A multi-scale wavelet and spectral kurtosis-guided framework for noise-resistant rolling bearing fault diagnosis
Yurui Jia, Dechen Yao, Jianwei Yang, Limin Jia, Tao Zhou, Quanyu Long
Abstract
The health status of rolling bearings is essential for ensuring the safe and stable operation of industrial equipment. However, vibration signals acquired under actual operating conditions are often contaminated by strong background noise. This contamination makes it difficult for traditional fault-diagnosis models to effectively extract subtle fault features, leading to low diagnostic accuracy and poor noise robustness. To address this issue, this paper proposes a novel noise-resistant diagnostic network based on multi-scale wavelet convolution and kurtosis guidance (MWKG-Net). It is the first network to employ the lightweight Multi-Scale Wavelet Convolution Block (MWConv1D) to adaptively decompose one-dimensional vibration signals into multiple frequency subbands. This design not only enlarges the receptive field and achieves preliminary noise suppression, but also captures multi-scale fault information precisely, while reducing model parameters and computational overhead. Furthermore, we design a Kurtosis-Guided Feature Re-weighting (KGFR) module that quantifies the prominence of fault impulses by computing the spectral kurtosis of feature maps and dynamically re-weights feature channels. This enhances fault-related impulse components and further suppresses noise and unrelated harmonics. By combining wavelet-based multi-resolution analysis with kurtosis-based impulse sensitivity, the proposed MWKG-Net significantly improves the extraction of robust fault features under strong noise. Moreover, the model’s lightweight design makes it suitable for deployment on resource-constrained industrial platforms. Experimental results on the MFS-MG, BJTU-RAO and Urban Rail Train datasets demonstrate that, compared with state-of-the-art methods, our model maintains very high diagnostic accuracy even under extremely low signal-to-noise-ratio conditions, showcasing exceptional noise robustness, computational efficiency, and generalization capability.