Research And Design Of Rail Crack Detection System Based On Deep Learning

With the aggravation of the railway transportation task,the pressure of the rail gradually increases.Therefore,the aging and damage of the rail is inevitable,especially the surface of railways which have been used for a long time is prone to crack.In order to ensure the safety of railway operations,it is of great significance to detect and deal with the rail crack defects in time.The traditional detection methods can not meet the needs of the rail detection in terms of both resource consumption and time.This thesis proposed a rail crack detection model based on deep learning,and uses Raspberry pi3A+ for embedded platform transplantation,which is developed and applied to the rail detection software system.The main research contents are as follows:(1)This thesis investigates and analyzes the research status of the rail detection at home and abroad,and summarizes the related research content of Convolution Neural Network(CNN).In the field of target detection,the single-stage and two-stage deep learning detection algorithms are compared and analyzed.And we specifically analyze the series of Region-based CNN(R-CNN)networks and the series of You Only Look Once(YOLO)networks.This paper analyzes and compares the advantages of the fourth generation network YOLOv4 compared with other detection algorithms,which lays a theoretical foundation for the establishment of the rail crack defect detection model.(2)The rail data set is processed.Different crack data are selected and labeled as the data set used in the network model,and then the data set is processed according to the training requirements of YOLOv4 network.Meanwhile,in order to improve the effect of the network model and reduce the misjudgment rate,we carry out some processing methods,such as data enhancement,expanding the data set and adding confrontation samples.Based on this data set,the YOLOv4 model is trained to adjust and optimize the network according to the characteristics of detection category.The optimal network model is obtained by clustering algorithm and anchor frame calculation.Its detection accuracy can reach 88.6%,and it only takes 182 millisecond to detect an image on average using the experimental platform.Then,the Raspberry pi platform was used to transplant the network model and build a small detection device.According to the characteristics of the platform,the model is optimized and accelerated to realize the optimization of the model.The optimized model only takes 2.1 seconds to detect an image on the Raspberry pi 3A+ platform.(3)Based on the establishment of rail crack detection model and the transplantation of Raspberry pi,the software platform is developed on QT Creator with C++ and Python mixed programming relying on the "Track Circuit Simulation Platform" project.Through the design of four modules of rail crack detection system based on deep learning,the data set processing function,network model preview function,network model detection function and network model training function are realized.Ultimately,the development of rail crack detection system based on deep learning is completed and tested one by one,which proves that the system can realize the intelligent detection of rail crack,The efficiency of rail crack detection is improved.There are 61 figures,17 tables,and 42 references.