Insulator State Perception Of Overhead Transmission Lines Based On UAV Inspection Images

In recent years,as the society’s demand for the quality of electric power supply continues to increase,ensuring the safe and stable operation of transmission lines has become increasingly important.Insulators play an important role in fixing wires and insulating conductors in overhead lines,so real-time and accurate identification of overhead line insulators and their defects is the primary task of power inspection work.This thesis focuses on the unmanned aerial vehicle(UAV)inspection scenario of overhead lines and studies the deep learning-based insulator object recognition and defect detection methods.The research contents of this thesis include:(1)A study on an insulator detection method based on the Yolov5 s network,which achieves faster and more accurate recognition of overhead line insulator targets.To improve the training effect of the object detection network,a deep learning-based single-image super-resolution reconstruction insulator data augmentation method was studied,and experiments were conducted to investigate the impact of different initial learning rates on network performance indicators.Based on the Yolov5 object detection network,comparative experiments were conducted on five models with different network depths and convolutional kernel numbers to analyze the detection performance indicators and generalization ability of the Yolov5 s network on the homemade insulator dataset.(2)A lightweight improved insulator detection network based on Yolov5 s was proposed.To achieve fast and accurate positioning of insulators in UAV inspection scenarios,a lightweight improved network Yolov5s-GG-S2 based on GhostNetV2 was designed.The feature extraction module of the Yolov5 s backbone network was replaced by GhostNetV2,and the concept of ghost convolution was introduced to reduce the redundant computational complexity in the convolution process.The spatial shift attention mechanism S2-MLPV2 based on multilayer perceptron was introduced to compensate for the accuracy loss caused by the lightweight network,and the dynamic focusing WIo U target box regression loss function was introduced to replace the CIo U function used in the Yolov5 original network.The effectiveness of each module mentioned above was verified through ablation experiments.Comparative experiments on different models showed that the Yolov5s-GG-S2 network designed in this thesis had reduced parameters and computation,while improving detection accuracy and inference speed.(3)A Yolov8-based insulator defect detection model was proposed,which accurately identifies and classifies small target insulator defects in overhead lines.Firstly,the Yolov8 network was analyzed in detail,and the model complexity was reduced by parallel bottleneck layers in the backbone network and removing the front convolution layer before upsampling in the connection layer.The detection head used a decoupled head structure to separate the insulator defect detection and classification and positioning tasks to improve network performance.Comparative experiments were conducted on Yolov8 s and Yolov5 s networks,and the detection ability of Yolov8 s network for small target insulator defects was found to be superior to that of Yolov5 s network by comprehensively analyzing accuracy indicators,complexity indicators,and F1 scores.Based on the above research,this thesis has implemented an insulator detection network based on Yolov5 s and designed a lightweight insulator detection network that is easy to carry for UAV inspection scenarios.The deep learning-based insulator data augmentation method was also studied.A Yolov8-based insulator defect detection network was proposed.This thesis provides theoretical methods and experimental verification for the state perception of overhead line insulators based on UAV inspection images,with clear engineering application value and scientific research significance.