A trackable multi-domain collaborative generative adversarial network for rotating machinery fault diagnosis

Obtaining sufficient balanced data is tricky in practical rotating machinery fault diagnosis tasks. It is a pressing real-world problem to accurately diagnose faults from imbalanced data. Generative adversarial networks have become a prevailing method to address this issue. However, its complex training mechanism and opaque architecture induce a credibility crisis, resulting in users not trusting the output completely. Therefore, a trackable multi-domain collaborative generative adversarial network (TMCGAN) is proposed for rotating machinery fault diagnosis. The core contribution of TMCGAN is achieving globally interpretable generation and credible classification, which encompasses three specific points. Firstly, a multi-domain collaborative adversarial strategy is built to sequentially learn key feature information of the signal from different domains, thereby achieving comprehensive training for multi-domain cooperative energy supply. Secondly, parallel frequency loss is designed to incorporate multi-dimensional frequency detail information, thus enriching the feedback and forming a more efficient closed loop for adversarial training. Finally, the streaming tracking factor is developed to elucidate the internal working mechanism, providing real-time tracking feedback to explain the underlying decision-making rationale, thereby enhancing interpretability. Two case studies demonstrate that the classifier empowered by TMCGAN achieves excellent performance in rotating machinery fault diagnosis, while also maintaining high credibility.