Research On Rolling Bearing Fault Diagnosis Method Under Few Sample Condition

In recent years,the deep learning method with the background of big data has made remarkable achievements in the field of mechanical fault diagnosis,which provides an important way for the research of intelligent fault diagnosis of equipment.However,these studies are realized on the basis of sufficient training data,and the diagnostic effect of the model is greatly affected by the amount of data.In the actual working environment,due to various constraints,it is unrealistic to collect and mark sufficient fault data for each bearing state.Specifically,when the fault samples are far less than the normal samples,the data distribution will be unbalanced;when the absolute number of various bearing data is very small,it will be a small sample.Unbalanced phenomenon and small samples are collectively referred to as the problem of few samples.Therefore,how to construct a reliable fault diagnosis model under the condition of few samples is an urgent problem to be solved.In this paper,the following work is carried out to solve the problem of bearing fault diagnosis when it is difficult to collect rich data in practical production applications.(1)An unbalanced sample fault diagnosis method based on VAE-GAN and FLCNN is proposed.Aiming at the problems of poor generalization ability and low diagnosis accuracy caused by unbalanced distribution in bearing fault diagnosis,this paper studies it from two aspects of data and algorithm.At the data level,a VAE-GAN sample augmentation model combining variational auto-encoder and generative adversarial network is proposed to expand the fault samples;At the algorithm level,a FLCNN model based on Focal Loss and onedimensional convolutional neural network is proposed to complete bearing fault identification.The experimental results show that the method in this paper can effectively improve the bearing fault diagnosis effect under the condition of unbalanced samples,the Recall and F1-score value can reach 0.93 and 0.92 respectively on the public data set with unbalanced ratio of 1:10.(2)A small sample fault diagnosis method based on improved Siamese network is proposed.In view of the situation that only a small amount of bearing data can be obtained under certain conditions,the current rolling bearing fault diagnosis model based on deep neural network is difficult to be fully trained,which will produce serious over fitting problems.Based on the idea of metric learning,the current mainstream deep learning fault diagnosis models are difficult to be fully trained,which will cause serious over fitting problems.Based on the idea of metric learning,this paper proposes an improved Siamese network fault diagnosis method.This method adds a classification branch to the Siamese network,and replaces the commonly used Euclidean distance metric with a network metric.First,the fault samples are input into the feature extraction network shared by two parameters in pairs,and the bearing signal data are mapped to the low dimensional feature space by using long-term memory network and convolution neural network;Then,the similarity of the extracted sample features is measured through the relationship measurement network,and the sample features are input into the classification network to complete the bearing fault identification.The verification results on the laboratory data set show that,compared with methods such as the unimproved Siamese network and classical convolutional neural network,this method can achieve higher fault diagnosis accuracy and better generalization performance under the condition of small samples.When there are only 10 training samples in each category,the accuracy of the model can be83.6%;when there are 100 samples in each category,the accuracy can reach 98.6%.