Application Of Data Augmentation Method In Fault Diagnosis Of Rotating Equipment Using Variational Auto-encoder

In the field of fault diagnosis of rotating equipment,the problem of data imbalance is usually encountered,that is,the amount of fault data,especially catastrophic failure data or unexpected mechanical failure,is very small,while the amount of normal state data is very large.When the commonly used data-driven fault diagnosis method predicts unbalanced data,the result is that the prediction accuracy of the majority of samples is often very high,while the accuracy of the minority of the samples will be very low.Will have fatal consequences.For example,in the industrial field,if the fault state of the rotating equipment is misdiagnosed as the normal state,it may cause serious economic losses and even casualties.Therefore,in order to solve the problem of data imbalance in the fault diagnosis application of rotating equipment,and to improve the classification accuracy of the data-driven fault diagnosis method on the unbalanced data set,the following research is carried out in this paper:(1)Aiming at the problem of unbalanced categories of unbalanced data sets,this paper proposes a data enhancement method based on variational autoencoders.This method uses the Gaussian hidden variables of the variational autoencoder to sample minority vibration signals to generate simulation data.Demonstration experiment results show that,compared with the traditional data enhancement method,the data enhancement method based on the variational autoencoder produces a synthetic signal close to the distribution characteristics of the real sample,which proves the feasibility of this method.(2)In order to further improve the ability of the data enhancement model to restore the minority data features,this paper proposes a data enhancement method based on the Wasserstein autoencoder on the basis of the variational autoencoder.The Wasserstein autoencoder can make the latent variable encoding reflect differently distributed input samples to a greater extent,and generate more realistic data.The experimental results of data enhancement show that compared with the variational autoencoder,this method generates a synthetic signal closer to the characteristics of the real sample,which reflects the superiority of the Wasserstein autoencoder data enhancement method.(3)In order to improve the classification accuracy of data-driven fault diagnosis methods on unbalanced data sets,this paper proposes a fault diagnosis method(WAE＿CNN)combining Wasserstein autoencoder and convolutional neural network.This method mixes the artificial signal generated by the Wasserstein autoencoder with the original unbalanced data to form a class-balanced training set,which is finally used for convolutional neural network training to realize fault recognition.According to the experimental conclusions,the fault diagnosis method proposed in this paper produces a better classification effect than other methods,which verifies the effectiveness of the WAE＿CNN method.