Uncertainty-aware statistical regularization for domain-generalizable bearing fault diagnosis under unseen conditions.

In predictive maintenance for smart manufacturing, intelligent bearing fault diagnosis in industrial field settings faces critical challenges under variable operating conditions. The need for adaptive production scheduling leads to continuous fluctuations in rotational speed, load profiles, and environmental parameters, resulting in dynamic distribution shifts in acquired vibration data. Existing cross-condition fault diagnosis approaches primarily achieve domain generalization through distribution alignment of multi-source condition data. However, the interference of condition-related information can potentially lead to insufficient stability from the aligned features. Moreover, the forced alignment of data distributions under different conditions may lead to the omission of some important fault information. To address these challenges, an uncertainty-aware statistical regularization framework is proposed. First, a new domain alignment mechanism is designed to map the statistical vectors of data from various conditions into a unified distribution, thereby providing a stable basis for fault diagnosis. Second, an uncertainty-aware dynamic domain perturbation mechanism is developed to generate diverse fault features through domain perturbation, allowing for a more comprehensive understanding of the variation patterns of fault features under different conditions. By combining these two strategies, it is possible to fully utilize feature diversity while maintaining feature stability, thereby more accurately identifying faults. Experimental results show that our method outperforms the existing methods on both the ZJUT dataset and the PU dataset.