Abstract

In intelligent fault diagnosis of rotating machinery with limited fault data, many studies have employed generative adversarial networks (GANs) to create synthetic data. However, GANs frequently face issues like training instability mode collapse when GANs are trained with small amounts of data, and stochastic nature of GANs in latent vectors sampling often results in generation of low-quality and less diversity, which potentially decreases the performance of fault diagnosis models. To tackle the training instability and uncertainty in latent vectors sampling, this paper introduces two new methods: a physicsguided GAN and a novel sampling approach. The proposed GAN model addresses training instability through additional use of spectra information, adversarial spectral loss, and a specialized model design. And the proposed sampling approach reduces uncertainty in latent vectors sampling by eliminating low-quality data while ensuring data diversity. The proposed GAN model with sampling method is validated on both rotor and rolling element bearing testbeds and the results reveal that the proposed synthetic signal generation framework significantly improves classification accuracy. Moreover, discussion about the impacts of the sampling on fidelity and diversity are performed. The results suggest that proposed synthetic signal generation framework marks a step forward in applying deep learning and GANs to industrial applications