Enhancing wind turbine blade damage detection with YOLO-Wind.

This study presents an enhanced YOLOv8n framework for wind turbine surface damage detection, achieving 83.9% mAP@0.5 on the DTU dataset-a 2.3% improvement over baseline models. The architecture replaces standard convolutions with depthwise separable convolutions (DWConv) to optimize computational efficiency without compromising detection accuracy. The C2f module is restructured by integrating MobileNetV2's MBConv blocks with efficient channel attention (ECA), which improves gradient flow and enhances discriminative feature extraction for sub-pixel defects. Furthermore, a newly added P2 detection layer enhances multi-scale defect recognition in complex environments. Cross-dataset evaluations validate the model's robustness and adaptability, with 15.2-16.3% mAP@0.5 improvements in blade damage detection and a 3.1% accuracy gain in agricultural defect identification. Although the performance on steel surface datasets varies, the systematic integration of DWConv, MBConv, and ECA mechanisms establishes a methodological advancement for industrial vision applications. The proposed framework demonstrates an effective trade-off between precision and efficiency, contributing to real-world computer vision solutions for wind energy infrastructure maintenance.