proposed.The second chapter is literature review.This chapter mainly summarizes and analyzes the historical research on yarn quality prediction technology.The development process of yarn quality prediction technology is from mathematical statistics technology to grey theory and then to machine learning technology.Machine learning technology mainly includes support vector machine and neural network.Through the analysis of contribution,the biggest problem of yarn quality index prediction is how to improve the prediction accuracy and wide applicability,thus pointing out the direction for the research goal of this paper.The third chapter is the design and implementation of yarn data warehouse.Firstly,a data warehouse is designed according to the theme of yarn quality index prediction.Then,a specific rotor yarn is taken as an example to establish a case data warehouse to provide support for subsequent yarn quality prediction.The fourth chapter is based on CNN-BP depth neural network yarn quality prediction model.CNN-BP model is a deep neural network yarn quality prediction model without considering the influence factors of process timing on yarn quality.CNN is an optimization of the input original eigenvalues,and BP is a regression fitting of the optimized eigenvalues.The input of CNN-BP model is that the fiber performance indexes and process parameters form an N×N matrix in sequence.After 2 times of 2-D convolution kernel from 1×1 to N×N,the optimization of eigenvalue combination from part to whole is realized.Finally,the optimized eigenvalue is fitted by BP neural network,thus the corresponding yarn quality indexes are predicted.The parameter configuration of CNN-BP model in rotor spinning is obtained through the experiments and analysis of input eigenvalue sorting,fiber performance index,process parameters,convolution kernel,activation function,number of BP neurons and other factors.Training and testing with 200 sets of rotor yarn data sets show that the prediction accuracy of CNN-BP model is higher than that of traditional BP neural network and linear regression yarn quality prediction methods.On this basis,considering that CNN-BP model has many parameters and 200 sets of data sets cannot fully train the model parameters,this paper proposes to train CNN-BP model by using generated countermeasure network(GAN)to integrate 20,000 sets of data based on the original 180 sets of data used for training.The experimental results show that the prediction accuracy of CNN-BP model trained by 200180 sets of data sets is higher than that of CNN-BP model trained by 180 sets of data sets only.The fifth chapter is the research of yarn quality prediction model based on CNN-GRNN depth neural network.At present,there are still some difficulties in collecting spinning data in the industry,and small sample data sets are often obtained.In order to solve this problem,CNN-BP model has been improved accordingly.First,CNN is partially improved by changing the original two-dimensional convolution to one-dimensional convolution,that is,one-dimensionalconvolution from 1 to N,so that the combination optimization of eigenvalues from part to whole can be more fully realized.Secondly,the BP neural network with more global fitting parameters is changed to the GRNN neural network with only one super parameter for local fitting.Through the discussion and analysis of CNN-GRNN input fiber index parameters,process parameters,convolution kernel parameters,super parameters of GRNN network and so on,various parameters of CNN-GRNN model can be finally determined.The experiment was also carried out on 200 sets of data of rotor spun yarn examples.The results show that the prediction accuracy of CNN-GRNN model on 180 sets of training sets is higher than that of CNN-BP model,which is like that of CNN-BP model using 20180 sets of training sets.Chapter 6 is a yarn quality prediction model based on CNN-LSTM depth neural network.This model is a CNN-LSTM model based on the assumption that the influence of textile processing timing on yarn quality is considered.CNN optimizes the input characteristic value through one-dimensional convolution and pooling.LSTM fits the optimized fiber performance index and process parameters according to the processing time sequence,thus realizing the prediction of yarn quality index.The effects of input fiber performance index,process parameters,convolution kernel parameters,pool kernel parameters,LSTM cell number,LSTM layer number and other factors on prediction accuracy are studied,and the lattice parameters of CNN-LSTM model are determined.Experiments on spinning data sets show that the prediction accuracy of CNN-LSTM model is higher than that of CNN-GRNN model,which means that CNN-LSTM model has the highest prediction accuracy among the three depth neural network models proposed in this paper.Of course,this needs to be supported by relevant process parameter data in spinning process.Chapter 7 is the application and verification of the three models proposed in this paper in yarn quality prediction of other spinning systems.In this chapter,four test schemes are proposed to verify the validity of the three deep neural network models.The test data sets of the four schemes are ring spinning data set with dynamic process data,ring spinning data set without dynamic process data,wool spinning data set with dynamic process data and vortex spinning data set without dynamic process data.Experiments on ring spun cotton yarn data sets and wool spun yarn data sets with dynamic process timing show that the CNN-LSTM model has the highest prediction accuracy among the three depth neural network models proposed in this paper on ring spun and vortex spun data sets without dynamic process data,only CNN-BP model and CNN-GRNN model participate in the experiments.The experimental results show that CNN-GRNN model has high prediction accuracy.Through four groups of experiments,it can be concluded that on the yarn data set with dynamic process data,the three models proposed in this paper can be predicted,of which CNN-LSTM model has the highest prediction accuracy.On the yarn data set without dynamic process sequence,it is only suitable for CNN-BP model and CNN-GRNN model.Among them,CNN-GRNN model is suitable for small sample data set due to its special network structure,and its prediction accuracy is higher than CNN-BP model.Chapter 8 is the conclusion.The application verification of the three kinds of deep neural network yarn quality index prediction models proposed in this paper in various spinning systems shows that they are all effective.At the same time,it is also found that the application of the threemodels is applicable.CNN-LSTM model performs best on yarn data set with dynamic process data.CNN-GRNN model is suitable for the situation where there is no dynamic process yarn data in a small sample.CNN-BP model is applicable to large samples without dynamic process.The three models based on deep learning neural network proposed in this paper are an exploration on the prediction of yarn quality.The research has certain guiding and reference significance for the application of this technology in the field of yarn quality,especially the introduction of key process parameters,which has good effect.Of course,there is still a need for further research on the selection of spinning process parameters and the determination of training parameters in the future.