Class-weighted alignment and reliable self-training for cross-domain fault diagnosis under class-imbalanced conditions

Abstract

Domain adaptation has been extensively applied in cross-domain fault diagnosis. However, practical industrial systems still face a major unresolved challenge: the number of fault samples is limited, and their features are weak or indistinct. Such class imbalance between normal and fault states often leads to misalignment in conventional domain adaptation methods and hinders the extraction of domain-invariant representations. To address this issue, a novel fault diagnosis method named class-weighted alignment and reliable self-training (CWA-RST) is proposed. The core idea of this method is to introduce adaptive class-related weights to reduce the cross-domain scale discrepancy of same-class fault features. The feature extractor integrates a convolution neural network, a bidirectional gated recurrent unit, an attention module, and metric learning to enhance feature representation. A class-weighted alignment module is designed to amplify weak features in the minority class, enabling more effective cross-domain distribution alignment. Additionally, the reliable self-training mechanism is constructed to automatically select trustworthy pseudo-labeled samples from the target domain, thereby refining the model and enhancing its cross-domain adaptation capability. Extensive experiments on two public datasets and one real-world industrial system demonstrate that the proposed method achieves superior cross-domain diagnostic performance under class-imbalanced conditions.