Defection Of Rail Surface Light Band And Research On Defect Recognition Method

As an important infrastructure in China’s transportation system,railway plays an important role in China’s transportation field.With the passage of time,the rail will suffer losses in different degree,including the rail surface spalling,depression,wave wear damage and so on.These damages will lead to the rail surface is not smooth,and then lead to the light band on the rail surface is abnormal.The state of rail light band shows the smoothness of the track.Traditional methods of light band recognition are easily disturbed by the external environment,and it’s difficult for them to recognize the complex light band.Furthermore,the light band with serious defects greatly affects the safety of the passengers.In this paper,the light band on rail surface is studied.Based on the method of computer vision,this paper not only identifies it,but also makes quantitative analysis,and identifies the serious defects on the rail.In order to understand the abnormal condition of the rail,the light bands on rail surface are identified firstly.A multi-label convolution neural network(MLCNN)based on deep learning was proposed,which predicted three properties including breadth,position and uniformity of light band.Three parallel classifiers were used to speed up the classification and reduce redundant features.The experimental results show that the average accuracy and recall of MLCNN are more than 96%,which are better than that of related traditional methods.This model can identify all kinds of anomalies in the light bands.Then,the light bands with abnormal breadth and position are analyzed quantitatively.A method of light band extraction from whole to local was proposed.Firstly,the overall position of the light band was extracted.Then the image was segmented according to the threshold value.The maximum interclass variance method of particle swarm optimization based on empirical threshold was used to determine the threshold.Combined with the overall position of the extracted light band,the precise position of the light band was obtained by scanning the image line by line.Finally,the central position and width of the light band was calculated.The fitting degree and accuracy are above 96%.The rail can be processed in time by the width or position deviation of the light band.For rail light band images with serious defects,a multi-channel defect detection network combining defect extraction and defect classification was proposed,which improved recall rate of defect detection.Defect extraction network located the defect positions,and located defect regions were sent to the classification network.The defects were divided into two categories: squat and spalling.The experimental results show that the recall of MCDDN for severe defect detection reaches more than 98%,which verifies the effective-ness of the proposed model.