Fault diagnosis of wind turbine rolling bearings with two-stage fusion of multi-channel adaptation and multi-domain correlation

Abstract As a critical component of wind power generation systems, the fault diagnosis accuracy of fan rolling bearings directly impacts equipment operational reliability and economic efficiency. However, traditional multi-channel fusion methods predominantly employ fixed weighting strategies, which struggle to adapt to the dynamic changes in channel importance under complex operating conditions. Furthermore, they fail to fully leverage the intricate intrinsic correlations between fault features across the time domain, frequency domain (FD), and time- FD, resulting in suboptimal diagnostic accuracy. To address these challenges, this paper proposes a two-stage fusion fault diagnosis method featuring multi-channel adaptation and multi-domain correlation. First, a data-level adaptive fusion module based on channel attention is constructed. Through an end-to-end learning mechanism, it dynamically optimizes weights across channels, avoiding the limitations of manual prior assumptions. Second, a heterogeneous neural network is designed to concurrently extract features from the time, frequency, and spatiotemporal domains. Finally, a Transformer fusion module is introduced to establish cross-domain feature correlations. Leveraging self-attention mechanisms, it captures synergistic mapping relationships among multi-domain features. Through learnable token-like generation, it produces global discriminative representations for precise fault classification. Validation on the Huazhong University of Science and Technology public dataset demonstrates a diagnostic accuracy of 99.20%. The method maintains 87.35% accuracy under −4 dB strong noise conditions and exhibits only a 7.85% performance degradation during cross-dataset transfer to the Shandong University of Science and Technology dataset, significantly outperforming existing comparative methods. Ablation experiments and interpretability analysis further validate the effectiveness of each module and the consistency between model decisions and physical mechanisms, providing a robust and reliable solution for intelligent wind turbine bearing diagnosis.