Deep Learning for Track Quality Evaluation of High-Speed Railway Based on Vehicle-Body Vibration Prediction

Shuai Ma,Liang Gao,Xiubo Liu,Jing Lin

doi:10.1109/access.2019.2960537

Abstract

Track quality evaluation is fundamental for track maintenance. Around the world, track geometry standards are established to evaluate track quality. However, these standards may not be capable of detecting some abnormal track geometry conditions that can cause considerable vehicle-body vibration. And people gradually realized that track quality evaluation should be based not only on track geometry but also on vehicle performance. Vehicle-body vibration prediction is beneficial for locating potential track geometry defects, and the predicted accelerations can be used as an auxiliary index for assessing track quality. For this purpose, this paper gives a method to predict vehicle-body vibration based on deep learning, which represents one of the newest areas in artificial intelligence. By integrating convolutional neural network (CNN) and long short-term memory (LSTM), a CNN-LSTM model is proposed to make accurate and point-wise prediction. To achieve the optimal performance and explore the internal mechanism of the model, structural configurations and inner states are extensively studied. CNN-LSTM can take advantage of the powerful feature extraction capacity of CNN and LSTM, and outperforms the fully-connected neural network and the plain LSTM on the experimental data of a high-speed railway. In detail, CNN-LSTM has superior performance in predicting vertical vehicle-body vibration below 10 Hz and lateral vehicle-body vibration below 1 Hz. Moreover, analysis shows that the predicted vehicle-body acceleration can act as a performance-based evaluation index of track quality.

Highlights

With the development of high-speed railway (HSR), people are paying more attention to the safety and ride comfort
For the purpose of promoting track quality evaluation, a convolutional neural network (CNN)-long short-term memory (LSTM) model is proposed for the point-wise prediction of HSR vehicle-body vibration by using track geometry
Analysis shows that CNN can learn shape features contained in track geometry waveform, and LSTM is capable of learning the sequential information of vehicle-body acceleration

Summary

INTRODUCTION

With the development of high-speed railway (HSR), people are paying more attention to the safety and ride comfort. Signal processing-based models share the merits of calculation efficiency and conceptual clarity, but they are inherently linear models and only applicable to constant speed conditions The latter takes advantage of intelligent models, such as artificial neural network and support vector machine, to recognize complex patterns and nonlinear relationships between track geometry and vehicle response [4], [9]. Convolutional neural network (CNN) and long short-term memory (LSTM) are classic deep learning models designed for image and sequential data respectively, and have already been adopted in safety surveillance [14], disease prediction [15], human kinematics interpretation [16], health condition monitoring [17], and vehicle fault diagnosis [18], [19] These models tend to demonstrate better performance than traditional machine learning models, for example, some deep convolutional nets have even surpassed human-level performance in visual recognition tasks [20].

LSTM FOR SEQUENTIAL FEATURE LEARNING

CONCLUSION