Research On The Application Of Deep Learning In Rail Profile Detection

With the increasing improvement of railway network construction,the railway operation mileage in China is increasing,and the railway related detection workload is also increasing.In order to change the current situation of manual rail detection by railway field maintenance personnel,improve the efficiency and accuracy of railway rail maintenance,and effectively ensure the safe and stable operation of railway transportation,the miniaturized and convenient non-contact rail contour detection equipment has become a hot spot in the research of railway maintenance detection equipment.In recent years,there have been many researches on rail profile detection methods based on line structured light in China.Among them,the method using the arc of rail waist and rail bottom as the registration feature has a better detection effect,but most of them stay in the research on the profile of standard rail section,ignoring the blocking characteristics of rail waist such as splint,fastener profile and ballast in the actual environment of railway site.With the improvement of hardware computing power and the rapid development of deep learning technology,image classification and target detection technology are becoming more and more mature.In order to adapt to the complex railway field environment,this paper realizes the rail contour image recognition function based on deep learning technology,and designs the corresponding rail wheel contour measurement software system.This paper mainly studies the rail contour recognition algorithm based on deep learning,and uses the deep learning target detection algorithm to study the rail contour image recognition.Faster R-CNN and YOLOv5 s target detection model are used for rail contour image recognition respectively.The comparative experiment is completed on the self-made rail contour image data set,and the experimental results are analyzed.The experimental results show that the YOLOv5 s target detection model has better detection accuracy and higher detection efficiency,but there is still room for improvement for rail contour image recognition.The model structure can be further optimized and the model can be lightweight to improve the detection efficiency.The improvement and optimization of YOLOv5 s is completed based on Ghost Net.In order to lighten the model and improve the detection efficiency,the network model is optimized by using Ghost Net structure and hard swish activation function,which effectively reduces the amount of parameter calculation of the model;Focalloss loss function is used to reduce the impact of uneven sample distribution on the network model and improve the attention of loss function to difficult samples.The experimental results show that the improved YOLOv5 s algorithm has significantly improved the reasoning speed,and the model size is far smaller than the original YOLOv5 s model,which achieves the purpose of lightweight model and improving detection efficiency.Based on the improved YOLOv5 s rail profile recognition model,a non-contact rail profile detection software is built.The rail profile recognition function is added to the rail profile detection process to immediately detect whether the rail profile image contains the arc characteristics of rail waist and rail bottom necessary to complete the wear measurement,adapt to the railway site working environment and improve the efficiency of rail profile detection.In order to verify the feasibility and universality of the research content in this paper,field tests were completed in the maintenance section of large track maintenance machinery of Chengdu Railway Bureau of ordinary speed railway and Jishou east station of Zhangji Huaihua line of high-speed railway,and ideal test results were obtained.