Multi-Domain Joint Subclass Alignment Adaptive Network for Rotating Machinery Cross-Condition Fault Diagnosis

Unsupervised domain adaptation (UDA)-based methods have been widely used for cross-condition fault diagnosis in rotating machinery. However, these methods typically focus on aligning the marginal distributions between the source and target domain, and neglect the fusion of information from multiple source domains. These limitations reduce classification accuracy in the target working conditions. To address these issues, this paper proposes a novel multi-domain joint subclass alignment adaptive network (MJSAAN). First, a subclass interval module is introduced. When combined with the local maximum mean discrepancy (LMMD) loss, it facilitates subclass alignment across different domains and improves the separability between subclasses. Next, a multi-source domain adversarial module is developed to integrate information from multiple source domains, enhancing the coverage of diagnostic knowledge in the target domain. Finally, the LMMD loss is employed to compute confidence scores for multiple models in the target domain, improving diagnostic accuracy. The effectiveness of MJSAAN is validated through experiments on bearing and planetary gearboxes across various conditions.