Multidomain Class-Imbalance Generalization With Fault Relationship-Induced Augmentation for Intelligent Fault Diagnosis

Domain generalization-based fault diagnosis (DGFD) has attracted considerable attention due to its potential to extend diagnostic knowledge to previously unseen operational conditions or machinery. However, the collected data in real-world situations exhibit severe class imbalance, which decreases the generalization ability of diagnostic models. Therefore, this article proposes a fault relationship-induced augmentation framework (FRAF) for multidomain class-imbalance generali- zation in fault diagnosis. A new data augmentation perspective that captures invariant interclass relationships across domains is developed. Relationship mappers transform normal samples into fault samples belonging to corresponding domains, which increases the sample number of fault classes and transfers the diversity of normal samples to fault samples. Extensive empirical analysis based on cross-working conditions and cross-machine tasks suggests the superiority of the proposed method.