Research On State Inspection Method Of Cotter Pin On Railway Catenary Based On Deep Learning

Cotter pin is an essential part of catenary support suspension device,It is widely used in nuts of various joints to avoid their loose or falling off.Additionally,its state has an important impact on the stability of catenary.At present,the development of intelligent state detection technology of cotter pin is limited.Therefore,the state detection of cotter pin is still rely on manual inspection,which is time-consuming,labor intensive and inefficient.In order to solve the above problems,the state inspection method of catenary cotter pin is studied.In addition,in order to know exactly which catenary the cotter pin of defect status belongs to,the method of recognition of catenary pole number plate is also studied.In view of poor imaging quality in high-speed catenary inspection vehicle,a location algorithm of pole number plate based on morphology transform and a segmentation method of character based on projection matrix are proposed.Firstly,the candidate areas of the number plate are obtained by morphology transform,and the structural features and HOG&SVM(Histogram of Oriented Gradient and Support Vector Machine)are used to choose them.Secondly,the upper and lower boundaries of the number plate are extracted by affine transformation and two horizontal projections.Then,the candidate segmentation points are obtained by using the extreme points of the projection matrix,and the variation coefficient of the distances is analyzed to obtain reasonable character segmentation points.At last,the characters are recognized by deep residual network.Test experiment results in normal-speed railway indicates that the accuracy of pole number location and character segmentation are 96.9% and 95% respectively.Compared with the traditional vertical projection segmentation method and the adaptive projection segmentation method,the accuracy of character segmentation is improved by 19.9% and 6.5% respectively.Aiming at the problems of high resolution images and changeable structures of catenary suspension monitoring device(4C),a fast positioning method of catenary cotter pin based on Tiny-YOLOv3 is proposed.This method uses fixed sliding window to scan high resolution catenary images.Because of the low efficiency of sliding window and the high similarity of images taken by the same camera,The perceptual hashing algorithm is used to calculate the similarity between images,and the candidate regions are used between images with high similarity.In addition,in order to avoid the problem of overlap between candidate regions with close distances,Floyd algorithm is used to merge candidate regions with close distances.The validity of this method is verified by testing the image of catenary with different structures.The numbers of normal samples and defect samples of the cotter pin are extremely unbalanced,so it is impossible to train classifiers to judge the state of cotter pin.In order to solve this problem,a state detection method of cotter pin based on Deep Labv3+ model is proposed.This method does' t need defect samples and can judge the loose state of the cotter pin.Firstly,the trained Deep Labv3+ model is used to extract the semantic information from the image of the cotter pin,including the head and tail information of the cotter pin.Then,the connecting areas between the head and the tail of the cotter pin are analyzed,and the state of the cotter pin is judged according to the analysis results.The test results of the cotter pin in different states show that the method is effective and can partly solve the problem of loose state detection of the cotter pin which can't be solved by the traditional algorithm.