Abstract
Rotating machinery continuously generates unlabeled monitoring data during long-term operation, and the data distribution dynamically evolves with changing operating conditions. Conventional static offline diagnostic methods lack continuous adaptation capability, making it difficult to simultaneously handle newly emerging fault categories and cross-condition distribution shifts. This paper investigates class-incremental unsupervised domain adaptation for fault diagnosis (CIUDA-FD). In this setting, a diagnostic model is required to incrementally learn emerging fault categories from an unlabeled target-domain data stream while retaining previously acquired diagnostic knowledge. To address this challenge, this paper proposes a reliable shared-class identification and selective adversarial alignment (ReSCI-SAA) framework. Specifically, a tri-criteria shared-class identification (TriSCI) module with confidence fusion is first designed to identify reliable shared classes by jointly considering cumulative prediction probability, intra-domain clustering compactness, and inter-domain similarity. Based on the identified shared classes, a pseudo-label-guided selective adversarial alignment strategy is developed to align reliable class distributions across domains and reduce negative transfer from unreliable target samples. Furthermore, a prototype-accumulated knowledge distillation mechanism is introduced to mitigate catastrophic forgetting during class-incremental adaptation. Experimental results on the public CWRU bearing dataset and a self-built motor fault dataset demonstrate that the proposed method achieves superior average diagnostic performance over the best-performing existing method across multiple transfer tasks.

This publication has 25 references indexed in Scilit: