Research On Estimating Car-body Acceleration With Track Irregularity Based On Deep Learning

With the continuous improvement of high-speed railway train running speed,track quality evaluation is required to be more accurate and faster.The existing evaluation methods generally use the measured track inspection data to calculate the track quality index to evaluate the track geometry and guide the track maintenance.Because the vibration state of vehicles is seldom considered in the evaluation process,some potential track defects that may cause vehicle accidents cannot be identified according to the existing track status evaluation standards.Car-body vibration acceleration is not only a comprehensive reflection of the input excitation of track irregularities and the characteristics of vehicle dynamic response,but also an intuitive embodiment of vehicle ride comfort.The track quality evaluation combined with the track irregularity and the car-body response can further identify the track defects effectively,which requires the establishment of a prediction model that can effectively link the track geometry with the car-body vibration response.At present,most scholars have carried out simulation analysis on track geometry and car-body dynamic response based on vehicle-track coupling dynamics model.However,the calculation efficiency of the traditional dynamics simulation model is low.The vehicle suspension,track stiffness and other parameters are greatly different from the real situation under the limitation of a variety of assumptions,which leads to the inconsistency between the simulation and the measured results.According to the time series and periodicity of the track inspection data,a Convolution Neural Network-Long short Term Memory(CNN-LSTM)model is proposed to estimate the acceleration of the car-body by taking the track geometric irregularity as the input.The main research work of this paper is as follow:(1)The coherence of vehicle acceleration and track irregularity was analyzed by using the coherence function.The author identified the irregularity components with strong correlation with vehicle response were identified,so the input and output of the estimation model were determined.The variation characteristics of different irregularities in space frequency domain are studied by means of track spectrum analysis.In time domain and frequency domain,the influence of train speed on vertical and lateral acceleration of the vehicle body is investigated.(2)An approach based on Approximate Bayesian Computation(ABC)is proposed to optimize vehicle suspension parameters,the estimated values of the suspension parameters are introduced into the multi-body dynamics model to predict the car-body acceleration,which is compared with the real-world car-body acceleration.The results show that compared with the calculation results of the original vehicle parameters,the results obtained by using the optimized vehicle parameters are more consistent with the measured data.(3)A CNN-LSTM combined model was established according to the temporal and spatial data transfer characteristics of track irregularity and vehicle acceleration.The model takes the track irregularity and train speed as inputs,uses CNN network to extract waveform features from track irregularity data,and uses LSTM network to transfer time series data,so as to estimate vehicle acceleration.In order to predict vehicle body acceleration under different working conditions,the vertical measured data were divided into 8 working conditions according to different track types and different infrastructure under track,The lateral measured data were divided into 6 working conditions according to different track types and different line shapes.Multi-body dynamics simulation model,LSTM model and CNN-LSTM combined model were used to estimate the acceleration under different working conditions,and the estimated results were compared with the measured data.The results show that CNN-LSTM model has higher computational efficiency and more accurate estimation results.