| As an important part of the power network,overhead transmission lines are exposed to the outdoors all year round,and the components in the lines,including insulators,will wear out or even fail directly over time.If these faulty components are not detected and eliminated in a timely manner,the safety of the power system will be seriously jeopardized,so regular inspection and maintenance operations are required to ensure the stable operation of transmission lines.In recent years,the technology of civil UAVs represented by DJI has become more and more mature,and it is a good choice to replace manual inspection by UAV inspection.But the number of aerial insulator images taken by UAV inspection is extremely large,and the traditional method cannot effectively and quickly process these images,so the current hot deep learning becomes the best choice.In this paper,we take the insulator images taken by UAV inspection as the research object,and use deep learning as the research method to realize the detection of insulators and their defects in the aerial images in real time during UAV inspection.To address the problem of insufficient feature extraction capability of VGG,the backbone network of Faster R-CNN,the fusion of residual network and feature pyramid network is used as the feature extraction network,and the average accuracy of insulators and their defects reaches 0.986 after the improvement of the feature extraction network.But the detection speed is still slow.For the problem of slow computation of data redundancy in Faster R-CNN,the ordinary convolution is replaced by deep separable convolution,and the SE channel attention module is introduced into the network,and the activation function in the network is improved.Finally,it is experimentally concluded that the average accuracy of the improved network is slightly decreased,but the detection speed is greatly improved to reach 32.05 FPS.In order to further improve the accuracy of the network,the anchor frame selection strategy is improved.Because of the large difference between insulator and defect data,the labeled data of insulator and defect are extracted separately,and then the two parts of data are clustered separately using Kmeans++ clustering method to obtain the improved anchor frame.The average accuracy of the improved network is improved through experiments.The average detection accuracy of the network is 0.983 and the detection speed is 32.13 FPS,which can effectively detect insulators and their defects on the used data set and meet the requirements of real-time detection on the used equipment. |
