Detection And Recognition Of Railway Catenary Dropper Based On Deep Learning

Railway catenary is a transmission line that supplies power to trains,and its operating status is of great significance to the safe operation of railways.catenary dropper is an important suspension part of catenary.As the mechanical vibration of trains during operation and wear between parts,it is easy to appear Slack,breakage and other conditions affect the safe running of the train.The traditional manual inspection and manual screening of fault images have problems of low efficiency,long time,and high labor intensity,which cannot meet the needs of rapid and efficient detection.Therefore,it is of great significance to develop an intelligent detection method of catenary dropper.The research work of this paper is based on the real-time monitoring of the contact network status in the 3C system.The experimental data in this paper is derived from the contact network images collected by the 3C vehicle-mounted high-definition camera group,taking the dropper in the image as the research object,using deep learning and image processing technology realizes the intelligent detection positioning and status recognition of the dropper,main tasks as follows:In the process of making the experimental data set of the catenary dropper detection network,first use Label Img to annotate image data.At the same time,in order to reduce the labor to spend a lot of time annotating sample data,the Albumentations enhancement library is used to quickly enhance the labeled image data.In the production process of the experimental data of the catenary dropper state recognition network,first write a program to batch intercept the dropper area from the marked dropper image data,and then manually filter to obtain a large number of normal state samples and a small number of true fractures and relaxation Samples;In order to solve the problem of sample imbalance,first part of the pseudo-fault sample data is made manually,and then the amount of pseudo-fault samples is expanded by image enhancement methods.In the catenary dropper detection,YOLO v3 is used as the basic network,and the network is improved according to the dropper detection requirements.The first is to combine the dropper characteristics with K-means clustering to obtain a specific prior for the dropper detection Frame,to achieve accurate positioning of the dropper area.For the detection of dropper,considering the particularity of the size of the prior frame of the dropper,the prediction from the original three-scale feature map is improved to the two-scale feature map.Prediction,and finally filter out the final prediction results through confidence and NMS to achieve fast and efficient detection of catenary dropper.In the state recognition of catenary dropper,Squeezenet network is used as the basic network to identify the state of the catenary dropper.The process of recognizing the state of the dropper is to first cut out the area of the detected dropper from the contact image,and then input the dropper into the state recognition network to complete the classification of the state of the dropper.During the training process of the catenary dropper detection network and the state recognition network,they are respectively trained by transfer learning,which solves the limitations of insufficient model training with large parameters and accelerates the network convergence speed.