Research On Surface State Detection Method Of Catenary Insulator Based On Deep Learning

With the rapid development of high-speed railway,China’s railway system has also put forward higher requirements for the operation and maintenance of railway equipment.As an important part of the electrified railway,the effective and reliable monitoring and maintenance of the catenary is the premise to ensure the safe operation of the railway.Nowadays,the catenary suspension state detection and monitoring device(4C device)has been applied to the field of catenary state monitoring by the railway company.However,this technology is to detect the state of catenary components by artificial view and traditional image processing methods.This method has low detection efficiency and unstable detection results,and it is difficult to meet the requirements of intelligent development of railway systems.As an important component of the catenary system,the insulator has been in a complex environment for a long time and is affected by strong electric fields and various external forces.There may be breakdown inside the insulator,and the surface of the insulator may be damaged,flashover,surface foreign matter and other abnormalities state,thus affecting the normal operation of the insulator.In order to accurately and efficiently detect the state of the insulator,this thesis mainly takes the wrist-arm rod porcelain insulator in the catenary image collected by the 4C device as the research object,and studies the detection of two typical abnormal states on the surface of the insulator.First of all,this thesis introduces the relevant knowledge of convolutional neural networks and deep learning target detection algorithms.In order to accurately and effectively identify insulators of different scales in catenary images,the Faster R-CNN algorithm with higher recognition accuracy is selected as the insulator location and identification algorithm.The feature extraction network and Non-Maximum Suppression(NMS)algorithm of the original algorithm are optimized and improved to improve the detection.Compared with other algorithms,the results show that the improved Faster R-CNN algorithm can significantly improve the detection accuracy of catenary insulators.Secondly,using the grayscale texture features of the insulator images,the insulators can be effectively classified into two types: normal surface state insulators and surface state abnormal insulators by SVM algorithm.Then,according to the periodic characteristics of the grayscale on the surface of the insulator,the grayscale integral projection method is used to identify and classify two abnormal types of insulator tile defects and foreign matter inclusions between insulator tiles.Finally,the YOLO v5 algorithm is applied to the insulator surface state detection task.Because the combined structure of Feature Pyramid Network(FPN)and Path Aggregation Network(PAN)is used in the network structure of YOLO v5 algorithm,it can obtain feature maps with rich semantic and location information.The Mosaic data enhancement method is also used by it,which randomly scales to expand the number of small objects,improves the robustness of the network,and makes the YOLO v5 algorithm outstanding in the detection of small objects.At the same time,in order to solve the problem of few samples of abnormal state on the surface of insulator,Deep Convolutional Generative Adversarial Networks(DCGAN)is used to generate insulator images of abnormal surface state to expand the data.By detecting the surface state of the insulator under the same experimental conditions as other algorithms,the experimental results show that the YOLO v5 algorithm has a higher recognition accuracy for various abnormal states on the surface of the insulator than other algorithms.