Continuous FMD-Net: A Physics-Driven Algorithm-Unrolling Network with Spatio-Temporal Dual Sparsity for Adaptive Fault Diagnosis

Abstract

Extracting and demodulating weak fault signatures from rotating machinery under heavy background noise remains a significant challenge. Conventional methods such as Feature Mode Decomposition (FMD) rely on manual parameter tuning and are highly susceptible to mode mixing. To address these limitations, this paper proposes Continuous FMD-Net, a fully differentiable framework for adaptive bearing fault diagnosis. The network unrolls the FMD algorithm into deep neural layers and incorporates three key designs: (i) a trainable time-domain spatial attention mask that adaptively determines the effective window length L of each channel’s filter via backpropagation; (ii) a channel gating mechanism that prunes redundant channels to automatically optimize the number of modes K; and (iii) a cross-correlation orthogonality penalty, embedded in a physics-informed joint loss function, that enforces mutual repulsion among parallel channels to effectively suppress mode mixing. Experiments on the Case Western Reserve University (CWRU) bearing dataset demonstrate that the proposed network achieves an accuracy approaching 100% under 0 dB noise, and reaches 90.13% on the −8 dB strong-noise blind test set, significantly outperforming all compared methods. Convergence curves and UMAP visualizations further confirm that the network learns highly decoupled and discriminative fault features. This work provides a feasible pathway for integrating physical demodulation priors with end-to-end deep learning.