Small-Amplitude Hunting Interval Prediction For High-Speed Trains Based On Data Imbalance

The hunting motion generated by the high-speed train can seriously affect the safety.The current research on hunting motion is mainly for the classification and identification of hunting instability,but there is a small-amplitude hunting state from normal to hunting instability during the train operation,and the prediction of small-amplitude hunting state can provide early warning of hunting instability.Therefore,the research in this thesis is focused on the prediction of small-amplitude hunting of high-speed trains.Since the trains are in normal driving condition for a long time,it is difficult to collect the fault samples of hunting motion,which makes the existing data have the characteristics of imbalance and leads to the problem of model overfitting and lack of generalization when using machine learning methods for prediction,so this thesis will first carry out the research of data imbalance.On the other hand,the prediction of high-speed train time series is a nonlinear prediction problem,and the traditional point prediction method is highly random and does not have stability.Therefore,this thesis will carry out the study of interval prediction of small-amplitude hunting.For the above problems,the main research contents of this thesis are as follows.1.Based on the data imbalance problem,this thesis improves the original encoder structure on the basis of Conditional Variational Autoencoder-Generative Adversarial Network(CVAE-GAN,Conditional Variational Autoencoder-Generative Adversarial Network),and puts the self-attention encoding layer as the first input layer of temporal samples,a new deep learning fault generation model(CTrans VAE-GAN,Conditional Transformer Variational Autoencoder-Generative Adversarial Network)is constructed,and the addition of this self-attention layer makes The addition of this self-attention layer allows the original model to learn the before-and-after dependence of the temporal input,and effectively enhances the generation quality of the model by first building the self-attention matrix and then performing the feature extraction of the intermediate encoding.After the experimental generation comparison,it is found that the proposed method outperforms the existing deep learning fault generation model in terms of maximum mean difference and sample kurtosis value index.The original high-speed train hunting monitoring model is insufficiently trained due to data imbalance.In this thesis,the well-trained CTrans VAE-GAN model is used to enhance the original data set,and the enhanced data is used for classification.Experiments show that the classification accuracy using the proposed CTrans VAE-GAN model after data augmentation reaches 97.5%,which is better than other comparative methods.2.Based on the interval prediction problem,this thesis improves the model structure and optimization method of the upper and lower bound interval estimation(LUBE)model,proposes an interval prediction model structure based on full self-attention layer,and designs a constructive interval optimization method combining prediction evaluation indexes on this basis.After experiments on public datasets and high-speed train online monitoring data,the combined performance of the method in terms of interval evaluation indexes and training time is proved to be better than other methods.In this thesis,the Transformer model is applied to signal feature extraction with time series,based on the attention mechanism to provide before and after time information for any time point in the input time series,and to obtain the association of each hidden layer with the contextual hidden layer,so as to achieve automatic feature extraction of time series signals.The model allows more parallelization of the input samples compared to recurrent neural networks,reducing the training time while gaining performance improvements.The designed optimization method for constructing intervals can dynamically adjust the width of the constructed intervals by the changes of the statistical interval coverage(PICP,PI coverage probability)and the interval coverage width(PINAW,PI normalized averaged width)after each iteration,so as to establish the corresponding upper and lower prediction limits label,which can be updated by calculating the minimum mean squared error(MSE,Mean Squared Error)loss function and then the parameter update through the back-propagation method of gradient descent.Finally,the experimental comparison verifies the improvement of the self-attention interval prediction model based on self-attention prediction relative to the traditional method.3.This thesis combines the proposed classification and prediction methods to construct a new hunting motion monitoring scheme and develop a monitoring software based on Py Qt5.After simulation experiments,the software can first complete real-time classification of highspeed trains’ operation status and predict the identified small-amplitude hunting motion status on this basis,which has some practical application significance.