基于车辆响应的高速铁路轨道几何状态评估方法  

Track Geometry State Assessment Method for High-speed Railway Based on Vehicle Response

作　　者：马帅 刘秀波 张彦博 陈茁 赵东全 MA Shuai;LIU Xiubo;ZHANG Yanbo;CHEN Zhuo;ZHAO Dongquan(Infrastructure Inspection Research Institute,China Academy of Railway Sciences Co.,Ltd.,Beijing 100081,China)

机构地区：[1]中国铁道科学研究院集团有限公司基础设施检测研究所,北京100081

出　　处：《铁道学报》2023年第11期117-127,共11页Journal of the China Railway Society

基　　金：国家重点研发计划(2018YFE0207100);中国国家铁路集团有限公司科技研究开发计划(K2022T002)。

摘　　要：轨道几何状态科学评估对保障高速铁路列车平稳、安全运行具有重要意义。基于高速综合检测列车多次检测数据,利用卷积神经网络、注意力模块和长短时记忆网络,分别学习数据的波形特征、注意力权值、长距离空间依赖关系特征,建立CBAM-CNN-LSTM车辆动态响应预测模型。该模型通过输入轨道几何、运行速度和车型预测不同工况下的车辆动态响应,进而利用预测的车辆动态响应评价轨道几何状态。研究结果表明,建立的模型能够有效预测车体振动响应,根据我国某高速铁路两种车型综合检测列车检测数据的验证结果,车体横向、垂向加速度的均方根预测误差分别为0.004 g、0.009g,相关系数分别为0.608、0.793;利用预测的车辆动态响应评估轨道状态,能够有效识别引起车体振动加剧的轨道几何不利状态或隐形病害。此外,模型内部的注意力权值有助于分析挖掘导致轨道状态不良的轨道几何参数类型和位置信息。The scientific assessment of track geometry state is of great significance in ensuring the stable and safe operation of high-speed trains.Based on the track inspection data of comprehensive inspection trains from several inspection runs,a CBAM-CNN-LSTM vehicle dynamic response prediction model was established using the convolutional neural network,convolutional attention module and long short-term memory to learn the waveform characteristics,attention weights,long-span spatial dependency of the data respectively.By inputting track geometries,operation speed and vehicle type,the model can automatically predict vehicle-body accelerations under different conditions.The vehicle-body acceleration predictions can then be used to assess track geometry state.The results show that the proposed model can effectively predict vehicle-body vibrational responses.According to the validation results on the inspection data from two types of comprehensive inspection trains of a high-speed railway in China,the root-mean-square-errors of the lateral and vertical vehicle-body acceleration predictions are 0.004 g and 0.009 g respectively,with respective correlation coefficients of 0.608 and 0.793.Adverse states of track geometry or hidden defects where abnormal vehicle dynamic response occurs can be effectively identified by using the predicted vehicle-body accelerations as track state assessment indices.In addition,the attention weights inside the model are helpful in analyzing and excavating the categorical and positional information of track geometry parameters that lead to poor track state.

关 键 词：轨道几何状态 车辆响应 评估方法 高速铁路 卷积神经网络 长短时记忆网络 注意力模块 

分 类 号：U216.3[交通运输工程—道路与铁道工程]