Research On Rail Damage Recognition Based On Deep Learning

As China’s high-speed railway technology advances at an unprecedented pace,its stability,punctuality and other advantages are deeply loved by the public.Now the Fuxing EMU on the Beijing-Shanghai and Chengdu-Chongqing lines has run out of 350km/h operation speed,and China’s EMU has become a new business card representing China’s technology.At the same time,the extremely high operation speed also lays a hidden danger for traffic safety.To guarantee the efficient and safe operation of high-speed trains,it is necessary to conduct realtime and periodic detection of rail damage.Such a huge amount of work is obviously not possible by human beings;Non-destructive testing technology and traditional machine learning The accuracy of detection technology is often poor,and its anti-interference ability is weak.To address this issue,this paper focuses on the rail damage detection method based on deep learning and conducts a series of research work in four areas: the basic theory of rail damage and deep learning,the deployment of the rail damage detection system based on deep learning,the improvement of the deep learning network framework,and the design of the rail damage detection platform.(1)Firstly,the purpose and practical significance of rail surface damage detection were introduced,and the research status of traditional non-destructive testing technology and machine learning-based testing technology at home and abroad were expounded.Then the rail structure and damage are studied,and the main damage and the reason of the damage are analyzed.Secondly,the basic principle of object detection based on deep learning and var Io Us popular deep learning frameworks are analyzed.(2)The reason why YOLOv5 is selected as the basic algorithm in single-stage object detection is discussed,and the workflow,basic environment deployment and principle of YOLOv5 are described in detail.In view of the lack of high-quality rail damage DTAa sets,this study conducted field visits on a large number of railway lines in Jiangxi Province,and selfmade a rail surface damage DTAa set named My-Rail,which contains 800 high-resolution and high-quality RGB rail damage images,including a variety of dense,overlapping and unobvious rail damage images.And a variety of interferences are added manually,which can restore the real and complex rail damage detection scene to the maximum extent.Based on the My-Rail DTAaset,this study uses the YOLOv5 algorithm to train the YOLOv5 m model,and gives the evaluation indicators of the model and the test results of the model damage identification and detection experiment,and points out the shortcomings of the model.(3)Aiming at the problems of insufficient sensitivity,recall and robustness of YOLOv5 small targets,this study makes a series of improvements on YOLOv5,and proposes a CmyYOLOv5 network for rail damage small target recognition.Improvements over YOLOv5include: An improved Self-Attention mechanism was added to the Backbone network,and SPPF(Spatial Pyramid Pooling Faster)with modified activation function was used to replace SPP(Spatial Pyramid Pooling)in YOLOv5.Aiming at the problem of insufficient sensitivity of small objects,a small object detection layer is added to the YOLOv5 model,and a neck network Cmy-Bi FPN that can obtain more small object features is proposed to replace the PANet in YOLOv5 network.Finally,PLDIo U positioning damage function is used to further accelerate the speed of model positioning and improve the positioning accuracy of the model.The experimental results show that the Cmy-YOLOv5 m model trained by Cmy-YOLOv5 has better detection performance than the YOLOv5 m model trained by YOLOv5.The mean average Precision(Map)is increased by 44.8%,the Recall rate is increased by 34.4%,the F1-Score is increased by 24.3%,the precision is increased by 9.8%,and the detection speed is increased by8.50%.Compared with other models,Cmy-YOLOv5 m has better Map value,and its mean average Precision(Map)reaches 0.801,Recall is 0.887,F1-Score is 0.893,and precision is0.901.It is verified that the proposed scheme can provide a more favorable guarantee for rail damage detection.(4)Completed the design of the rail damage detection platform,including the design of the overall frame and the dynamic simulation experiment of the detection platform;And the selection of industrial camera,optical lens and lighting subsystem.The results show that the designed detection platform can ensure the operation of the image recognition system.