International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD2023)

    The imbalance of bearing fault samples can bring about the problems of the unstable learning process of the classification model and low classification accuracy. A Wasserstein generative adversarial network model (VAE-WGAN-GP) that fuses a variational auto-encoder and data with Gradient Penalty (GP) is proposed in this work. First, the structure of the generator is improved to extract the hidden variables by feature coding through the encoding-decoding structure to extract the latent information; Then, the training process adopts the Wasserstein distance to measure the difference between the generated samples and the real samples score, and introduces the GP term so as to improve the stability of the fault diagnosis model; Finally, the fake samples with real features are generated by a game between the generator and the discriminator. The experimental results show that the proposed method can generate high-quality bearing fault samples and improve the fault diagnosis accuracy under imbalance conditions.