Cross-view selective adaptation and rectification for open-set domain adaptive fault diagnosis of rotating machinery

Abstract

In practical industrial scenarios, the scarcity of labeled data and unpredictable novel faults under dynamic working conditions severely degrade the performance of transfer fault diagnosis. Robustly classifying shared faults while accurately isolating unknown faults under such open-set conditions remains a formidable challenge. To address this issue, a Cross-view Selective Adaptation and Rectification (CSAR) framework is proposed, featuring three coordinated components: Boundary Unknown Separation (BUS), Momentum-weighted Selective Adaptation (MWSA), and Cross-view Consistency Rectification (CVCR). Specifically, built upon a parallel time-frequency architecture, the BUS module establishes an explicit boundary within the temporal stream representation to facilitate the isolation of unknown faults. Concurrently, the MWSA mechanism facilitates cross-domain distribution alignment in the spectral stream by integrating an entropy-aware dynamic weighting strategy and an unknown rejection loss, effectively preventing negative transfer. Furthermore, to seamlessly couple the dual representations, the CVCR module ensures cross-view consistency through dynamic information interaction, thereby mitigating semantic drift. Extensive experiments on the JNU, BJTU, and SDUST mechanical testbeds demonstrate the effectiveness and superiority of the proposed CSAR framework compared to existing open-set diagnostic methods.