针对风力涡轮机表面缺陷类型多、尺度差异大与特征提取困难等问题,提出了改进YOLOv7(you only look once version 7)算法用于风力涡轮机表面缺陷检测.首先,采用渐进金字塔网络(asymptotic feature pyramid network,AFPN)替换YOLOv7 网络中的路径聚合特征金字塔网络(path aggregation feature pyramid network,PAFPN),解决了多尺度融合过程中特征丢失和退化问题,并降低了模型复杂度;其次,采用扩充的高效聚合网络(efficient layer aggregation network-wide,ELAN-W)模块替换了AFPN中的基础模块,提高了模型的特征提取能力;最后,在颈部网络输入端以卷积和空间组增强(spatial group-wise enhance,SGE)注意力机制构建了卷积注意力模块,提升了模型对检测目标的定位能力和检测性能.实验结果表明,改进YOLOv7 算法对风力涡轮机表面缺陷检测的平均精度均值、检测速度分别达到了 85.4%、133.0 帧/s,相较于原版YOLOv7 算法分别提升了 1.8%、17.7%.该研究成果能够有效地提升风力涡轮机表面缺陷检测性能.
Surface Defect Detection of Wind Turbines Based on Improved YOLOv7 Algorithm
Directing at the problems that the surface defects of wind turbines involve various types,significant differences in scale and difficulty in feature extraction,etc.,an improved YOLOv7(you only look once version 7)algorithm was proposed for wind turbine surface defect detection.Firstly,the asymptotic feature pyramid network(AFPN)was used to replace the path aggregation feature pyramid network(PAFPN)in the YOLOv7 neck network,which solved the problem of feature loss and degradation in the multi-scale fusion process and reduced the model complexity.Secondly,the efficient layer aggregation network-wide(ELAN-W)module was used to replace the introductory module in AFPN,which improved the feature extraction capability of the model.Finally,convolution and spatial group-wise enhance(SGE)attention mechanisms were used to build a convolutional attention module to improve the model's positioning ability and detection performance of detection targets.The experimental results showed that the mean average precision and detection speed of the improved YOLOv7 algorithm for surface defect detection of wind turbines reached 85.4%and 133.0 frames/s,respectively,which were improved by 1.8%and 17.7%compared to the original YOLOv7 algorithm.The research results effectively improved the surface defect detection performance of wind turbines.
万方数据
维普
