Diagnosis Method Of Bearing Fault By Visualization Of Vibration Signal Based On Deep Learning And Transfer Learning In The Case Of Few Sample

Bearing is an extremely important part in rotating machinery,which has a complex,harsh working environment and long working time,so it has become one of the most fault-prone parts in rotating machinery.The traditional bearing fault diagnosis method mainly uses one-dimensional vibration signal,which has many feature information,complex screening,incomplete feature extraction and heavy workload,and the signal is easily disturbed by noise.To a certain extent,it limits the further improvement of the accuracy of bearing fault diagnosis and the further optimization of diagnosis efficiency.The fewer sample will lead to the model can not fully extract features,can not train a reliable and effective,strong generalization of the diagnosis model,affecting the effect of the diagnosis model.In this paper,the imaging method of bearing vibration signal is studied firstly,the bearing vibration signal collected by sensor is encoded into two-dimensional image by image representation method,and the image is input to the feature extraction network of deep learning.bearing fault diagnosis based on vibration signal visualization is realized.Then,based on the method of visualization of vibration signals,the bearing fault diagnosis in the case of few sample is further studied.The details are as follows:(1)Aiming at the problem that the extraction of bearing fault feature information is incomplete,the workload is heavy,and the signal is easily disturbed by noise,based on the collected bearing vibration signal data and Case Western Reserve University(Case Western Reserve University,CWRU)bearing open data set,the bearing vibration signal is characterized by vibration signal visualization,and five kinds of two-dimensional images are obtained.Then the two-dimensional image is input into the Residual Network(Residual Network,Res Net)for training,and finally the experimental verification is carried out.From the aspect of image conversion quality--images of Gramian Angular Field(Gramian Angular Field,GAF),such as angular sum field(GASF),angular difference field(GADF),Markov Transition Field(Markov Transition Field,MTF),the original binary Recursive Plots(Recurrence Plots,RP)and the improved color map proposed in this paper,are compared.The results show that using Res Net34 as the fault category recognition method,the average accuracy of the transformed image of the RP color representation method on the two data sets is 89.428% and 98.764% respectively,reflecting the superiority of the proposed RP color image representation method.(2)Aiming at the problems of insufficient training,poor reliability and weak generalization caused by the lack of training samples.Combined with the advantages that Transfer Learning(Transfer Learning,TL)can reduce the dependence of the model on samples and accelerate the speed of convergence,the Swin Transformer-TL method with Self-Attention(Self-Attention)is introduced into the two-dimensional image data set encoded by RP color image representation method.The global feature information of RP color image is extracted and compared with Res Net,Dense Convolutional Network(Dense Convolutional Network,Dense Net),Vision Transformer and their respective TL models.The results show that the average recognition rate of the Swin Transformer method combined with Transfer Learning is the highest,and the highest accuracy of the test set is 97.1%.The stability of the RP color image representation method is further verified on different models.(3)In order to solve the problem that the training samples of the bearing fault diagnosis model are insufficient,which leads to insufficient feature extraction,the Generative Adversarial Network(Generative Adversarial Network,GAN)is used as the image data enhancement method,and the Depth Convolution Generated Adversarial Network(Deep Convolution GAN,DCGAN)and WGAN-GP(Wasserstein GAN-Gradient Penalty,WGAN-GP)model are used to expand the RP color image samples from the bearing vibration signal.The generated image is used to replace the original training set image to train the fault diagnosis model.The results show that WGAN-GP can better learn the distribution and characteristics of existing image data,generate high simulation data,and the generated images can optimize the training effect of fault diagnosis model,and improve the effect of fault diagnosis in the case of few sample.