To address the limitations of traditional rolling bearing fault diagnosis methods—such as inadequate feature extraction, limited noise robustness, and weak generalization under variable working environments—this study proposes a fault diagnosis framework that integrates parallel axial attention into a ResNet architecture. First, continuous wavelet transform (CWT), known for its inherent noise immunity, is employed to convert vibration signals into time–frequency images, providing a noise-suppressed representation of fault characteristics. Convolutional layers are then applied to reduce image dimensionality and computational complexity. A parallel axial attention module is subsequently introduced to independently capture feature dependencies along the temporal and frequency axes, enhancing the model’s ability to focus on discriminative fault-related regions while filtering out irrelevant noise. ResNet serves as the backbone network for deep feature learning and classification. Experiments on the Case Western Reserve University bearing dataset show that the proposed method achieves an average diagnostic accuracy exceeding 99.67% under multiple operating regimes. Notably, it maintains an accuracy above 95% even in high-noise environments with signal-to-noise ratios (SNRs) ranging from −4 dB to 4 dB, significantly outperforming several existing convolutional neural network-based approaches. This demonstrates the strong anti-noise capability and robustness resulting from the synergistic combination of time–frequency analysis and attention mechanisms. Furthermore, cross-dataset validation using the Southeast University bearing dataset confirms the strong generalization ability of the method. These results indicate that the proposed approach exhibits excellent diagnostic performance and practical applicability in noisy and complex industrial environments.
Wang et al. (Sun,) studied this question.