Imbalanced Fault Diagnosis Method for Fan Bearing Based on Generative Model

Abstract

In the historical data of fan bearing, the ratio of fault data is much well below the regular data, and the fault diagnosis model established predicated upon this has low generalization ability. In response to this issue, this paper proposes an imbalanced fault diagnosis method for fan bearing based on generative model. Predicated upon one-dimensional convolution and one-dimensional deconvolution, the encoder and decoder were constructed respectively, and then the conditional variational autoencoder was established and trained to learn the data distribution corresponding to different bearing states. Then, the trained conditional variational autoencoder is used to generate fault samples, so that the number of fault samples for each type is equally consistent with the number of regular samples. Finally, the vibration signal samples corresponding to different bearing states with balanced sample numbers were input into the multi-scale residual convolution network to train the bearing fault diagnosis model. The superiority of this method were verified using the CRWU bearing test dataset. The results show that the proposed fault diagnosis framework has preferable applicability and accuracy for class-imbalance.