A Novel Small Samples Fault Diagnosis Method Based on the Self-attention Wasserstein Generative Adversarial Network

In the current industrial production process, fault data of rotating machinery are often difficult to obtain, and a small amount of fault data can lead to insufficient training of the model and reduced diagnostic accuracy. In addition, the generative adversarial network as a data generation model has the disadvantage of unstable training leading to poor quality of generated data. To address the shortcomings mentioned above, this paper proposed a rotating machinery fault diagnosis method based on the self-attention gradient penalty Wasserstein generative adversarial network under small samples. First, the Wasserstein distance with gradient penalty and self-attention mechanism was introduced into the generative adversarial network model to construct the self-attention gradient penalty Wasserstein generative adversarial network (SA-WGAN-GP) to solve the problems of poor quality of simulation data and unstable training. Then, a convolutional variational autoencoder (Con-VAE)-based sample screening strategy was constructed to realize the screening of high-quality simulation data. Finally, the original data and the screened simulation data were fed into the classifier to achieve fault diagnosis. Experimental validation was performed on Case Western Reserve University and laboratory self-test rolling bearing data. The experimental results show that the method can generate higher-quality simulation data and obtain higher diagnostic accuracy compared with other small-sample methods.