UDDGN: Domain-Independent Compact Boundary Learning Method for Universal Diagnosis Domain Generation

Domain generalization (DG)-based intelligent fault diagnosis (IFD) methods have effectively addressed challenges posed by domain shifts when confronting unseen target domains. However, these DG-based methods assume identical fault modes in the source and target domains, which is unrealistic in complex real engineering scenarios. In practical applications, target fault modes are often unpredictable, leading to an unknown label space relationship between the source and target domains. To address this limitation, we introduce a novel scenario termed universal diagnosis domain generation (UDDG). The UDDG requires accurate identification of target samples corresponding to labels within a common label set while effectively rejecting unknown fault modes. To address this more challenging and practical problem, we propose the UDDG network (UDDGN). The UDDGN employs a supervised contrastive loss module via manifold mix-up and a compact boundary module based on classwise adversarial training to establish domain-independent compact boundaries for each mode in the source domain. When a target mode emerges, classification adheres to these compact boundaries; samples not satisfying any criteria are classified as “Unknown.” Extensive experiments on two test rigs validate the effectiveness and superiority of our proposed algorithm.