Weight Alignment Prototype Contrastive Network for Rotating Machinery Continual Fault Diagnosis Under Class Imbalanced Data

Continual fault diagnosis in fault-type incremental learning scenarios has attracted increasing attention due to the practical difficulty of collecting all fault types of target equipment in advance. However, most existing intelligent fault diagnosis methods rely on a batch learning paradigm, which lacks support for the incremental incorporation of new fault types and suffers from performance degradation. Moreover, the dynamic emergence of fault types inevitably aggravates data imbalance over time. To tackle these challenges, this study proposes a novel weight alignment prototype contrastive network (WAPCN), which enables effective continual fault diagnosis under both class imbalance and fault-type incremental conditions. The core innovation lies in the integration of non-learnable, mutually orthogonal class prototypes, which are generated via the Gram-Schmidt orthogonalization process on a unit hypersphere prior to each phase. These prototypes are incorporated into a prototype contrastive framework to guide the learning of highly representative and discriminative features. To address the evolving imbalance, a dynamic margin adjustment mechanism is designed to enforce intra-class compactness and maintain inter-class separation as new fault types emerge. Furthermore, a weight alignment strategy is introduced to rectify classifier bias and enhance recognition performance for new classes. Experimental results on multiple bearing datasets demonstrate the effectiveness and superiority of the proposed WAPCN compared to existing approaches.