Physics-Informed Uncertainty-Aware Dynamic Balancing Learning Framework for Unsupervised Universal Domain Generalization Diagnosis Scenario

Conventional domain generalization (DG) assumes fully labeled source domains and consistent label spaces across domains. However, in industrial diagnostics, incomplete labeling is common, giving rise to the unsupervised universal DG (UUDG) challenge. This article proposes a physics-informed, uncertainty-aware dynamic balancing learning framework for UUDG bearing fault diagnosis scenarios. The physical knowledge-guided pseudo-label generation strategy can reduce the impact of noise on the fault feature frequency offset and obtain physical pseudo-labels for the source domain and target domain sample sets. The physical knowledge-assisted uncertainty-aware pseudo-label dynamic updating strategy can provide dynamic pseudo-labels with gradually increasing confidence for model training, further reduce noise interference, and improve the generalization performance of the model. To maximize the generalization performance of the model while ensuring the accuracy of the model, the weight dynamic balancing strategy for data-physical loss term based on stability testing and loss variance quantification is proposed to break through the accuracy-generalization dilemma, thereby obtaining a robust diagnosis model for the UUDG scenario. Comparative experiments, ablation experiments, noise resistance tests, computational efficiency analysis experiments, and model interpretability experiments are carried out to comprehensively analyze the diagnostic performance of the proposed method. Compared with the existing DG diagnosis methods, the proposed method achieves more convincing results.