Research On Prediction Method Of Train Derailment Coefficient Based On Image Processing

Train derailment accident has the characteristics of strong destructiveness,great harmfulness,serious loss and wide influence range,so train derailment prevention has become an important subject of concern in the field of railway transportation safety.On the one hand,the railway is gradually developing towards the direction of high speed and heavy load,which leads to the increasing interaction between wheel and rail and affects the smoothness of operation,so it makes derailment accidents more likely.On the other hand,the existing theoretical system of train derailment analysis is not perfect and mature,and the train operation environment detection system can not keep up with the requirements of high speed development,which can not effectively reduce or even avoid the occurrence of derailment accidents.Therefore,it is imperative to form a recognized standard that can reasonably judge whether train derailment occurs.Considering the complex derailment mechanism and many factors affecting train derailment in this thesis,a prediction model of derailment coefficient based on multi-sensor fusion is constructed by using the camera to monitor the wheel-rail contact state of trains in real time.The purpose of this model is to put forward an effective method for real-time online evaluation of train operation safety,and provide reference for the upgrading of the train equipment involved in the automatic operation grade.The research contents are as follows:(1)This thesis studies four kinds of common rails(43kg/m,50kg/m,60kg/m,75kg/m)and three common tread types(conical tread,LM tread,LMA tread).According to the displacement between wheel and rail generated in the process of train snaking movement,combined with the position and number characteristics of wheel/rail contact points,the wheel/rail contact models in switchless section and switch section were established.Based on the conclusion of wheel lift curve with time put forward by China Academy of Railway Sciences,the safety grade range is established according to the different distance between wheel and rail center line.(2)In order to obtain better wheel/rail edge segmentation curves from dynamic wheel/rail contact images,a calibration method combining radial distortion,eccentric distortion and thin prism distortion was proposed to correct the distortion of wheel/rail contact images.The basic idea of the algorithm is: firstly,the camera parameters are solved under the pinhole distortion model,and then the initial parameters are updated iteratively by combining the distortion model until the accuracy requirement is reached,and the distortion coefficient is calculated.Experimental results show that the proposed algorithm can not only solve the camera’s internal and external parameters more quickly and accurately,but also calculate the distortion coefficient of the distortion model.In addition,by artificially adding noise into the image,the algorithm has strong anti-noise performance and can be adapted to the dynamic environment of train operation.Finally,distortion correction of wheel/rail contact images in switchless section and switch section is carried out,and good correction effect is obtained.(3)In view of the characteristics of dynamic acquisition of wheel-rail contact images,such as small gray difference and complex background,this thesis proposes an algorithm based on generative adversarial network for edge curve segmentation of wheel-rail contact region.The generation model is the key to extract wheel-rail edge curve with higher accuracy,so the generation model is built with the U-shape idea.In order to enhance the sensitivity of the network to the output and adjust the weight of the network to improve the segmentation accuracy,the residual idea is introduced into the network.In addition,in order to enlarge the field of view of the receiving region,extended convolution is introduced.Experimental results show that the accuracy,sensitivity,specificity and other indexes of the improved model are improved,so it can complete the edge curve extraction of switchless section and switch section with high precision.(4)Considering that lateral relative displacement wheel and rail,rate of wheel load reduction,acceleration and speed parameters have great influence on the train running stability,a derailment coefficient prediction model based on multi-sensor data fusion is constructed in this thesis.In order to solve the problem that BP neural network will fall into local optimization,PSO algorithm is used to optimize it.Finally,the measured data collected in Lanzhou North Marshalling Station are used to verify the results.The results show that the PSO-BP derailment coefficient prediction model has high accuracy and fast prediction speed.