摘要
为了确保路基系统的安全性并获得可靠的结构响应设计,准确且高效地估算路基回弹模量至关重要。以路基回弹模量为研究目标,利用包含12种不同路基的样本数据,共计2813组实际回弹模量数据,运用机器学习算法进行了预测。首先,综合考虑了力学、土性、环境等多个参数,选取了加权塑性指数、干密度、剪应力、偏应力、含水量和冻融循环次数6个影响因素作为输入参数。随后,采用熵权法和数据归一化技术对数据集进行了预处理,以获取各个影响因素的权重系数,更准确地反映了对回弹模量的影响。在此基础上,采用了XGBoost算法对路基回弹模量进行预测,同时考虑到XGBoost算法超参数较多,选取较为困难,结合了贝叶斯优化(BO)算法对超参数进行调整。最后,进行了敏感性分析,调查了每个输入变量的相对重要性。结果表明:相比基于网格搜索的XGBoost,BO-XGBoost模型能够以较短的耗时准确预测路基回弹模量,其测试集的相关系数R2为0.9966,RMSE为1.53;相比基于遗传算法优化的人工神经网络模型、支持向量机模型及传统的LGP和Kim经验模型,BO-XGBoost模型均展现出了更优越的性能;输入参数敏感性分析结果表明,不同的输入参数对路基回弹模量的影响程度不同,其中干密度是影响XGBoost算法预测性能最显著的因素。
In order to ensure the safety of subgrade system and obtain reliable structural response design,accurate and efficient estimation of subgrade resilient modulus is important.Taking subgrade resilient modulus as the study object,using the sample data of 12 different subgrades,a total of 2813 groups of actual resilient modulus data are predicted by using machine learning algorithm.First,many parameters such as mechanics,soil properties and environment are considered,and 6 influencing factors are selected as input parameters,such as weighted plastic indicator,dry density,shear stress,deviating stress,water content and freezing-thawing cycles.Then,entropy weight method and data normalization technique are used to preprocess the data set to obtain the weight coefficients of each influencing factor and reflect the influence on the resilient modulus more accurately.On this basis,the resilient modulus of subgrade is predicted by using XGBoost algorithm,and considering that the XGBooST algorithm has too many super-parameters to select,the Bayesian optimization(BO)algorithm is combined to adjust the super-parameters of the algorithm.Finally,a sensitivity analysis is performed to investigate the relative importance of each input variable.The result shows that(1)compared with XGBoost,the BO-XGBoost model can accurately predict subgrade resilient modulus with shorter time-consuming,and its correlation coefficient R2 is 0.9966,RMSE is 1.53;(2)the BOXGBoost model has better performance than the artificial neural network model,the support vector machine model and the traditional LGP and Kim empirical models;(3)the sensitivity analysis result shows that different input parameters have different effects on the resilient modulus of subgrade,and the dry density is the most significant factor affecting the prediction performance of XGBoost algorithm.
作者
徐明
康雅晶
马斯斯
张鹤
XU Ming;KANG Ya-jing;MA Si-si;ZHANG He(Beijing Shougang Construction Investment Co.,Ltd.,Beijing 100865,China;Beijing Jian Zhuang Consulting Co.,Ltd.,Beijing 100161,China;State Grid Beijing Electric Power Company,Beijing 100051,China;Beijing Construction Engineering Quality Third Test Institute Co.,Ltd.,Beijing 100037,China)
出处
《公路交通科技》
CAS
CSCD
北大核心
2023年第11期51-60,共10页
Journal of Highway and Transportation Research and Development
关键词
道路工程
路基性能
XGBoost
路基回弹模量
贝叶斯优化
机器学习
road engineering
subgrade performance
XGBoost
subgrade resilient modulus
Bayesian optimization
machine learning
作者简介
徐明(1977-),男,吉林农安人,硕士,高级工程师.(33625227@qq.com);通讯作者:张鹤(1983-),男,河北辛集人,硕士,高级工程师.(44226929@qq.com)。
