Classification And Recognition Of DCT Production Defect Data Based On Auto-encoder Neural Network

The double clutch transmission(Double Clutch Transmission,DCT)occupies a large automobile transmission market in recent years because of its advantages in production and performance,so strict quality control of DCT products is an important link in the future DCT development.At present,the expert system(Expert System)is the most widely used in the on-line detection of the end of the production line of mechanical products.It is necessary to establish a complete fault feature library and judge the complex analysis and calculation of all kinds of signals collected.The final test results are not completely correct,and the false alarm rate is high,so manual aid distinguish is needed.Deep learning(Deep Learning,DL)is a advanced learning algorithm for artificial intelligence(Artificial Intelligence,AI).It can help to extract data features and carry out more accurate classification tasks.It is widely used in various fields including mechanical fault diagnosis.This paper mainly wants to realize end-to-end fault classification and recognition,reduce manual participation and improve detection efficiency and accuracy.Based on the real problem of DCT manufacturing defects at the end of enterprise production line,the research goal of this paper is to realize the automatic classification and identification of the original time domain data of manufacturing defects by combining the semi-supervised depth learning method of self-coding neural network.The main contents of this thesis are as follows:(1)The paper takes the DCT produced by a certain enterprise as the research object,and analyzes and classifies the vibration signals collected in the performance testing after the production and manufacture.Firstly,the vibration signal at the end of the actual DCT production line is collected,and the signal is analyzed,the production defects are classified and the characteristics were summarized.Secondly,the vibration signal is preprocessed,the sample size is increased by data expansion method,and the influence of different step size and sample length on the accuracy of classification and recognition is compared and analyzed.The experimental results show that the selection step size is 256 and the sample length is 4028,which is beneficial to the model training.Finally,the DCT original vibration data set and the DCT gear howling classification data set are established to pave the way for exploring more subdivision classes.In addition,the selected open source data set is introduced to build the service for the Auto-Encoder model and the later analysis and discussion model precision service.(2)An Auto-Encoder classification model based on one-dimensional vibration signals is built in this paper.The effects of loss function,optimization function and batch size and iteration number training parameters on network performance are analyzed.This paper analyzes the classification effect of the network for different data sets,and builds similar different types of models for experimental comparison.Finally,it is concluded that the Convolutional Auto-Encoder(CAE)classification network built by convolution layer pool layer has the best performance.The highest model and guess accuracy on label-free data sets of 50% are 96.06%.(3)In order to improve the accuracy of self-coding neural network in the classification and identification of DCT production defect data,we explore the data form and model optimization.The experimental results show that the data form of two-dimensional vibration image is used instead of one-dimensional data to compress the amount of data on the basis of preserving the original characteristics of the data,and the improvement of the ordinary Auto-Encoder to the two-dimensional convolutional Auto-Encoder also improves the ability of network to capture features and improve the accuracy of model recognition.Based on the Le Net-5 structure,the LCAE model is built and optimized.The mixed model LCAE＿CNN convolution Auto-Encoder and Convolutional Neural Network is obtained by improving the classifier.The prediction accuracy of the model is 1.51 percentage points higher than that of the LCAE.(4)There are many factors affecting the accuracy of the network model.In general,the larger the data size,the higher the prediction accuracy of the model.The experiment proves that expanding the image set size can further improve the prediction accuracy of the model and prove it by drawing the statistical chart of the confusion matrix;The AE and CAE of the LCAE 、LCAE＿CNN four network models based on one-dimensional DCT vibration data set are compared in chapter 4.The advantages of the LCAE＿CNN classification model proposed in this paper are obvious.The prediction accuracy of the proposed LCAE＿CNN classification model is 98.95%.