基于改进的YOLOv7和无人机航拍技术的风机叶片缺陷检测

Defect Detection of Fan Blades Based on Improved YOLOv7 and Drone Aerial Photography Technology

摘要: 提出一种基于改进YOLOv7算法的风电叶片表面缺陷检测方法。该方法通过改进YOLOv7模型，提高风机叶片缺陷检测算法的准确性和效率，使用可切换空洞卷积代替原始模型中的MPConv，强化模型对不同尺度缺陷的敏感程度。引入CoordATT注意力模块，增强模型整体对模糊特征和小目标特征的关注程度。替换CIoU坐标损失函数为Wise-IoU，提高模型检测能力的精确度。在自建风电叶片数据集上进行实验验证，结果表明改进YOLOv7模型的平均精度均值提高了1.8%，检测速度达到了57fps满足无人机巡检实时检测需求。通过对比实验，改进后的模型在mAP、FPS、Precision、Recall等性能指标下优于YOLOv5s，Faster R-CNN，SSD等模型。该方法提高了无人机自动化检测风机叶片缺陷的能力。

Abstract: A surface defect detection method for wind turbine blades based on the improved YOLOv7 algo-rithm is proposed. This method improves the YOLOv7 model and in order to improve the accuracy and efficiency of the fan blade defect detection algorithm, uses switchable cavity convolution to re-place MPConv in the original model, enhancing the sensitivity of the model to defects of different scales. The CoordATT attention module is introduced to enhance the overall model’s attention to fuzzy features and small target features. The CIoU coordinate Loss function is replaced with Wise-IoU to improve the accuracy of model detection capability. Experimental verification was conducted on a self-built wind turbine blade dataset, and the results showed that the average accu-racy of the improved YOLOv7 model increased by 1.8%, and the detection speed reached 57fps to meet real-time detection requirements. Through comparative experiments, the improved model is superior to YOLOv5s, Faster R-CNN, SSD and other models under mAP, FPS, Precision, Recall and other performance indicators. This method improves the ability of unmanned aerial vehicles to au-tomatically detect defects in fan blades.