Noise-robust multi-source domain adaptation for bearing fault diagnosis via hardness-aware noise injection and consistency learning

To address the coupled effects of strong noise and operating-condition variations on cross-domain bearing fault diagnosis, this paper proposes a multi-source unsupervised domain adaptation framework with hardness-aware noise injection. Firstly, a dual-path training mechanism with shared parameters is constructed. The original path transforms raw time-domain vibration signals by FFT for classification and domain alignment, whereas the noisy path injects Gaussian white noise in the time domain according to sample hardness and uses the resulting spectra to learn noise-robust representations. Sample hardness, which reflects the learning difficulty of a sample, is estimated from predictive uncertainty or prediction inconsistency among multiple subnetworks. Secondly, a network architecture composed of a shared feature extractor and multiple source-domain-specific subnetworks is designed. Normalization, efficient attention aggregation, and deformable convolutions are combined to enhance feature stability and extract hierarchical features, while multi-level maximum mean discrepancy (MMD) alignment is imposed on original-path features to reduce domain discrepancy. Hierarchical feature consistency and prediction consistency constraints are introduced to jointly optimize the two paths at both the representation and output levels. Finally, experiments on the PU, HUST, and CWRU datasets under signal-to-noise ratios (SNRs) from −6 dB to 4 dB (extended to −8 dB on CWRU) demonstrate more robust target-domain performance, especially under low-SNR conditions. Ablation and visualization results further verify the effectiveness of hardness-aware noise allocation, consistency learning, and multi-level alignment.