A multi-band enhanced dynamic learning method for robust fault diagnosis under signal and label noise.

Accurate fault diagnosis of rotating machinery from vibration measurements faces challenges from real-world data imperfections, especially the concurrent presence of measurement noise and label noise. Measurement noise corrupts vibration signals and blurs fault patterns, whereas label noise compromises supervision reliability and induces biased optimization in deep diagnostic models. To address this compound noise challenge, this paper proposes a robust framework named MSD-DCSFNet. The framework applies a multi-band signal decomposition module combining variational mode decomposition and the fast Fourier transform to construct a structured multi-channel spectral representation. This step enhances the visibility of harmonics, sidebands, and impact-related bands under low signal-to-noise conditions. A deep collaborative synergy feature network featuring a dual-branch architecture subsequently learns complementary spectral-structural cues and impact-transient cues, producing highly discriminative features. A dynamic asymmetric cross-entropy loss is also introduced to improve training robustness. It adaptively adjusts sample weights according to prediction confidence, suppressing high-risk supervision from corrupted labels and unstable measurements. Extensive experiments on the CWRU dataset, the SEU dataset, and a self-built dataset demonstrate that the proposed framework maintains over 90% diagnostic accuracy even under a 0 dB signal-to-noise ratio combined with a 60% label noise rate. These results validate its practical value for diagnosis using imperfect industrial data.