Research On The Intelligent Identification Technology Of Dropper Damage In High-Speed Railway

The harsh working environment of the catenary droppers will have different degrees of damage due to the long-term high frequency mechanical vibration of vehicles,which will lay hidden dangers for the normal operation of high-speed railway.Therefore,it is especially important to identify the damage of the catenary droppers.Although there are already scholars apply the intelligent algorithm to the fault detection of droppers,but the accuracy rate is low and it fails to meet the actual work needs.To solve this problem,this thesis conducts a study on the intelligent identification algorithm of high-speed railway droppers damage in order to further improving the accuracy of identification,providing timely guidance to the maintenance of the catenary and eliminate safety hazards.The main research contents and conclusions are as follows:(1)Using the catenary droppers images collected by 4C system as the research object,the collected data are processed to extract the characteristic parameters of droppers and establish the database of droppers damage.It includes things like improving the visual effect of the image by using CLAHE for low light image processing;It also expands the number of samples by rotation,random color,random noise,etc.(2)As for the defects of broken and apparently slackened droppers are detected directly in the localization phase with the improved YOLOv5 s algorithm.For possibly normal droppers the algorithm,firstly localizing them from the original map,and then the localized droppers area is cropped from the original map to reduce irrelevant background interference for further detection of the droppers parts.The improved algorithm based on YOLOv5 s combines a large number of advanced network design advantages,such as Swin transformer,CBAM,Bi FPN,etc.Experimental results show that the improved YOLOv5 s algorithm has better performance in both accuracy and speed on the task of droppers detection.m AP@0.5and m AP@0.5:0.95 reaches 98.3% and 79.2%,respectively,which are 5.6% and10.3% improvement compared to the original YOLOv5 s algorithm.(3)The droppers which are judged to be probably normal are cropped according to the target frame,and the difficulty of judging subtle defects and faults which may appear.Jumping connection,CBAM and Dropout are added to the original GANomaly as optimization measures.The improved GANomaly is used to detect anomalies in the contact wire clip part,dropper wire part,and load-bearing cable clip part.The method only learns the data distribution of normal samples,and determines whether the input image is defective by comparing the difference between the input image and the image reconstructed by algorithm.The experiment has compared multiple models,it is found that the improved GANomaly is more effective in detecting abnormalities in the droppers,and the AUC for the detection of the three parts of the droppers are not less than 0.75.