Contrast-enhanced adversarial domain generalization network with data augmentation and Bayesian inference for...

Advanced fault diagnosis techniques typically rely on large amounts of labeled data from known domains, yet collected datasets often suffer from significant class imbalance. In addition, unseen domain data, which arises from variations in operating conditions or equipment heterogeneity, usually lacks sufficient prior knowledge, leading to a marked decline in model performance. To address these issues, this study proposes a contrast-enhanced adversarial domain generalization framework that integrates data augmentation and Bayesian inference for imbalanced fault diagnosis of rolling bearings under diverse scenarios. Specifically, a correlation-guided adaptive mixup method is developed to alleviate class imbalance in source domains by adaptively adjusting sample weights according to their similarity. A feature extractor based on multiscale pinwheel-shaped convolutions and spatial-channel collaborative attention, together with a parallel multi-classifier training architecture, is then designed to enhance feature learning from multiple source domains. To further strengthen generalization, an inter-domain contrastive loss is incorporated into adversarial training, encouraging the model to capture more robust domain-invariant representations. Finally, a Bayesian fusion mechanism with dynamic weighting is introduced to integrate the complementary strengths of different classifiers for accurate recognition of unseen domain data. Two rolling bearing datasets are employed to construct cross-condition and cross-machine experimental tasks. Comparative results demonstrate that the proposed approach achieves superior diagnostic accuracy and strong generalization capability, thus providing a reliable solution for industrial equipment health monitoring.