Robust bearing fault diagnosis using wavelet-enhanced multi-scale CNN with SE attention

Abstract Reliable bearing fault diagnosis is vital for the safe operation of rotating machinery. However, noise contamination in vibration signals remains a persistent barrier to accurate detection. This paper proposes a wavelet-enhanced multi-scale convolutional neural network with squeeze-and-excitation attention (WMS-SE-CNN), explicitly designed for noise-resilient diagnosis. The framework integrates three key innovations: (i) Gaussian noise augmentation before wavelet decomposition to emulate real-world disturbances and strengthen robustness, (ii) parallel convolutional branches with kernel sizes of 3, 5, and 7 to capture both impulsive transients and long-range oscillatory fault patterns, and (iii) a lightweight channel-wise SE recalibration module that highlights fault-relevant features with minimal computational overhead. Evaluations on two heterogeneous datasets—the Shandong University (SDU) dataset and the Case Western Reserve University (CWRU) benchmark—demonstrate not only high diagnostic accuracy but also strong cross-dataset generalization under noisy conditions. In contrast to prior CNN–wavelet or attention-based approaches that primarily address clean or single-domain data, WMS-SE-CNN advances both noise robustness and practical feasibility for industrial fault diagnosis.