FE-YOLO11n: a feature-enhanced detection model for multi-class surface damage of wind turbine blades

During operational service, wind turbine blades (WTBs) are exposed to a variety of external environmental factors leading to four primary types of surface damage: pinholes, oil contamination, leading-edge corrosion, and surface cracks. To address the challenges of high computational resource consumption and limited adaptability to diverse damage morphologies in traditional object detection methods, this paper proposes a feature-enhanced detection model named FE-YOLO11n. To enhance the edge feature extraction capabilities of the backbone network in the model, the Sobel Filter Downsampling (SFD) module is proposed. Furthermore, to address the high false detection rate of varying damage patterns caused by the fixed receptive field in the backbone network, the Variable Receptive Field (C3k2VRF) module is incorporated. Lastly, to increase the backbone network’s focus on irregularly shaped damage, the Inverted Residual Multi-Scale Attention (C2IRMA) module is proposed. Experimental results demonstrate that FE-YOLO11n achieves a mAP of 83.3% in detecting four types of damage, outperforming the baseline model by 1.9%. The model operates at a computational cost of 6.5 GFLOPs and achieves a detection speed of 119.7 FPS, meeting the requirements for real-time detection. Compared to mainstream object detection methods such as YOLOv12 and Hyper-YOLO, FE-YOLO11n demonstrates superior generalisation capabilities and robustness.