Research On The Detection Algorithm Of Track Surface For Track Inspection Trolley Based On Deep Learning

The rail is the infrastructure of the railway track,which bears the huge pressure exerted by the train,and transmits the pressure to the sleeper bed,in addition to playing a role in guiding the direction of the train,so the rail must be maintained and overhauled in time to ensure good condition so that the train can run safely.With the increasing service life of rails,the problem of rail damage has become increasingly prominent,and the manpower and material resources invested in ensuring the safe operation of railways are increasing day by day.The automatic detection technology of track damage has received more and more attention.The traditional rail surface defect detection method are prone to missed detections,false detections and other situations,and it is difficult to achieve rapid,accurate and timely detection and other needs.In view of the above problems,this paper first compares three classical deep learning-based object detection algorithms,and trains and tests the rail surface defect dataset.Secondly,the method of improving the backbone network and optimizing the loss function is used to optimize the model,which greatly improves the detection accuracy.The following is the main content of the paper:Use machines instead of human eyes to do measurement and discrimination-machine vision,and design track inspection trolleys according to the design requirements of the Ministry of Railways.The inspection trolley needs to take track images at a speed of 15km/h while driving.The system integrates the light source and the camera,which can effectively improve the detection speed and reduce the probability of false detection.Firstly,the inspection trolley is designed to meet the needs of the railway public works operation site,and the inspection trolley is equipped with the image dynamic acquisition function.In order to solve the problem of residual image caused by shooting during the process,the acquisition system adopts a line scan camera,which can improve the detection efficiency and reduce the false detection.In terms of system design,Faster RCNN algorithm,SSD algorithm and YOLOv5 s algorithm,which are representative of deep learning algorithms,are applied to rail defect detection,and experiments are carried out on the premise of introducing the main network structures of each algorithm.The main defect types in the experimental data were surface abrasion,peeling off block,rail corrugation,fatigue crack,and tread crush.For the undisclosed rail dataset,this paper produces 1000 track surface defect datasets for algorithm training.Through experimental analysis and comparison,the Faster RCNN algorithm has the highest detection accuracy,but the slowest detection speed.Compared with Faster RCNN algorithm,the detection speed of SSD algorithm is significantly improved,but its missed detection rate is higher.YOLOv5s is most suitable for rail surface detection but less accurate than Faster RCNN,and introduces the main performance of the attention mechanism module.Under the premise that the experimental background remains unchanged,the feature extraction capability is improved by adding the SE module to the YOLOv5 s backbone network.According to the characteristics of rail surface defects,an improved YOLOv5 s algorithm based on loss function is proposed,and α-IOU Loss is used instead of CIOU Loss,so as to improve the accuracy of defect location.The m AP of the optimized YOLOv5 s rail defect identification algorithm is 95.5%.