基于车辆响应的桥梁区段轨道高低不平顺快速估计算法  

Fast Estimation Algorithm for Track Irregularities in Bridge Sections Based on Vehicle Response

作　　者：支洋 ZHI Yang(Department of Track,Communication&Signaling and Power Supply,CHINA RAILWAY,Beijing 100844,China)

机构地区：[1]中国国家铁路集团有限公司工电部,北京100844

出　　处：《铁道建筑》2024年第10期20-25,共6页Railway Engineering

基　　金：中国国家铁路集团有限公司科技研究开发计划(N2023G014)。

摘　　要：基于高速铁路轨检数据,采用Pearson相关系数及相干函数,研究桥梁区段轨道高低不平顺与车辆响应的相关关系。利用贝叶斯优化(Bayesian Optimization,BO)对时间卷积神经(Temporal Convolutional Neural,TCN)网络进行改进,确定最佳感受野大小,提出基于贝叶斯优化的时间卷积神经(BO-TCN)网络算法,利用该算法对轨道高低不平顺进行估计,并与传统循环神经网络算法的准确率及计算效率进行对比。结果表明:车体垂向加速度能够反映桥梁跨长及其2~4倍频(32、16、10、8 m)和轨道板长及其2倍频(6.45、3.30 m)引起的周期性轨道高低不平顺;以车体垂向加速度作为BO-TCN的输入特征,可实现3 m以上波长轨道高低不平顺的准确估计;相比长短期记忆(Long Short Term Memory,LSTM)网络和门控循环单元(Gated Recurrent Unit,GRU),利用BO-TCN算法估计的轨道高低不平顺与实测值吻合度更高,且训练速度可达LSTM、GRU的20倍以上。Based on high speed railway track inspection data,Pearson correlation coefficient and coherence function were used to study the correlation between track irregularities and vehicle response in bridge sections.Using Bayesian Optimization(BO)to improve the temporal convolutional neural(TCN)network,determine the optimal receptive field size,and propose the temporal convolutional neural network algorithm based on Bayesian Optimization(BO-TCN).This algorithm was used to estimate longitudinal level irregularities,and compared with the accuracy and computational efficiency of traditional recurrent neural network algorithms.The results indicate that the vertical acceleration of the vehicle body can reflect the periodic track irregularities caused by the bridge span length and its 2~4 time frequency(32,16,10,8 m),as well as the track slab length and its 2 time frequency(6.45,3.30 m).By using the vertical acceleration of the vehicle body as the input feature of BO-TCN,the accurate estimation of the irregularities of tracks with wavelengths of over 3 m can be achieved.Compared to Long Short Term Memory(LSTM)networks and Gated Recurrent Units(GRU),BO-TCN could estimate track irregularities with higher accuracy than measured values,and the training speed can reach more than 20 times that of LSTM and GRU.

关 键 词：高速铁路 贝叶斯优化 时间卷积神经网络 车辆响应 轨道不平顺 车体垂向加速度 

分 类 号：U213.6[交通运输工程—道路与铁道工程]