Research On Automatic-Inspection System Of Rail-Surface Defects Based On Deep Learning

As an important supporting part of the general speed railway and high-speed railway,the rail is responsible for the major task of train’s operation safety.In the early,manual inspection by railway workers has become the past,with the development of computer hardware and software,rail surface defect detection by Deep Learning method gradually grow up,it can realize no contact,efficient and online detection,some people choose this method instead of the traditional manual method.Therefore,in this paper,we designed an automatic inspection system based on Deep Learning,which can realize the real-time detection of rail track flaws.The detection hardware system consists of five parts,namely image acquisition system,defect positioning system,energy consumption management system,walking system and data storage system,and the software part includes image pre-processing algorithm,rail surface defect detection algorithm and human-computer interaction platform.Firstly,we designed the proposal of this automatic detection system,and we selected and calculated hardware of each system,along with the software design requirements,software structure composition and software development tools are put forward.Secondly,due to the process of rail surface image acquisition,rail inspection system mechanical jitter,uneven road,encoder acquisition level error affect the image quality and the interference of sleeper,subgrade will seriously affect the accuracy of defect detection,so we need to carry out image pre-processing operation,including rail surface defect area positioning,image enhancement and data enhancement.The vertical projection method is used to determine the area of rail surface.The image enhancement is to make the image defect more prominent and make the detection task simpler.In the image enhancement,Non-Local Mean filtering(NLM)method is used to achieve background smoothing and MSRCP algorithm for image enhancement.In order to improve the robustness of network,we use the improved DCGAN generation model to expand the datasets,which can avoid under-fitting and heighten generalization rate.Moreover,due to the rail surface defects present multi-scale,small difference between background and prospect of images and there is a lot of noise,traditional defect detection algorithm has become more and more difficult to make online inspection come true,so we put forward a online inspection algorithm based on improved YOLOX,the algorithm can achieve96.1% accuracy and the FPS value of 49.74,which is a more accurate and real-time better rail surface defect detection algorithm.Then,in order to collect the real-time data during the operation of the vehicle,a data acquisition circuit is designed,which can collect the power quantity of the power supply,the mileage of the vehicle and the position information of the rail surface defects.Through the combined positioning method of encoder mileage positioning and GPS+BDS satellite positioning,we realize the defect positioning,and the combined positioning method is higher in the positioning accuracy than the individual positioning method,and finally use the standard Kalman filter method for data fusion,to achieve the high-precision positioning of the car.Finally,on the actual railway track,we test the image acquisition part,energy consumption management system,walking system,rail surface defect positioning system,surface defect test results,and design human-computer interaction platform which can real-time display the status information of rail inspection system,the experimental results show that the designed system can actualize autokinetic and fast detection,and can realize the car positioning and information record,finally,through timely notify the maintenance personnel can we ensure the safety of train operation.