The detection of wind turbine blades (WTBs) damage is crucial for improving power generation efficiency and extending the lifespan of turbines. However, traditional detection methods often suffer from false positives and missed detections, and they do not adequately account for complex weather conditions such as fog and snow. Therefore, this study proposes a WTBs damage detection model based on an improved YOLOv8, named AUD-YOLO. Firstly, the ADown module is integrated into the YOLOv8 backbone to replace some conventional convolutional down-sampling operations, decreasing the parameter count while boosting the model's capability to extract image features. Secondly, the model incorporates the UniRepLKNet large convolution kernel with the C2f module, enabling it to learn complex image features more comprehensively. Thirdly, a lightweight DySample dynamic up-sampler substitutes the nearest-neighbor interpolation up-sampling method in the original model, thereby obtaining richer semantic information. Experimental results show that the AUD-YOLO model demonstrates outstanding performance in detecting WTBs damage under complex and adverse weather conditions, achieving a 3% improvement in the mAP@0.5 metric and a 6.2% improvement in the mAP@0.5-0.95 metric compared to YOLOv8. Moreover, the model has only 2.5M parameters and 7.2 GFLOPs of computational complexity, this adaptation renders it appropriate for implementation in environments with constrained computational capacity, where precise detection is critical. Lastly, a mobile application named WTBs Damage Detection system is designed and developed, enabling mobile-based detection of WTBs damage.