Estimation Of High Speed Railway Vehicle Response From Track Irregularity Using Deep Learning Techniques

Track smoothness affects train safety and passenger comfort in high-speed railways.Scientific evaluation of the track quality is of great significance to maintain healthy track state and ensure trains run safely.The current track quality evaluation method based on track geometric amplitude cannot accurately reflect vehicle response,and it is difficult to find ections where poor vehicle response is caused by the composite track disease.Therefore,it is essential to establish an efficient vehicle response estimation model and use track irregularity to estimate the vehicle response to assist in evaluating track state.This paper established a response estimation model for high-speed railway vehicles based on deep learning technology,and realized the estimation of vehicle body acceleration,wheel-rail force and wheel load reduction rate based on track irregularity.The research work mainly includes:(1)The correlation between track irregularity and vehicle response is analyzed.Based on the measured and simulated data of track irregularity and vehicle response on the high-speed railway,the coherence function was used to analyze the significant wavelength band of the coherence between the track irregularity and the vehicle response.In the data preprocessing part,the wavelength components below 2 meters of the measured data were filtered out based on wavelet decomposition and reconstruction theory.(2)A CA-CNN-MUSE(Coordinate Attention-Convolutional Neural NetworksMulti-Scale Attention)vehicle response estimation model was proposed,which realized the estimation of vehicle body acceleration,wheel-rail force and wheel load reduction rate based on track irregularity.The model used Convolutional Neural Networks(CNN)to learn different wavelength features of track irregularities;and focused on important channel dimensions and key segment locations by introducing Coordinate Attention(CA)into CNN;at the same time,the model extracted sequence features at different scales and captured the long-term and short-term trends of sequences in parallel by taking advantage of Multi-Scale Attention(MUSE)composed of Multi-head Self Attention mechanism and Depthwise Separable Convolution.The experimental results showed that: on the measured-simulation data set,the correlation coefficient of this model was improved by10.42% compared with Long Short-Term Memory(LSTM);compared with CNN-LSTM,the correlation coefficient was improved by 4.69 %.(3)The CA-CNN-MUSE vehicle response estimation algorithm was transplanted based on the SE5 AI platform.The results indicated that the estimation accuracy of the CA-CNN-MUSE model on the SE5 AI platform was the same as that on the GPU platform.