Abstract In recent years, researchers have extensively studied deep learning to diagnose faults in rotating machinery. However, existing methods often fail to adequately capture and express fault information. The absence of sensitive features limits the model’s ability to extract discriminative characteristics, particularly in environments where signals are subject to substantial noise interference or where fault samples are scarce. To address the aforementioned issues, this paper proposes a fault diagnosis method based on multi-scale wavelet-weight initialisation and an adaptive gain mechanism (MSWAG). This method first employs multiple wavelet basis functions to initialise (MWTI) the convolutional kernel weights in the neural network’s first layer. By utilising a channel-wise convolutional architecture, it integrates the complementary information extracted by different wavelets during feature extraction. Secondly, an adaptive gain mechanism (AGM) is introduced that automatically learns a scaling factor for feature amplitudes, thereby achieving non-linear enhancement of key features and noise suppression. Finally, a multi-scale feature extractor (MSFE) is constructed to comprehensively capture discriminative information within fault signals by adapting perception strategies to different frequency components. Experimental results based on real rotating machinery data demonstrate that MSWAG exhibits superior diagnostic performance under both intense noise and sparse sample conditions, significantly outperforming existing mainstream methods. This finding highlights the method’s potential for enhancing key features and for practical applications.
Hu et al. (2026) studied this question.