A noise-invariant strategy for fault diagnosis in wind turbine main shaft bearings: A comparative study of Attention-based CNN models

Vibration monitoring is lauded for its efficacy in condition monitoring of turbine drivetrain components. However, the slow rotational speed of the main bearing obscures fault features in ambient noise. This study advanced a noise-invariant strategy for fault diagnosis in wind turbine main shaft bearings using CNN models with attention mechanisms. A baseline model is first trained, followed by the development of CNN models with different attention mechanisms to compare their performance under noisy vibration signals. The channel attention-based CNN model outperforms all other models, achieving the highest fault diagnostic accuracy (99.3%), precision (99.4%), recall (99.3%), and F1 score (99.3%). This suggests that emphasizing the most relevant feature channels significantly improves CNN fault classification under noisy conditions. Spatial Attention also contributed to improved performance but was slightly less effective compared to Channel Attention. The CBAM model, although a powerful attention mechanism, introduced unnecessary complexity and resulted in marginally lower performance due to the redundancy of combining both channel and spatial attention. Given the harsh operating conditions of turbine main shaft bearings and the susceptibility of turbine main bearing data to ambient noise, this study recommends integrating channel attention into CNN-based main bearing fault diagnostic models to improve fault detection accuracy.