Fault diagnosis of wind turbine main shaft bearings based on lightweight attention residual network

Abstract Reliable fault diagnosis for wind turbine main shaft bearings faces dual challenges: existing deep learning models demand excessive computational resources for industrial deployment, and weak fault-induced acoustic emission signals under low-speed/heavy-load conditions are prone to noise interference, thereby degrading diagnostic reliability. To address these challenges, we propose SEDSNet, a lightweight residual network for bearing fault diagnosis. SEDSNet integrates depthwise separable convolutions and Squeeze-and-Excitation (SE) modules within a lightweight residual framework. By replacing standard convolutions, the architecture significantly reduces computational complexity while maintaining diagnostic accuracy. The SE module adaptively recalibrates channel-wise features to amplify fault-related patterns, while skip connections stabilize gradient flow. A streamlined four-layer residual structure ensures efficient feature extraction and inference. Experimental results demonstrate that the proposed method outperforms existing techniques in noisy conditions and across different operating scenarios. Ablation studies further validate the effectiveness of each component. Transfer learning experiments also confirm the model’s ability to generalize across different datasets. This research provides a computationally efficient and noise-robust solution for wind turbine bearing monitoring, bridging the gap between diagnostic accuracy and deployment practicality. The proposed methodology offers significant value for predictive maintenance systems in renewable energy infrastructure.