Research On Image Recognition Method Of Rail Fastener Defects Based On Convolutional Neural Network

Spring bar is an important part of the track system.Its function is to fix the rail in the correct position and prevent the lateral and longitudinal displacement of the rail.In the process of train operation,there will be frequent impact and friction between the rail and the hub.With the accumulation of time,the fastener status will gradually deteriorate and a series of faults will appear.Fastener fault detection has always been an important work of Railway Public Works Department.At present,China’s railway departments have equipped a series of track inspection cars and comprehensive inspection cars,but in the actual application process,the problems such as insufficient number of inspection cars,inflexible use,some inspection cars do not have the ability of fastener fault detection and so on are exposed.China’s railway departments urgently need a kind of accurate,efficient,portable and low-cost track surface defect detection equipment.In recent years,with the continuous development of deep learning technology,excellent object detection algorithms are emerging,and defect detection technology based on machine vision is widely used.This paper introduces the common fastener defect detection equipment at present.Aiming at the problems exposed in its application,a fastener inspection car which can detect fastener defects in real time is designed,and the key parts are selected.An image acquisition unit based on linear array camera is designed in this paper.In order to obtain the image of fastener without residual 15km/h,and to obtain the complete image of fastener without distortion,A fastener location unit which can predict the scanning area of linear array camera and the relative position of fastener is designed in this paper.Based on the analysis of the working environment and working mode of fastener inspection system,the basic theory of deep learning and the mainstream target detection algorithm are studied in this paper.Finally,the YOLOv3 algorithm is selected to realize fastener defect detection.The improved YOLOv3 algorithm based on Mobile Net is put forward to replace the YOLOv3 Darknet53 backbone network and re-clustrate the size.The detection speed is greatly improved and can reach 24.0 FPS.on the JETSON AGX XAVIER edge computing platformFinally,the image acquisition experiment and fastener defect detection test are designed in this paper,and the function of fastener image acquisition and fastener defect real-time detection are tested respectively.The experimental results show that the fastener inspection system designed in this paper has the ability to detect fastener defects in real time,but some details still need to be optimized and improved.