An improved YOLOv11n-DWR-CAA method for multi-type defect detection of wind turbine blades in UAV-based inspection

Abstract Reliable detection of wind turbine blade damage is essential for ensuring the safe and efficient operation of wind farms. This study proposes a UAV-based intelligent inspection framework supported by deep learning. A dedicated dataset was constructed through image cleaning, normalization, and augmentation, covering both blade condition recognition (normal/abnormal) and three representative defect types, namely scratches, cracks, and breakage. For blade condition recognition, comparative experiments across multiple convolutional neural network architectures identified VGG-16 as the best-performing model, and the final optimized VGG-16-based classifier achieved a test accuracy of 96.65%. For defect-type detection, an improved YOLOv11n detector was developed by integrating SIoU loss, context anchor attention, and dynamic weighted residual feature fusion to improve small-defect localization, suppress background interference, and strengthen multi-scale feature representation in UAV-acquired blade images. Results from both ablation and comparative experiments indicated that the improved YOLOv11n detector achieved a mAP@0.5 of 87.9%, surpassing the lightweight YOLOv3 variant, YOLOv5n, YOLOv6n, YOLOv8n, and YOLOv10n. These findings indicate that the proposed method improves detection accuracy, robustness, and efficiency under the reported experimental setting, and provides a useful technical reference for UAV-based intelligent blade inspection research in the wind energy sector.