Dynamic Mask Cepstrum-Enhancement Network (DMC-EN): A dual-task physics-guided method based on cyclostationary features of bearing signals

Abstract

While deep learning has advanced bearing fault diagnosis, most models operate as black boxes, treating vibration signals as generic data and failing to integrate fundamental physical principles. To address this limitation, this paper introduces the Dynamic Mask Cepstrum-Enhancement Network (DMC-EN), a novel dual-task, physics-guided framework. Specifically, a physics-informed embedding strategy is introduced, which leverages the prior knowledge of signal periodicity to generate an adaptive mask by combining data-driven features and structural scores. This mask is applied in the cepstral domain to enhance fault-related features. The processed features are then channeled into a custom backbone network that performs two synergistic tasks: (1) accurate fault classification and (2) reconstruction of a physically coherent, enhanced time-domain signal. The entire learning process is governed by a composite loss function, where a physics-informed regularizer based on the envelope spectrum sparsity of the reconstructed signal—not only guides the reconstruction task but also provides feedback to the upstream mask generation. This unique, closed-loop constraint forces the network to learn representations that align with the cyclostationary nature of bearing fault signals. This holistic integration is the key to our method’s robustness. Extensive experiments on three public bearing data sets demonstrate that this integrated approach is robust in noisy and small-sample scenarios, confirming the efficacy and superiority of our proposed framework.