Swin TransCNN: A Dual-Branch Hybrid Mode for Weak Fault Feature Identification of Pitch Bearings

Pitch bearings are critical rotating components within wind turbines, consistently subjected to combined wind loads and blade centrifugal forces. This operational regime renders them particularly susceptible to premature damage. Under the low-speed, high-load conditions characteristic of wind turbine operation, the resulting fault signals exhibit weak energy signatures and are prone to significant noise contamination. Conventional diagnostic methods often struggle to effectively capture both localized transient details and global structural patterns within these signals concurrently. To address this limitation, this paper proposes Swin TransCNN, a novel dual-branch hybrid model that integrates convolutional operations for local feature extraction with a global attention mechanism. The Local Feature Branch employs multi-layer residual convolutions to extract high-resolution details, enhancing sensitivity to subtle fault signatures. The Global Feature Branch utilizes a shifted-window mechanism to model long-range dependencies within the time-frequency representations of the signals, improving structural pattern capture. A dedicated feature fusion block, incorporating dual attention mechanisms (global and local attention), hierarchically integrates the complementary features from both branches, enhancing discriminative power and ultimately improving fault identification performance. Furthermore, to advance research on fault mechanisms and diagnostic methods under realistic conditions, this paper details the development of a specialized fault simulation test rig. Experimental results confirm that this method effectively extracts subtle fault features and significantly enhances diagnostic accuracy for wind turbine pitch bearings.