基于卷积神经网络的城市轨道交通轨道几何不平顺指标预测  

Prediction of Urban Rail Transit Track Geometry Irregularity Indicators Based on Convolutional Neural Networks

作　　者：张梓鸿 赵正阳 王文斌 李洋 李沅军 ZHANG Zihong;ZHAO Zhengyang;WANG Wenbin;LI Yang;LI Yuanjun(Urban Rail Transit Center,China Academy of Railway Sciences Group Co.,Ltd.,100081,Beijing,China;不详)

机构地区：[1]中国铁道科学研究院集团有限公司城市轨道交通中心,北京100081 [2]石家庄铁道大学交通运输学院,石家庄050043

出　　处：《城市轨道交通研究》2023年第10期109-115,共7页Urban Mass Transit

基　　金：北京市自然科学基金-丰台轨道交通前沿研究联合基金项目(L221001);中国铁道科学研究院集团有限公司院基金重点课题(2021YJ188)。

摘　　要：目的:城市轨道交通轨道几何是评价轨道平顺性的重要指标。鉴于城市轨道交通轨道几何不平顺指标预测中,传统方法仅对部分因素进行了量化,且预测精度较低;而神经网络法较为简单,不能较好地拟合输入、输出数据,且主要通过时间序列对数据进行预测,预测精度有限。提出采用卷积神经网络对轨道几何不平顺指标进行预测。方法:从数据选取、数据集构建、模型训练及模型评估等方面介绍了轨道几何不平顺预测的整体流程。分别建立左高低、右高低、左轨向、右轨向、轨距、水平和三角坑等7项指标的基于卷积神经网络的轨道几何不平顺指标预测模型,从检测项、时间及里程等3个维度对该模型训练的数据集进行整合处理。介绍了卷积神经网络的内部结构,从学习速率、激活函数、损失函数、Dropout方法等角度对模型进行了优化。选取某城市轨道交通线路的检测数据作为评估数据集,通过均方误差对轨道几何不平顺7项指标的预测结果进行了评估。采用单隐层BP神经网络、双隐层BP神经网络、多元回归分析及卷积神经网络等方法对三角坑轨道几何不平顺的均方误差进行对比。结果及结论:三角坑的均方误差为0.0184,其他指标的均方误差均在0.02左右;相比单隐层BP神经网络、双隐层BP神经网络和多元回归分析,卷积神经网络法的均方误差分别降低了73.29%、71.80%和664.81%,说明卷积神经网络具有更好的拟合能力。基于卷积神经网络的城市轨道交通轨道几何不平顺预测模型能较好的预测轨道几何指标的变化趋势。Objective:Urban rail transit track geometry is a significant indicator for evaluating track smoothness.Considering the limitations of conventional methods in predicting urban rail transit track geometry irregularity indicators,which only quantify certain factors and yield low prediction accuracy,and the shortcomings of neural network methods in fitting input-output data effectively and limited accuracy due to the primarily predicting data through time series,the use of CNN(convolutional neural networks)is proposed for predicting track geometry irregularity indicators.Method:The overall process of predicting track geometry irregularities is presented,covering aspects such as data selection,dataset construction,model training,and model evaluation.Separate prediction models based on CNN are established for seven indicators:left elevation,right elevation,left alignment,right alignment,gauge,level,and turnout pit.The dataset for model training is integrated and processed from three dimensions:detection items,time,and mileage.The internal structure of CNN is outlined,and optimization of prediction models is performed from perspectives including leaning rate,activation function,loss function,and the Dropout method.Detection data from an urban rail transit line are employed as assessment dataset,and the prediction results of the seven track geometry irregularity indicators are assessed using MSE(mean squared error).The MSE of triangular pit track irregularities are compared between single-hidden-layer BP(backpropagation)neural networks,double-hidden-layer BP neural networks,multiple linear regression analysis,and CNN.Result&Conclusion:The MSE of turnout pit irregularities is 0.0184,and the MSE of other indicators are around 0.02.Compared to the previous three,the MSE of CNN decrease by 73.29%,71.89%,and 664.81%,respectively,indicating superior fitting capability of CNN.The prediction models for urban rail transit track geometry irregularities based on CNN can effectively forecast the changing trends of track geometry ind

关 键 词：城市轨道交通 轨道几何不平顺指标 卷积神经网络 

分 类 号：U213.213[交通运输工程—道路与铁道工程]