Research Of Rolling Bearing Fault Diagnosis Method Under Unbalanced Sample Conditions

Fault diagnosis methods based on data-driven and deep learning have become the mainstream in the field of fault diagnosis and have achieved great success.However,the vast majority of fault diagnosis research is still in the scope of supervised learning.In the actual industrial production process,due to various reasons,fault samples often have insufficient labels or even missing labels.In addition,because the probability of occurrence of each type of fault is not the same,it may be difficult to collect samples of a certain type of fault or the probability of occurrence is low.It is often difficult for researchers to collect enough fault sample data for this type of fault.It can still be analyzed and diagnosed in some ways,such methods are called few-shot learning or extreme zero-shot learning.This thesis is to diagnose the fault of the rolling bearing under this condition.The main research contents of this paper are as follows:Aiming at the problem of few-sample fault diagnosis,a bearing fault diagnosis method based on GAN(Generative Adversarial Network)and expanding the fault sample set is designed.The one-dimensional bearing fault sample is converted into a twodimensional matrix image by means of image conversion,through the training of the generator and discriminator in the generative confrontation network,the real fault sample is expanded,and the impact of sample imbalance on diagnosis is reduced.Then it is diagnosed by(Convolutional Neural Networks).The accuracy rate reached 92.80% in the experiment on the CWRU(Case Western Reserve University)data set,indicating that it is feasible to use GAN to expand the data of rolling bearings.Aiming at the problem of unbalanced fault samples,a fault sample generation method based on DCGAN(Deep Convolution GAN)is designed.STFT(Short-time Fourier Transform)is used to convert the one-dimensional fault sample signal into a twodimensional fault sample time-frequency diagram,and the generator and discriminator of the DCGAN used are respectively optimized to enable faster convergence and generate higher quality images Extended sample.The diagnostic results on the CWRU dataset show that,compared with sample expansion using GAN before expansion,using DCGAN and STFT for expansion can achieve an accuracy rate close to 100%.Aiming at the problem of zero-sample fault diagnosis,a fault diagnosis method based on fault attribute description is designed.The time-frequency image is input into the deep residual convolutional neural network to extract the characteristics of the visible class samples,and then use these characteristics and attributes to train the attribute learning network based on the convolutional neural network,so as to realize the prediction of the unseen class samples.Also using the CWRU dataset for experiments,the recognition rate for different attributes can reach 96.4%.