Fault diagnosis method for imbalanced bearing data based on W-DCGAN

Abstract

In order to solve the problem of highly imbalanced bearing fault data collected due to the limitation of bearing working conditions, which causes the problem of reduced accuracy and stability of fault diagnosis. A fault diagnosis method for imbalanced bearing data based on wasserstein-deep convolution generative adversarial network (W -DCGAN) is proposed. Deep convolution generative adversarial network (DCGAN) is based on the traditional generative adversarial network (GAN) and introduces the convolutional neural network into the unsupervised learning training to improve the generative effect of the generative network. Wasserstein generative adversarial network (WGAN) introduces Wasserstein distance loss into GAN to enhance training stability. W-DCGAN is a combination of WGAN and DCGAN. The Wasserstein distance loss and the feature extraction capability of deep convolutional neural network are utilized in W-DCGAN to steadily generate high-quality samples as expanded data for bearing fault diagnosis applications. Experiments were conducted with the bearing dataset from Western Reserve University. It is shown that the improved diagnosis performance is mainly due to the balanced dataset by the signals generated by W-DCGAN. In the case of data imbalance, the proposed method can handle the imbalanced bearing fault diagnosis more effectively and can improve the accuracy of fault diagnosis.