MS-PAF: a physics-aware self-supervised multi-sensor alignment framework for rotating machinery fault diagnosis under complex operating conditions

Deep-learning-based fault diagnosis of rotating machinery in complex industrial environments is hindered by inter-sensor spectral inconsistency, non-stationary time–frequency coupling, and limited labeled data. To address these challenges, this paper proposes a Multi-Sensor Physics-Aware Alignment Framework (MS-PAF) that integrates physics priors into a self-supervised learning paradigm for robust multi-sensor feature fusion. A dual-branch time–frequency decoupling feature extractor is first developed, where adaptive frequency basis functions are introduced to jointly characterize broadband impulsive components and narrowband modulation patterns. A cross-sensor frequency alignment module is then designed to mitigate domain heterogeneity caused by sensor placement and transmission-path differences. Furthermore, a physics-aware self-supervised strategy constrains the attention mechanism through a spectral-consistency loss, enabling alignment of intrinsic fault signatures shared across sensors. The proposed method is evaluated on three public datasets, including a laboratory gearbox dataset, the CUMT pitch-bearing dataset, and the Tsinghua bearing–gear compound-fault dataset. Experimental results demonstrate that MS-PAF consistently outperforms state-of-the-art baselines in both accuracy and robustness. Ablation studies further verify the effectiveness of the proposed physics-consistency constraint under limited-label conditions.