A novel interpretable physics-informed adaptive algorithm unrolling network for rolling bearings fault diagnosis

Bearing fault diagnosis plays a crucial role in ensuring the stable operation of mechanical equipment and minimizing downtime losses. Data-driven deep learning methods offer superior computational efficiency but suffer from poor interpretability and potential overfitting due to their black-box nature. To address this issue, a novel Physics-informed Adaptive Fast Iterative Shrinkage-Thresholding Network is proposed that leverages algorithm unrolling to achieve comprehensive interpretability while integrating mathematical model stability with deep learning capabilities. First, physics-informed convolution kernels based on Sinc functions are employed to effectively extract feature information within the fault frequency band range. Second, the Iterative Shrinkage-Thresholding Algorithm Network is improved by introducing sub-band adaptive thresholds and implementing multi-layer algorithm unrolling, enabling rapid convergence while extracting deep features. To achieve effective feature fusion at different unrolling stages, a Gaussian Context Transformer is incorporated. The proposed network was validated on multiple datasets. Experimental results demonstrate that this method enhances network interpretability and reduces parameters while exhibiting strong robustness and noise resistance in noisy environments and cross-working-condition scenarios.