Physics-Informed Optimization of Quality Factors in Resonance-based Sparse Signal Decomposition for Motor Bearing Fault Diagnosis
To address the limitations of existing Resonance-based Sparse Signal Decomposition (RSSD) - which relies on manually selected quality factors - and data-driven methods lacking physical interpretability, this paper proposes an optimization method incorporating physical information for adaptive fault feature extraction in motor bearing fault diagnosis. This method integrates bearing spring-damping fault impact model into the loss function, establishing a physical correlation constraint between the quality factor and the system damping ratio to achieve adaptive optimization of the quality factor. Utilizing truncated unrolling and gradient approximation to ensure effective gradient backpropagation, resolving the coupling challenge between inner-layer RSSD optimization and outer-layer network parameter updates. Experimental results from the motor bearing fault diagnosis test bench demonstrate that the proposed method yields quality factors consistent with bearing dynamic characteristics under different rotational speed, outperforming empirical RSSD and genetic algorithm approaches in fault feature extraction.