PIFC-SABiLSTM: physics-inspired feature constraint with self-attention bidirectional LSTM for interpretable fault diagnosis of rolling bearings

Rolling bearings are critical components of rotating machinery. The operational status of these bearings directly impacts the stability and safety of the entire system. However, conventional data-driven models frequently demonstrate deficiencies in generalization and robustness when detecting early, subtle fault features. The present paper puts forth a novel approach to fault diagnosis, namely PIFC-SABiLSTM (Physics-Inspired Feature Constraint with Self-Attention BiLSTM). PIFC-SABiLSTM is a data-driven model that incorporates physics-inspired feature constraints with a self-attention mechanism and a bidirectional long short-term memory (BiLSTM) network. This integration aims to address the limitations of conventional data-driven models, particularly their poor generalization and limited robustness. This approach is predicated on the transformation of the statistical patterns of rolling bearing vibration characteristics during steady-state operation into differentiable and physically feasible domain constraints. A Physics-Inspired Feature Constraint (PIFC) layer is constructed, where a projection operator guides the learning process of the network to conform to actual vibration physics and suppress the generation of non-physical features. This approach effectively decouples and fuses multi-scale time-frequency features by integrating the global dependency modeling capability of self-attention mechanisms for periodic impact patterns with the bidirectional temporal capture capability of Bidirectional Long Short-Term Memory (BiLSTM) for transient events. The experimental findings on the MFPT (Mechanical Faults Prevention Technology) bearing dataset demonstrate that the proposed method achieves 98.70% accuracy and an F1 score exceeding 98%, thereby significantly outperforming existing mainstream methods. This research proposes a novel "data-mechanism" synergistic optimization approach for high-precision, intelligent fault diagnosis in complex industrial environments. The approach demonstrates significant practical value for engineering applications.