Intelligent Recognition Method For Rail Damage Defects Of Heavy-load Railway Based On B-scan Image

As the speed of railway transportation is getting faster and the loading is more and more heavy,the number of rail internal damage is also increasing rapidly.At present,the internal damage detection of rails mainly uses the cooperation of rail damage detection vehicle and small rail damage detector to collect B-scan images,and then,the images have to be played back manually for damage detection,so it is workload and low efficiency.In particular,the long-term pressure on the heavy-load railway leads to the rail internal minor damages developing rapidly once they are formed,and the whole thing has a great potential safety risks.Based on this background,this paper analyzes and studies the intelligent detection method of rail damages based on heavy-load railway B-scan images,in order to improve the efficiency and accuracy of rail damage detection and ensure the safety of railway transportation.The research achievements and innovations of rail damage detection are as follows:Firstly,a detailed investigation on the current status of rail damage detection and object detection tasks at home and abroad has been done.Ultrasonic rail damage detection and common damage map identification methods are also analyzed.In order to ensure the accuracy of damage detection,the principle and characteristics of generating the rail damage detection data(B-scan images)are analyzed in detail.In addition,the label Img tool is used to complete the production of the sample labels.Secondly,a rail damage detection method MFDet based on multi-scale feature fusion is proposed.In this method,Faster R-CNN based on region proposal with high accuracy is selected as the main network.According to the damage characteristics in B-scan images,a multi-scale feature fusion structure TFPN is proposed.Besides,MLRM strategy for multi-layer feature mapping of candidate regions is proposed to further enhance the ability of feature expression and improve the accuracy of multi-scale rail damage prediction.After preprocessing the B-scan images,the relevant experiments have been performed.The experimental results show that the overall detection effect of the MFDet network model is better than Faster R-CNN and the mainstream single-stage object detection algorithm.Finally,In order to improve the accuracy of rail damage detection further more,a multi-stage cascade detection model MCDet is proposed which is based on MFDet.First,due to the existence of various external interference factors,the collected B-scan images have more clutter and complex background,which affect the accuracy of model detection.MCDet adopts a multi-stage training method to achieve adversarial learning between clutter features and damage features,so it can extract damage features more accurately and improve model detection performance.Second,due to irregular rail damage,the deviation in the regression position of the damage bounding box is common,and the difference is greater if the deviation mapped to the actual railway.MCDet uses cascade detectors with the different IOU threshold for each stage of the detector to control the quality of samples,which can locate the damage location more accurately and classify the damage.Experimental results show that the accuracy of the MCDet model is significantly improved with acceptable detection speed.