MFYOLO: Improved UAV lightweighting algorithm for wind turbine blade surface visibility damage detection

Abstract

Ensuring the structural integrity of wind turbine blades in alignment with dual-carbon objectives is crucial for optimizing wind power generation. To address the challenge of identifying small and complex defect features while maintaining a lightweight and efficient algorithm, this study introduces an innovative image input methodology. This approach combines the original image, a super-resolution enhanced image, and a defect-magnified image, resulting in significant improvements in average accuracies (mAP50 and mAP50–95) of 94.7 % and 83.3 %, respectively. The proposed lightweight algorithm, MFYOLO, employs multilayer aggregated depth-separable convolutions for deep feature extraction, alongside spatial pyramid pooling to capture multi-scale details and reduce feature map dimensionality. Dynamic convolution and a generalized reverse bottleneck structure adjust the size and sampling strategy of the convolution kernel, while self-attention mechanisms and convolution module grouping effectively reduce the computational burden. Knowledge distillation further enhances MFYOLO's performance without increasing the parameter count. Compared to the benchmark, memory usage (RAM), computational demand (GFLOPs), and parameter count (Params) are reduced by 48.2 %, 47 %, and 48.9 %, respectively, while maintaining stable mAP50 and significantly improving mAP50–95 and accuracy. The method achieves 156 frames per second (FPS), with robust performance validated on multiple datasets.