MDHP: a reliable unsupervised domain adaptation framework for noise-robust fault diagnosis in rotating machinery

Abstract Unsupervised domain adaptation (UDA) aims to solve the problem of transferring knowledge from a labeled source domain to an unlabeled target domain. However, in actual industrial scenarios, background noise can cause the feature distributions of the source domain and the target domain to shift, thereby leading to degradation in the migration performance. To address this issue, this paper proposes a multi-level denoising and historical prior enhancement (MDHP) UDA framework. Firstly, by combining discrete wavelet transform with residual shrinkage mechanism, adaptive threshold adjustment is achieved in the frequency domain, and this process is combined with dynamic convolution cascading to construct a multi-level denoising network, achieving noise suppression and enhancing the robustness of feature representation. Secondly, an innovative feature enhancement network and gated feature selection module are designed. This module enhances the target domain features using historical prior features and achieves domain feature decoupling through a learnable dual-gated mechanism, effectively eliminating the prior bias of the source domain while retaining the inherent attributes of the target domain. Finally, based on vibration signal data from bearings and gears, experiments were conducted by simulating real industrial measurement scenarios with different signal-to-noise ratios and operating condition transfers. The results show that the MDHP framework demonstrates superior diagnostic performance compared to other typical domain adaptation methods.