Study On Correlation Model Between Track Irregularity And Vehicle-body Vibration Acceleration By Using Transformer

In the past decade,China’s high-speed railway has developed rapidly.Track structure inevitably degraded after years of high-intensity service.As a result,some tracks are in irregular state and affect the stability and safety of the train operation.The conventional track inspection train used to detect the state of tracks is expensive,time-consuming and occupies a lot of operating resources,while the vibration acceleration response of high-speed train when running on the track is easy to measure.If the correlation model between track irregularity and the vehicle-body vibration acceleration response can be established,a lot of manpower and material resources can be saved when the vehicle-body vibration acceleration response is used to evaluate track irregularity.And the estimation of the vehicle-body vibration acceleration response from track irregularity can assist in the evaluation of track state,which is of great significance.Therefore,this paper applies the most effective Transformer network in the field of natural language processing to establish the correlation model,puts forward the corresponding improvement strategy according to the characteristics of this subject.The main contents of this paper include:(1)An estimation model of track height irregularity based on Transformer Encoder network is established.In order to solve the problem of low precision of network modeling caused by direct application of Transformer Encoder,an input dimension augment strategy is proposed.And for the purpose of improving the precision of estimation,the masked multi-head attention mechanism is applied in this paper.Compared with the previous work based on LSTM network and GRU series networks,the mean absolute error of track height irregularity and long-wave height irregularity are reduced by 11.2% and 9.30% respectively by the estimation model proposed in this paper without significantly increasing the number of model parameters on the measured dataset of domestic high-speed railway.The correlation coefficient ρ increased from0.931 to 0.946 and from 0.906 to 0.922,respectively.(2)An estimation model of the vehicle-body vibration acceleration based on Transformer network is established.In order to solve the problem that the estimation model is difficult to implement Decoder Embedding effectively,an effective Decoder Embedding scheme is designed for this research subject.And for solving the Exposure bias phenomenon in the process of model training,a solution called Scheduled Sampling for Transformers is adopted in the process of training.Compared with the previous work based on LSTM network and GRU series networks,Transformer network built in this paper improves the estimation precision of vehicle-body vibration acceleration on the domestic high-speed railway measured dataset,and the mean absolute error of vertical and horizontal vehicle-body vibration acceleration decreases by 7.21% and 2.92%,respectively.The correlation coefficient ρ increased from 0.791 to0.820 and 0.733 to 0.746.