Research And Implementation Of Insulator Detection Method Based On The Improved Region Proposal Networks

As an important device in the power system,the insulation state directly affects the stability of the entire power system and the reliability of the power supply.Therefore,power inspection is an indispensable means to ensure the safe operation of the power grid.The emerging UAV inspections can obtain a large number of detailed inspection images through the equipped high-definition cameras and image transmission equipment.These inspection data are only analyzed and processed manually.The workload is huge and the efficiency is low.There are deviations caused by the experience and quality of the staff.Therefore,it is necessary to introduce an intelligent identification method.This paper combines the needs of the National Natural Science Foundation funded project“Research on target detection and fault identification based on deep feature representation”,and studies from the following aspects:(1)This paper uses the convolutional neural network to realize the detection of insulators,and solves the problems of poor robustness,poor generalization ability and low accuracy of traditional detection algorithms.Firstly,by studying the characteristics and extensive application of convolutional neural networks,combined with engineering requirements and hardware support,VGG16 is used as the feature extraction network and RPN as the detection method,and the network model suitable for this topic is constructed.Secondly,the UAV is used to collect samples of glass and ceramic insulators on different lines and times,and artificially expand and combine 3D artificial images of insulators to carry out sample expansion as a training sample.Then this paper chooses the open source Tensorflow as a tool,combined with the relevant adjustment technology to improve the network structure and optimize in the training process.By automatically learning the nature of the insulator features,the insulator detection in the complex aerial background is achieved,the training accuracy is 85%,and the test accuracy is 78%.(2)This paper proposes an improved region candidate network(Region Proposal Networks)for improving the detection accuracy of insulator targets in aerial images.In the case of incomplete insulator samples,the insulator training samples are expanded and improved by means of interception,rotation,mirroring and artificial synthesis;clustering statistics of the labeled boxes of manually labeled insulator samples are obtained,and the aspect ratio of the labeling frame is obtained.Distribution of the candidate frame size;layer-by-layer analysis of the feature extraction network VGG16,merging the second,third and fifth layer feature maps for insulator target recognition;changing the loss function to achieve dynamic adjustment The proportion of negative samples,thus solving the problem of imbalance between positive and negative samples during training.The experimental results show that compared with the original regional candidate network,the improved regional candidate network proposed in this paper can detect the insulator target in the aerial image more effectively,and the accuracy of the improved regional candidate network is improved by nearly 5 percentage point.(3)This paper tests and verifies the insulator test.Firstly,the insulator detection test was carried out for different backgrounds,different types and different quantities,and compared with the traditional DPM and HoG-based SVM algorithms,and the performance of the network was analyzed through visualization effects.Then the insulator detection algorithm under different backgrounds is verified.The verified insulator detection effect has reached the engineering requirements,effectively reflecting the value of the inspection data,and helping to improve the efficiency and intelligence level of the power inspection.