A multihead deep convolutional transposed attention model for noise-immune diffusion with application to fault diagnosis

Accurate fault diagnosis of yaw motor bearings is crucial for wind turbine safety. However, deep learning models face dual challenges: scarce fault samples and significant label noise, which severely limit their performance. To address this, we propose a Noise-Immune Diffusion Model with Multi-depth Convolutional Head Transposed Attention (MDTANIDM). This method directly generates high-quality, accurately labeled fault samples from raw noisy data, bypassing error-prone label correction. It employs a confidence-based strategy to select samples and introduces a transposed attention mechanism that operates along the channel dimension rather than the spatial dimension, significantly reducing computational complexity from O(L2) to O(C2) while effectively capturing global channel dependencies. Evaluated on the CWRU bearing dataset under various label noise levels, our method achieves a superior diagnostic accuracy of 90.48% even with 40% label noise, outperforming existing approaches by 13%. This provides a robust solution for intelligent wind turbine maintenance.