UKANCNet: Multi-scale feature fusion with UKAN enhancement for micro wind turbine blade defect segmentation

Micro windmills, characterized by rapid installation and low wind-speed startup capabilities, are particularly suited for low-wind regions, providing reliable off-grid power. However, ensuring their operational stability remains challenging, necessitating efficient blade defect detection. Traditional encoder-decoder networks often suffer from information loss, a limitation addressed by multi-scale feature fusion. To further enhance defect localization, an advanced U-shaped Kolmogorov-Arnold network (UKAN) variant UKAN cross-scale network (UKANCNet) is designed for windmill blade defect identification. The key contributions can be summarized as follows: firstly, we introduce a fine-to-coarse feature integration method, which supersedes the conventional decoder with a hierarchical feature fusion mechanism to minimize information loss. Secondly, an attention mechanism is employed to suppress background noise while accentuating defect regions. Finally, an optimized loss function is used to enable automatic hyperparameter tuning and reduce computational overhead. Experimental results show that UKANCNet outperforms existing methods in various key indicators, notably reaching a remarkable 89.5% mean intersection over union (mIoU), and presents an efficient and precise solution for defect segmentation in micro and small wind turbine blades, significantly improving detection accuracy. By ensuring operational reliability, this work contributes to the stable performance of wind power systems and advances the intelligent development of renewable energy. Code is available at the following website:https://github.com/jiajia65/UKANCNet.