Prototype-Informed Dynamic Adaptive Network for Robust Cross-Operating-Condition Fault Diagnosis of Rolling Bearings

Abstract

Cross-domain fault diagnosis of rolling bearings remains challenging because of substantial shifts in feature distributions and condition-induced instability in intra-class structures. To address these challenges, a prototype-informed dynamic adaptation network to enhance diagnostic robustness under varying operating conditions is proposed. The proposed network employs a physics-prior-guided dynamic feature modulation mechanism, whereby physical priors serve as conditional cues during representation learning to facilitate the extraction of physically consistent features. Intermediate features are subsequently modulated using a physics attention fusion mechanism. To explicitly characterize condition-dependent structural variations among samples of the same class, multiple diagnostic prototypes are constructed for each fault category, and a cross-domain prototype-alignment constraint is imposed to stabilize the decision boundaries. In addition, dynamic domain-specific batch normalization is incorporated as a foundational module to mitigate inter-domain discrepancies during both training and inference, thereby providing a normalized feature space for the subsequent modulation and classification components. Cross-condition transfer experiments on the Paderborn University bearing dataset demonstrate that the proposed method consistently outperforms representative unsupervised domain adaptation approaches across diverse transfer scenarios. Its performance approaches the upper bound achieved through fully supervised target-domain training. Furthermore, ablation studies and interpretability analyses verify that each key module contributes to enhanced diagnostic stability and robustness under varying operating conditions.