Imbalanced domain adaptive incremental fault diagnosis method based on core focal anchor loss and adaptive prototype classifier

Incremental fault diagnosis is of great significance for the operation of equipment. However, the characteristic of across-domains imbalanced distribution of data leads the incremental model to favor the majority classes, resulting in a significant decline in diagnostic accuracy. Therefore, an imbalanced domain adaptive incremental fault diagnosis method based on core focal anchor loss and adaptive prototype classifier (CFAPC) is proposed. Specifically, in the initial stage, the focal loss is used to guide the model to focus on minority-class features, while anchor loss is employed to constrain the feature space and suppress minority-class drift. This strategy alleviates the decline in diagnostic accuracy by sharpening the decision boundaries between majority and minority classes. During the incremental learning stage, the method effectively adapts to imbalanced samples by retaining the initially trained network parameters and integrating an adaptive prototype classifier. This strategy avoids catastrophic forgetting in cross-domain imbalanced scenarios by constructing a fairer classification weight system. Besides, this classifier requires no backward iterative updates and can significantly enhance the efficiency of incremental fault diagnosis in imbalanced scenarios. Experiments performed on a wind turbine test rig show that the proposed method outperforms state-of-the-art approaches, especially in scenarios characterized by data imbalance and domain adaptation.