Analysis Of Insulator Defect Detection Method Based On Convolutional Neural Networks

As an important part of overhead transmission lines,insulators directly affect the safety and stability of power system.Because insulators are exposed to harsh natural environment for a long time,self-explosion and breakage often occur.For insulator faults,traditional methods mainly use manual inspection or helicopter inspection with the help of detection tools,which not only has poor inspection effect,but also can not guarantee the safety of inspection personnel.In recent years,various power grid companies have applied unmanned aerial vehicles(UAVs)to the inspection of transmission lines.It has gradually become a mainstream trend to obtain the working status of insulators in real time by analyzing insulator images collected by UAVs.The images collected by UAV patrol have the characteristics of wide field of view and complex background,in the actual transmission line images,the insulator defect target image area is small and low resolution,which leads to the low precision of insulator defect detection in the current target detection algorithm.In order to solve this problem,this thesis combines with the CNN technology to carry out a series of studies.The main work of this thesis is as follows:(1)It is difficult to collect insulator defect images in practical engineering at present,and there are not enough training samples,which leads to poor insulator defect detection effect.To solve this problem,this thesis constructs a new insulator defect data set PLIDD(Power Line Insulator Defect Data Set)based on UAV inspection data set and public insulator data set provided by a power grid company.(2)To solve the problem that the defect target of insulators is small,which leads to fewer defect features,an improved insulator defect detection algorithm based on YOLOv3(GC-SPPYOLOv3)is proposed.The improvement of this algorithm mainly has the following three points: Firstly,based on Darknet53,Ghost Module and dual attention module are introduced,which not only reduces the parameters of the model,but also improves the feature expression ability of the network;Then,a spatial pyramid pooling module is added in front of the three Detection head of YOLOv3,which makes full use of the multi-scale local feature information of the same convolution layer,realizes the fusion of local features and global features,and enriches the representation level of insulator defect targets;Finally,in order to reduce the negative effect of target size on the width-height loss function,the width-height loss function is optimized.Experiments show that on the PLIDD data set,the method proposed in this thesis has achieved good performance in terms of detection accuracy and speed.(3)The idea of super-resolution reconstruction is introduced to solve the problem that the insulator defect target is easy to be blurred or obscured,which leads to low feature resolution.In this thesis,super-resolution module(SRCNN)is added to SSD,YOLOv3 and their improved algorithms for experimental analysis,and it is found that the improvement effect on SSD is the most obvious,so an improved insulator defect detection algorithm based on SSD(S-ASFFSSD)is proposed.Firstly,in order to reduce the loss of insulator defect detection speed as much as possible,a super-resolution feature generation module is proposed to improve the resolution of shallow feature map responsible for small target detection,so as to improve the ability to express the characteristics of insulator defects.Then,in the prediction stage,the adaptive spatial feature fusion module is introduced,and the multi-scale feature maps of SSD are effectively fused by training and learning.The test on the PLIDD dataset proves that the method in this thesis has a significant advantage over the classic target detection algorithm in terms of detection performance.