Reliable bearing fault diagnosis under complex operating conditions is often hindered by the loss of critical information during feature extraction, particularly for weak fault signatures embedded in vibration signals. To address this challenge, this work proposes a parallel multi-domain deep learning framework that emphasizes the preservation and complementary exploitation of time-domain, frequency-domain, and time–frequency representations. The proposed framework integrates temporal modeling to capture long-range signal evolution, phase-aware frequency-domain analysis to preserve amplitude–phase coherence, and transient-enhanced time–frequency representations to highlight weak impulsive features in noisy environments. To effectively integrate heterogeneous representations, a dynamic self-attention-based fusion strategy is introduced, enabling adaptive interaction and importance reweighting among multi-domain features. Experimental studies conducted on bearing datasets from Huazhong University of Science and Technology and the University of Cincinnati demonstrate that the proposed method achieves diagnostic accuracies of 99.63% and 99.82%, respectively, significantly outperforming state-of-the-art deep learning and multi-domain diagnostic methods, with accuracy improvements exceeding 20% compared to representative baseline models. Furthermore, ablation and robustness analyses confirm that the coordinated preservation and fusion of multi-domain information significantly enhance diagnostic reliability and generalization performance under complex operating conditions.
Hu et al. (Thu,) studied this question.