基于机器学习的土壤不排水抗剪强度模型研究

Research on Machine Learning Based Soil Undrained Shear Strength Model

土壤抗剪强度在岩土工程中扮演着重要的角色,通常需要结合原位试验和实验室试验来进行测定。为了减少大型工程项目对繁琐且昂贵的实验室试验的依赖,探究了5种机器学习模型(人工神经网络、支持向量机、高斯过程回归、随机森林、XGBoost),利用静力触探试验数据来预测土壤不排水抗剪强度,以评估机器学习技术对减少实验室试验需求的潜力。结果表明,在5种模型中,XGBoost(R^(2)=0.76,MAE=10.79,RMSE=15.45)与ANN(R^(2)=0.76,MAE=10.87,RMSE=15.49)表现最出色,具有预测土壤不排水抗剪强度的潜力。然而,模型在组10折交叉验证中表现不佳,这表明尽管模型在训练时能够很好地处理特定场地数据,但在没有特定场地数据的情况下,泛化能力受到限制,凸显了准确描述土壤异质性的挑战。

In geotechnical engineering,soil shear strength plays a crucial role and is typically determined through a combination of in-situ tests and laboratory tests.To reduce the reliance of large-scale engineering projects on cumbersome and costly laboratory tests,this study investigates five machine learning models(namely the artificial neural network,support vector machine,Gaussian process regression,random forest and XGBoost)that utilize cone penetration test data to predict undrained soil shear strength.The results indicate that among the five models tested,XGBoost(R^(2)=0.76,MAE=10.79,and RMSE=15.45)and ANN(R^(2)=0.76,MAE=10.87,and RMSE=15.49)exhibit the most promising predictive capabilities for undrained soil shear strength.However,the models performed poorly in 10-fold cross-validation,suggesting that while they excel in handling site-specific data during training,their generalization capacity is limited in the absence of site-specific data.This underscores the challenge of accurately characterizing soil heterogeneity.