Deep learning-based methods have shown promising results in fault diagnosis, but research on interpretability and noise robustness still needs to be done. A multi-channel wide-kernel wavelet convolutional neural network is proposed to address these issues. Firstly, a first layer of multi-channel wide-kernel convolution is designed to fuse different weight information and suppress high-frequency noise. Secondly, a discrete wavelet transform block is designed to retain the low-frequency components of the discrete wavelet transform for signal denoising and feature dimension reduction. At the same time, Improved Balance Dynamic Adaptive Threshold is used to enhance the robustness of the model’s noise and the sparsity of features, making the model easier to optimize. Lastly, a power spectrum and normalized class activation mapping are designed to validate the post-hoc explanations of the model. The effectiveness and reliability of the Multi-Channel Wide Kernel Wavelet Convolutional Neural Network are verified through two gearbox datasets.

In practical scenarios, gearbox fault diagnosis faces the challenge of extremely scarce labeled data. Additionally, variations in operating conditions and differences in sensor installations exacerbate data distribution shifts, significantly increasing the difficulty of fault diagnosis. To address the above issues, this paper proposes a wavelet dynamic joint self-adaptive network guided by a pseudo-label alignment mechanism (MDJSN-DFL). First, the wavelet-efficient convolution module is designed based on wavelet convolution and efficient attention mechanisms. This module is used to construct a multi-wavelet convolution feature extractor to extract critical fault features at multiple levels. Secondly, to improve the classifier’s discriminability in the target domain, a transitional clustering-guided DFL is developed. This mechanism can capture fuzzy classification samples and improve the pseudo-label quality of the target domain. Finally, a dynamic joint mean square difference algorithm (DJSD) is proposed, which is composed of joint maximum mean square discrepancy and joint maximum mean discrepancy. The algorithm can adaptively adjust according to the dynamic balance factor to minimize the domain distribution discrepancy. Experiments on two different gearbox datasets show that MDJSN-DFL performs better in diagnostic scenarios under varying load conditions and different sensor installation setups, validating the proposed method’s effectiveness and superiority.