Enhanced Lightweight Multi-scale CNN with Resonance-dominant Frequency-band Learning for Robust Bearing Fault Diagnosis

Rolling bearings are essential components in rotating machinery, and reliable early fault detection is crucial for avoiding unexpected failures and reducing maintenance costs. While deep learning techniques have achieved notable progress, most prior studies have emphasized feature enrichment without thoroughly examining vibration spectral characteristics or establishing mechanisms to select the most informative time–frequency patterns. To exploit this physical insight, this study proposes an enhanced lightweight multi-scale convolutional neural network (CNN) with adaptive frequency-band learning for robust bearing fault diagnosis. Vibration signals are first band-limited to the resonance-dominant range and transformed using the Continuous Wavelet Transform with the Morlet wavelet, improving signal-to-noise ratio while preserving impulsive fault components. The proposed architecture integrates depthwise separable convolutions, multi-scale parallel branches, Squeeze-and-Excitation channel recalibration, and Atrous Spatial Pyramid Pooling to capture discriminative resonance-sensitive features with reduced computational complexity. Grad-CAM is employed to visualize the resonance bands contributing to classification decisions. Experiments on the CWRU dataset achieve accuracy above 99% with only 7.69 M parameters. Robustness under additive noise and cross-domain validation on the HUST dataset confirm strong generalization and domain invariance.