A dual-perspective joint domain generalization network for bearing fault diagnosis under unseen working conditions

Domain generalization methods hold significant promise for bearing fault diagnosis under unseen working conditions. By learning cross-domain invariant fault representations from multiple source domains, recent domain generalization fault diagnosis (DGFD) methods have achieved remarkable progress. However, when domain discrepancies are substantial, learning cross-domain invariant representations becomes challenging, severely limiting the applicability of current DGFD approaches. To address this issue, this study proposes a Dual-Perspective Joint Domain Generalization Network (DPJDG) that integrates class-level invariance and cross-domain invariance. Specifically, considering that all source domains share the same label space, a data augmentation strategy with semantic consistency is developed. This strategy enhances feature diversity and connects discrete subdomains of the same class. As a result, a unified domain is constructed, enabling smooth extraction of class-invariant features within the domain. For inter-domain invariance, cross-domain invariant features are extracted by minimizing the maximum mean discrepancy (MMD) loss among source domains. Furthermore, a dual-confidence weighting strategy, applied at both instance and domain levels, is introduced to quantify the generalizability of features from different attributes and source domains. Finally, supervised contrastive learning is employed to ensure effective fusion of class-invariant and domain-invariant features, thereby enhancing diagnostic performance. Experimental results on six diverse datasets validate the effectiveness and superiority of the proposed method in improving generalization performance across various diagnostic tasks.