基于经验约束神经网络的高速铁路路基累积变形预测研究  

Research on subgrade cumulative deformation prediction of high-speed railway based on empiricism-constrained neural network

作　　者：邓志兴 徐林荣[1] 李永威 王武斌[2] 苏谦[3] DENG Zhixing;XU Linrong;LI Yongwei;WANG Wubin;SU Qian(School of Civil Engineering,Central South University,Changsha 410075,China;National Engineering Research Center of Geological Disaster Prevention Technology in Land Transportation,Southwest Jiaotong University,Chengdu 611731,China;School of Civil Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]西南交通大学陆地交通地质灾害防治技术国家工程研究中心,四川成都611731 [3]西南交通大学土木工程学院,四川成都610031

出　　处：《铁道科学与工程学报》2025年第2期469-484,共16页Journal of Railway Science and Engineering

基　　金：铁路基础研究联合基金资助项目(U2268213);中国中铁股份有限公司科技研究开发计划(2023-重点-09);国家自然科学基金资助项目(42172322);陆地交通地质灾害防治技术国家工程研究中心团队建设项目(A0920502052401-452)。

摘　　要：掌握高速铁路路基变形发展对变形病害的控制与运营管理意义重大。目前,机器学习方法被广泛应用于累积变形预测,而传统机器学习路基累积变形预测模型存在时间分布外泛化性差的弊端。因此,提出一种基于经验约束神经网络(ECNN)的高速铁路路基累积变形预测方法。首先,基于工程现场或室内试验数据构建路基累积变形预测数据集,并划分为训练集和测试集;其次,基于训练集建立神经网络模型,综合测试集上的预测精度与误差、预测不确定性2个层次结果,确定最优神经网络预测模型;最后,利用最优神经网络模型驱动路基累积变形的数据信息,并以损失函数修正的方式嵌入累积塑性应变关系曲线(经验信息),实现对最优神经网络模型参数和损失函数的约束,完成ECNN模型构建。案例分析表明:双向门控循环单元(Bi-GRU)模型为最优的神经网络预测模型,拟合优度R~2达到0.972 59,扩展不确定性U_(95)和标准化平均差f_(smd)仅为0.015 6、0.181 09;相较于Bi-GRU模型,ECNN模型在预测精度与误差、预测不确定性2个层次均更优,表明考虑经验信息约束的ECNN模型具备更强的预测性能;ECNN模型相较于Bi-GRU模型具有优异的时间分布外泛化性能,当训练集覆盖的时间跨度较小时,可有效提高累积变形的预测精度。研究成果可为高速铁路路基累积变形预测提供新参考。Understanding the deformation evolution of high-speed railway subgrade holds significant importance for controlling deformation issues and managing operations effectively.Currently,machine learning methods are widely used for cumulative deformation prediction,while the traditional machine learning prediction models for cumulative subgrade deformation have the disadvantage of poor out-of-distribution generalization.Hence,a method for predicting cumulative subgrade deformation of high-speed railways based on empiricism-constrained neural networks(ECNN)was proposed.First,the cumulative subgrade deformation prediction dataset was constructed based on engineering field or laboratory test data and divided into training and test sets.Second,a neural network model was built based on the training set,and the optimal neural network prediction model was determined by combining the results of two levels of prediction accuracy and error and prediction uncertainty on the test set.Finally,the optimal neural network model was used to drive the data information of the cumulative subgrade deformation and embed the cumulative plastic strain relationship curve(empirical information)with loss function correction to realize the constraints on the parameters of the optimal neural network model and the loss function.This process completed the model construction of ECNN.The case study shows that the Bidirectional Gate Recurrent Unit(Bi-GRU)model is the optimal neural network model for high-speed railways,with the goodness-of-fit R2 reaching 0.97259,and the extended uncertainty U95 and the standardized mean deviation fsmd being only 0.0156 and 0.18109,respectively.The ECNN model is better in both prediction accuracy and error,and prediction uncertainty levels are comparable with the Bi-GRU model,indicating that the ECNN model considering the constraints of empirical information has stronger prediction performance.The ECNN model has excellent out-of-time-distribution generalization performance as compared to the Bi-GRU model,which can effe

关 键 词：高速铁路 累积变形预测 经验约束神经网络 神经网络 累积塑性应变关系 

分 类 号：U213.1[交通运输工程—道路与铁道工程]