Bi-Generator Cooperative Domain Adversarial Neural Network for Bearing Fault Diagnosis

Currently, cross-domain diagnosis methods for achieving reliable diagnosis under different working conditions have become a research hotspot. Adversarial training is a common approach employed in cross-domain fault diagnosis methods, aiming to minimize domain discrepancies. Specifically, the goal of such methods is to generate features that can deceive the discriminator. However, completely deceiving the discriminator is often difficult for the features generated by existing methods, i.e., the similarity between features across domains is insufficient. To address this limitation, we propose an approach called the Bi-Generator Cooperative Domain Adversarial Neural Network (BGC-DANN). Our method incorporates a confidence difference module, which guides two generators to produce features with distinct domain information. Additionally, we introduce a feature fusion module that combines the generated features from both generators, further perplexing the domain discriminator. Through comprehensive experiments, we demonstrate that the features generated by BGC-DANN exhibit superior domain-invariant characteristics and achieve higher diagnostic accuracy compared with alternative methods. These findings contribute to ensuring the reliable operation of mechanical equipment.