基于集成学习模型的正常固结土抗剪强度指标预测方法  

A data-driven model for predicting shear strength indexes of normally consolidated soils

作　　者：王钰轲 冯爽 钟燕辉[1,2,3] 张蓓 WANG Yuke;FENG Shuang;ZHONG Yanhui;ZHANG Bei(School of Water Conservancy and Transportation,Zhengzhou University,Zhengzhou 450001,China;National Local Joint Engineering Laboratory of Major Infrastructure Testing and Rehabilitation Technology,Zhengzhou 450001,China;Provincial and Ministerial Collaborative Innovation Center for Underground Engineering Disaster Prevention and Control,Zhengzhou 450001,China)

机构地区：[1]郑州大学水利与交通学院,河南郑州450001 [2]重大基础设施检测修复技术国家地方联合工程实验室,河南郑州450001 [3]地下工程灾变防控省部共建协同创新中心,河南郑州450001

出　　处：《岩土工程学报》2023年第S02期183-188,共6页Chinese Journal of Geotechnical Engineering

基　　金：国家自然科学基金项目(52109140);河南省优秀青年基金项目(232300421069);中原科技创新领军人才资助项目(234200510014);河南省交通运输厅科技项目(2022-5-5)。

摘　　要：抗剪强度指标的准确获取对工程设计具有决定性作用,目前试验方法确定的强度指标过于依靠工程经验,导致最终的取值具有不确定性。集成学习模型是机器学习的一个子类,在处理复杂的数据和任务时表现出强大的性能。为了能更准确地获得抗剪强度指标,以正常固结土为研究对象,利用不同的集成学习算法建立其抗剪强度指标的预测模型。通过均方根误差(RMSE)、可决系数(R^(2))、绝对值误差(MAE)评估不同模型的泛化能力,并采用Adaboost算法进行输入参数的敏感性分析。结果表明抗剪强度采用Adaboost、内摩擦角采用RF、黏聚力采用Adaboost算法具有最佳的泛化能力,其测试集R^(2)可分别达到0.925,0.965,0.942。敏感性分析结果显示,对抗剪强度影响最大的参数为干密度、含水率和法向应力;对内摩擦角影响最大的参数为曲率系数、黏粒含量和含水率;对黏聚力影响最大的参数为含水率、干密度和液限。本文所建立的预测模型可为工程中抗剪强度指标的选取及运用机器学习方法研究土体强度参数提供参考。The accurate determination of shear strength indexes is crucial for engineering soils.Currently,the strength indexes determined by the test method depend too much on engineering experience,which leads to the uncertainty of the final value.The ensemble learning is a subclass of machine learning that exhibits strong performance when dealing with complex data and tasks.To enhance the precision of the shear strength indexes,based on the normally consolidated soils,a model is established using the diverse ensemble learning algorithms for predicting the shear strength indexes of soils.Various models are assessed for their generalization capability using the root mean square error(RMSE),coefficient of determination(R^(2))and absolute value error(MAE).The Adaboost algorithm is employed for the sensitivity analysis of input parameters.The findings indicate that the Adaboost algorithm yields the best generalization for the shear strength the RF for the internal friction angle,and the Adaboost algorithm for the cohesion,achieving respective test-set R²values of 0.925,0.965 and 0.942.The sensitivity analyses reveal that the dry density,moisture content and normal stress exert the most significant influence on the shear strength,while the key factors for the internal friction angle are the coefficient of curvature,viscous grain content and water content.The water content,dry density and liquid limit are identified as the primary influencers on the cohesion.The data-driven model established herein offers guidance for selecting the shear strength indexes in engineering and investigating strength parameters of soils through the machine learning methods.

关 键 词：正常固结土 抗剪强度指标 集成学习 机器学习 敏感性分析 

分 类 号：TU43[建筑科学—岩土工程]