Open-set domain generalization method combining data augmentation and adaptive decision boundaries for fault diagnosis under unseen working conditions

Abstract

Domain generalization technology has attracted substantial attention in the fault diagnosis of unseen working conditions. However, most studies assume that the label spaces of source and target domains are consistent. In real-world applications, unknown faults may arise when the diagnostic method is applied to an unseen target domain. To address this, this study proposes a class-level boundary-based open-set domain generalization network (CBODGN). Specifically, the CBODGN leverages multiple source domains and employs adversarial learning to capture generalized domain-invariant features. Meanwhile, a centroid constraint module is designed to enhance the discriminability of these features by increasing intra-class compactness and inter-class separability of fault features. In addition, an adaptive fault class boundary module is designed to dynamically construct boundaries for known class faults based on the feature distribution of intra-class and inter-class, facilitating precise classification of known faults and accurately identifying unknown faults in the unseen target domain. Furthermore, a multi-domain feature augmentation module is incorporated to fuse fault information from multiple source domains, mitigating the impact of significant differences between source domains on learning generalized domain-invariant features. Subsequently, the proposed method is validated using two bearing datasets, with experimental results demonstrating superior diagnostic performance compared to existing approaches.