Prototype-guided bi-level adversarial domain adaptation network for intelligent fault diagnosis of rotating machinery under various working conditions

Abstract

Recently, deep learning (DL) models have proved to be effective in intelligent fault diagnosis of rotating machinery. However, due to diverse working conditions, the existence of domain shift is pervasive, which limits the popularization and application of traditional DL models. To address the above issue, an end-to-end prototype-guided bi-level adversarial domain adaptation (DA) network, which consists of a feature learner, health state classifier, source prototype learning module, domain-level discriminator and several class-level discriminators, is proposed for intelligent cross-domain fault diagnosis (CDFD). On the one hand, the feature learner and the domain-level discriminator compete with each other for a marginal-level DA. On the other hand, the feature learner and these class-level discriminators jointly play a minimax game for a conditional-level DA. Moreover, the prototypes learned by the prototype learning module are integrated into the bi-level adversarial DA, which facilitates the diagnostic knowledge transfer across domains. Extensive CDFD tasks on bearing and gearbox platforms are carried out for method validation. The results indicate that the proposed method is feasible and promising to promote intelligent fault diagnosis performance in engineering applications.