Cross-Domain Fault Diagnosis Method of Rotating Machinery Based on Classifier Prediction Consistency and Domain Generalization

To improve the accuracy and robustness of fault diagnosis with limited identically distributed labeled data and unknown diagnostic targets, this paper proposes a fault diagnosis method that integrates Classifier Prediction Consistency with Domain Generalization (CPC-DG). First, unlike traditional DG-based fault diagnosis, which primarily focuses on aligning global feature distributions and neglects inter-class boundary ambiguity, our method employs a dual-classifier adversarial training strategy, which refines the decision boundary by maximizing prediction divergence between classifiers on boundary samples and guides the shared feature extractor to learn more discriminative features, specifically keeping robust under shifted target domain. Second, to mitigate the noise associated with traditional pseudo-labeling techniques, which typically rely on the confidence outputs of a single classifier, we develop a pseudo-label generation and validation technique based on dual-classifier consensus. This approach assigns pseudo-labels only to unlabeled samples exhibiting high prediction consistency between classifiers by iterative updates to filter out unreliable labels that could impair model performance. Third, the boundary refinement and high-quality pseudo-labeling mechanisms are integrated with a global alignment strategy based on Maximum Mean Discrepancy (MMD) to minimize distribution discrepancies across multiple source domains. Lastly, extensive experiments on the CWRU and PU bearing benchmark datasets verify that the proposed method achieves superior diagnostic accuracy in scenarios characterized by label scarcity and an unknown target domain.