Research On Estimation Of Track Longitudinal Level Irregularity Based On Deep Learning And Vehicle Body Response

Rail transit has become an important mode of transportation in China.At present,China is the country with the fastest operation speed,the longest mileage and the largest scale of high-speed railway under construction in the world.Because the line will produce geometric deformation due to the factors such as train running and natural conditions,and threaten the safe operation of the train,the inspection and early warning of the geometric state of the track is a crucial task.At present,due to the high cost and less configuration of the comprehensive inspection vehicle,the inspection cycle of the running line is up to 15 days,and the track irregularity and overrun during the maintenance interval will seriously threaten the traffic safety,so it is necessary to strengthen the monitoring of the track operation status.This paper focuses on the intelligent estimation algorithm of Track Irregularity based on the deep learning and track dynamic inspection data.The algorithm synthesizes the vehicle vibration data obtained by the vehicle measurement unit to estimate the track geometric irregularity.Based on the research results of this paper,the measured vehicle body response is input into the estimation model,which can obtain the track irregularity status in real time and provide scientific basis for line maintenance.The main work and research results of this paper include:(1)The preprocessing of track dynamic inspection data is carried out.In this paper,the outliers in orbit dynamic inspection data are identified and eliminated based on the Pa??a criterion,which ensures the accuracy of the data used in the model.On this basis,the wavelet threshold method is used to denoise the orbit inspection data,and the setting of parameters such as wavelet basis function and threshold function of different observation components in orbit inspection data is determined through the evaluation indexes such as root mean square error and signal-to-noise ratio It provides support for the following research of track irregularity estimation model.(2)Based on the coherence analysis,the correlation variables of the estimation model are established,and the frequency-domain characteristics of the correlation variables and their components are studied to guide the track maintenance.The coherence analysis of vehicle body acceleration and geometric irregularity shows that the coherence function between longitudinal level irregularity and vertical acceleration of vehicle body is high,and the coherence function between them is more than 0.8 in the spatial frequency range of 0.029 ? 0.0358),so longitudinal level irregularity is the main factor causing vertical vibration of vehicle body,so the input variables of the model are vertical acceleration of vehicle body,and the predicted output variables are longitudinal level irregularity.Based on the Welch method,the frequencydomain characteristics of the dynamic inspection data are analyzed,and the different frequency band components of the dynamic inspection data are extracted.According to the evaluation and management method of the track irregularity state in China,the thirdorder approximation data is determined as the frequency band components that can effectively reflect the track overrun.(3)Based on the theory of deep learning,the track irregularity estimation model which is suitable for solving practical engineering problems is built independently.Based on the vehicle response as the model input,the track geometric irregularity estimation model based on cyclic neural network is established,which is trained by supervised learning.Based on the analysis of the results of multiple control experiments,a model training scheme with batch size of 65536 and number of training periods of 20 is proposed,and the optimal model structure suitable for the estimation of track geometric irregularities is proposed: dual loop layer LSTM neural network,the number of units in the first and second cycle layers are 32 and 64 respectively,the third layer of the network is linear full connection layer,and Tan is used in the LSTM units in the first and second cycle layers respectively H and relu activation functions;rmsprop algorithm is used to update the parameters;the dropout rate of loop layer input is 0.2,and the current dropout rate of loop layer internal loop connection is 0.7.(4)This paper studies the prediction performance of the track irregularity estimation model in a variety of practical application scenarios,and analyzes the influence of data frequency band and input variables on the prediction of the model.Based on the estimation model of track irregularity,the prediction effect of the model is studied in the whole line,different time points and different practical application scenarios.The results show that the root mean square error of the model is less than 0.0449 mm,which has high accuracy and strong generalization.In addition,compared with the original data,the prediction accuracy of the model in the low-frequency band is slightly improved.When there is a big difference between the prediction and the training line,increasing the model input can improve the prediction performance of the model,and the improvement degree of the model prediction accuracy by adding the vehicle body transverse load as the model input is greater than that by introducing the vehicle running speed;when the vehicle body transverse load and the vehicle running speed are introduced as the model input variables at the same time,the prediction accuracy of the model is the highest.