Research On Axle Positioning Of High-speed Railway Based On Deep Learning

With the rapid development of China’s high-speed railway technology,China has mastered the fundamental core technologies of high-speed railway in an all-round way.At the same time,related maintenance technologies for EMU trains are constantly being upgraded and enhanced.At present,the maintenance of motor trains in China has gradually developed from human inspection to machine inspection.First-class maintenance of rail trains is a maintenance task with short maintenance cycle and frequent maintenance in a complete maintenance service system in China at present.It directly influences the safety and efficiency of EMU trains in operation.At the same time,because the axle is a major supporting load part in the train running process,there will be higher requirements for maintenance ability and accuracy in the intelligent maintenance process.In this paper,the axle positioning problem in training primary inspection is studied.In the process of axle positioning research,due to the lack of relevant public data sets,this paper establishes an axle data set to solve this problem,which is based on the inspection pictures collected in different environmental conditions on the real scene,and on this basis,the work of manually labeling data is carried out.This axle data set can be used for deep learning object detection experiments,and provides powerful data support for subsequent experimental research.Aiming at the problem of EMU axle positioning,this paper starts with the object detection technology and edge detection technology based on meaningful learning,and puts forward different axle positioning methods according to the related problems in actual detection.It can summarizes as following:(1)Axle positioning method based on Yolo v4 object detection.Being dependent on the characteristics of axle data set,an axle location method combining deep learning object detection network,low illumination image enhancement,image horizontal line filtering and image morphology is proposed.The LECARM image enhancement technology is utilized to adjust the initial image pixel exposure,which improves the problems of minimal overall illumination and insufficient exposure of the collected original data in the actual detection environment.Yolo v4 object detection network is used to pre-select the initial position of the axle.In the actual research process,it is found that the object detection network is largely based on manual data tags when it completes the object location task.To solve this problem,the horizontal and vertical operator weights in Sobel operator are changed to extract the horizontal lines in the object area.And by combining the image morphology correlation method,the horizontal kernel is configured to locate the upper and lower edge lines of the axle respectively.The related experiments prove that the proposed axle positioning method based on Yolo v4 object detection can improve the problem that the object detection network relies heavily on manual tags,and at the same time improve the accuracy of axle positioning by using the object detection method to a certain extent.(2)Axle location method based on BDCN edge detection.This method is based on BDCN bidirectional cascade edge detection network and combined with improved HOG algorithm to locate the axle.This method reduces the influence of irrelevant pixels in the background on axle positioning to a certain extent,and improves the accuracy of axle positioning.Through comparative analysis of experimental data during the experiment,it is found that axle-like parts of the car body exist in very few images.The appearance of axle-like components will interfere with the results of axle positioning,resulting in the error detection of axle position.To solve this problem,this paper further proposes an axle location method which combines BDCN edge detection network with Yolo v4 object detection network.This method makes full use of the property that the object detection network can correctly screen the object axle,and avoids the problem of false detection caused by extracting wrong object information in the process of single edge detection.Finally,it is proved by relevant experimental results and experimental data that this method not only solves the problem of error detection of axles,but also improves the accuracy of overall axle positioning.