Research On Defect Detection Of Power Grid Inspection Insulator Based On Variable Convolution And Attention Mechanism

When the high-speed development that China has made in the field of Chinese economy,power construction is advancing vigorously,stable power grid supply has become an important guarantee for people’s good lives.However,due to the complex and diverse environment such as my country’s terrain and climate,transmission lines are exposed to nature’s harsh environment for a long time,and problems such as self-blasting damage are prone to occur.In order to ensure the stable and safe operation of my country’s the power grid,the transmission line must regularly arrange for staff to conduct a comprehensive inspection of the power grid.Due to the increasingly prominent contradiction between the growing scale of transmission lines and the incapable of rapid growth of manual overhaul capabilities,coupled with the increasing maturity of drone technology in recent years,the transition from "manual inspection" to "machine inspection" is imminent..Therefore,using UAV aerial photography to automatically detect defects in transmission lines has become a hot research issue.This thesis takes the important component insulators in the transmission line as the research object,first locates and then classifies the inspection images obtained by the drone,adjusts and optimizes on the basis of the existing Faster RCNN network,and realizes the detection and recognition classification of insulator defects.This article mainly does the following work:(1)Data set: Manually screen and label a large number of power grid inspection image data taken by drones.Then,aiming at the problem of insufficient negative samples,data expansion is carried out by rotating,horizontally flipping,and increasing contrast,so as to ensure the balance of data samples.(2)Insulator positioning: In this thesis,the two-stage target detection model Faster RCNN with higher accuracy is selected and optimized.The minimum convex set and minimum convex set loss function are used to replace the original IOU and loss function.According to the slender insulator The original anchor ratio is improved,and variable convolution is added to the RPN network of Faster RCNN,which increases the receptive field,obtains a better sampling effect,and improves the accuracy of insulator positioning in the grid inspection.(3)Insulator defect classification: The improved VGG16 classification network model is used to classify and identify the defects of the located insulator images,which are divided into normal and problematic categories.In view of the disadvantages of VGG16 network parameters and the weights more larger,global average pooling is used to replace the two methods of fully connected layer and network pruning to compress network parameters.The weight of the network is reduced,and the detection efficiency of the insulator defect classification and recognition is improved.Finally,through experiments on the grid inspection insulator images obtained by drone aerial photography,visual comparison before and after the improved method and related data analysis and building a simple grid inspection insulator defect detection interface,it can be concluded that the location of the grid inspection insulator is accurate There has been a certain improvement in the accuracy and recall rate of defect detection.