Research On Insulator Defect Detection Of Transmission Line Based On Deep Learning

Electric power is not only related to national economic security,but also closely related to the progress of human society.As an important equipment widely used in high-voltage transmission lines,insulators play an important role in electrical insulation and wire connection.Long term exposure to harsh environment may cause insulator wear,self explosion and other failures,which may cause a series of cascading failures,it brings about a serious threat to the stable and safe operation of electric power system.Compared with the traditional transmission line inspection method,UAV autonomous inspection technology is gradually favored by people for its simple operation,safe and reliable.At the same time,with the development of computer’s computing capability,machine learning,especially deep learning method,has achieved good results in dealing with this complex computer vision task,which is widely used in real-time visual translation,unmanned driving,video monitoring and other fields,providing a new idea for our transmission line insulator defect detection,However,the deep learning method is still in its infancy for defect detection of transmission line insulators.Therefore,this thesis optimizes and innovates the existing target detection model,and completes the task of defect detection of transmission line insulators from two aspects of insulator location and defect detection,the following work has been carried out for the small size of the defect.(1)Aiming at the problem of insulator data set of transmission line,this thesis selects the actual insulator pictures and the insulator pictures collected from the network for mixed annotation,and selects the insulators in different scenes for data annotation.Considering that the characteristics of insulators are relatively not obvious in the case of occlusion,this thesis only labels the insulators with relatively complete edge information without occlusion,and saves them in the most common VOC data set format.(2)In view of the relatively complex background of transmission lines and the wide acceptance range of UAV aerial photography,which leads to the low signal-to-noise ratio of insulators in the pictures,this thesis attempts to put forward a solution which combines insulator location and defect detection with cascaded convolution neural network.It can not only identify the insulator,but also give the specific location of insulator defects.Then an insulator detection algorithm based on improved YOLOv3 is designed,the idea of feature pyramid(FPN)is introduced to improve the robustness of insulator detection for multi-scale targets,and an IOU prediction branch is used to solve the mismatch between positioning accuracy and classification confidence.(3)In addition,in order to solve the problems of relatively unobvious insulator defect features and less defect samples,this thesis optimizes an insulator defect detection algorithm based on improved Faster R-CNN.On the basis of the network,a "top-down" feature module with up sampling is established,and the shallow features and deep features are multi-scale fused,the resolution of small target defect in the detection network is enhanced,and the IOU aware module is introduced to improve the mismatch between classification score and positioning accuracy,so as to complete the final insulator defect detection task.