{"title":"Prediction method for the dynamic response of expressway lateritic soil subgrades on the basis of Bayesian optimization CatBoost","authors":"","doi":"10.1016/j.soildyn.2024.108943","DOIUrl":null,"url":null,"abstract":"<div><p>Due to the limited features and poor accuracy of current methods for predicting the dynamic response of subgrades, this paper proposes an innovative approach that combines subgrade dynamic response field tests and machine learning (ML) technology. This method uses Bayesian optimization XGBoost (BO-XGBoost), Bayesian optimization LightGBM (BO-LightGBM), and Bayesian optimization CatBoost (BO-CatBoost) models to analyze the effects of physical properties and stress conditions on the dynamic stress, dynamic acceleration, and dynamic displacement of the subgrade. The optimal ML model was selected on the basis of the residuals, coefficient of determination (<em>R</em><sup>2</sup>), mean squared error (MSE), and mean absolute error of the prediction results. Using SHapley additive exPlanations (SHAP), the global importance, feature importance, and feature interaction behaviours of the optimal ML model input features were explained, and the main controlling features affecting the dynamic stress, dynamic acceleration, and dynamic displacement of the subgrade were obtained. The research results indicate that the prediction results of the BO-XGBoost, BO-LightGBM, and BO-CatBoost models for dynamic stress, dynamic acceleration, and dynamic displacement are mostly within the 10 % error range, and the <em>R</em><sup>2</sup> values of these three models are greater than 0.98. On the basis of the comparison results of the hyperparameter combinations, the objective of MSE (MSE<sub>obj</sub>), and the error evaluation metrics, the BO-CatBoost model yields the highest prediction accuracy, making it the optimal ML prediction model. This prediction method can quickly and intelligently obtain the main controlling features of dynamic stress, dynamic acceleration, and dynamic displacement, including depth (<em>H</em>), axle load (<em>P</em>), frequency (<em>f</em>), and moisture content (<em>w</em>). The boundary conditions for these four features are as follows: <em>H</em> > −1.3 m, <em>P</em> > 10 ton, <em>f</em> > 3.7 Hz, and <em>w</em> >18.1 %. The research results contribute to enhancing the service performance and lifespan of expressways.</p></div>","PeriodicalId":49502,"journal":{"name":"Soil Dynamics and Earthquake Engineering","volume":null,"pages":null},"PeriodicalIF":4.2000,"publicationDate":"2024-09-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Soil Dynamics and Earthquake Engineering","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0267726124004950","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, GEOLOGICAL","Score":null,"Total":0}
引用次数: 0
Abstract
Due to the limited features and poor accuracy of current methods for predicting the dynamic response of subgrades, this paper proposes an innovative approach that combines subgrade dynamic response field tests and machine learning (ML) technology. This method uses Bayesian optimization XGBoost (BO-XGBoost), Bayesian optimization LightGBM (BO-LightGBM), and Bayesian optimization CatBoost (BO-CatBoost) models to analyze the effects of physical properties and stress conditions on the dynamic stress, dynamic acceleration, and dynamic displacement of the subgrade. The optimal ML model was selected on the basis of the residuals, coefficient of determination (R2), mean squared error (MSE), and mean absolute error of the prediction results. Using SHapley additive exPlanations (SHAP), the global importance, feature importance, and feature interaction behaviours of the optimal ML model input features were explained, and the main controlling features affecting the dynamic stress, dynamic acceleration, and dynamic displacement of the subgrade were obtained. The research results indicate that the prediction results of the BO-XGBoost, BO-LightGBM, and BO-CatBoost models for dynamic stress, dynamic acceleration, and dynamic displacement are mostly within the 10 % error range, and the R2 values of these three models are greater than 0.98. On the basis of the comparison results of the hyperparameter combinations, the objective of MSE (MSEobj), and the error evaluation metrics, the BO-CatBoost model yields the highest prediction accuracy, making it the optimal ML prediction model. This prediction method can quickly and intelligently obtain the main controlling features of dynamic stress, dynamic acceleration, and dynamic displacement, including depth (H), axle load (P), frequency (f), and moisture content (w). The boundary conditions for these four features are as follows: H > −1.3 m, P > 10 ton, f > 3.7 Hz, and w >18.1 %. The research results contribute to enhancing the service performance and lifespan of expressways.
期刊介绍:
The journal aims to encourage and enhance the role of mechanics and other disciplines as they relate to earthquake engineering by providing opportunities for the publication of the work of applied mathematicians, engineers and other applied scientists involved in solving problems closely related to the field of earthquake engineering and geotechnical earthquake engineering.
Emphasis is placed on new concepts and techniques, but case histories will also be published if they enhance the presentation and understanding of new technical concepts.