MFTrans-DANN: unsupervised deep transfer learning model for noise-robust cross-domain diagnosis of rolling bearings

Abstract

Although deep learning has shown great promise in bearing fault diagnosis, standard models struggle to maintain accuracy in cross-domain applications subject to industrial noise and variable loads. To address this challenge, this paper proposes a lightweight transfer learning model MFTrans-domain-adversarial neural network (DANN) and applies it to the cross-domain fault diagnosis of rolling bearings under noisy backgrounds. Firstly, an adaptive morphological filtering module (AMFM) is proposed. Through a mechanism integrating learnable structuring elements with morphological operators, the AMFM adaptively adjusts the response weights of different branches according to the local features of the signal, thereby achieving dynamic noise suppression while effectively preserving fault-related features. Secondly, a lightweight Transformer encoding module is constructed. By interconnecting layers of multi-head self-attention, residual connections, normalization, and multi-layer perceptron, the lightweight Transformer encoding module not only simplifies the original Transformer architecture but also realizes the synergistic fusion of local nonlinear mapping and global feature interaction. Finally, the MFTrans-DANN transfer learning model is constructed based on the AMFM and the lightweight Transformer encoding module. Extensive experiments conducted on two distinct datasets validate the effectiveness of the proposed MFTrans-DANN. The results confirm that the model not only performs reliable fault diagnosis but also surpasses competing approaches in terms of accuracy and noise robustness. Particularly under low signal-to-noise ratio conditions, MFTrans-DANN achieves significantly better diagnostic performance compared to existing state-of-the-art methods.