Soil Undrained Shear Strength Prediction Based On XGBoost And LightGBM Model

In recent years,the field of Artificial Intelligent(AI)continues to heat up rapidly,and Artificial intelligence technologies such as machine learning have been rapidly developed and widely applied in various fields,such as image processing and speech recognition.With the rapid development of population and urbanization,the construction of huge geotechnical engineering is in full swing.In the geotechnical engineering site,geotechnical engineering survey work at the same time accumulated a large number of surveys and test data.Among them,the undrained shear strength is an important index of strength characteristic analysis and stability evaluation of viscous soil,which has an important influence on the design of highway and bridge foundation.The traditional method takes the data of a specific place and only considers the limited variables,and the empirical formula is established by using empirical relation.In this paper,based on the International Society for Soil Mechanics and Geotechnical Engineering(ISSMGE)'s Engineering Practice of Risk assessment and Management Committee(TC304)of CLAY/6/535 of data sets,creatively apply e Xtreme Gradient Boosting(XGBoost)and LightGradient Boosting Machine(LightGBM)ensemble learning algorithms to construct the undrained shear strength of cohesive soils model and analysis and research.The research focuses on the statistical analysis and preprocessing of data sets,the research,and implementation of the ensemble learning algorithm and model fusion.This study is an attempt to apply the "New Infrastructure Construction" centered on big data and artificial intelligence in the field of traditional geotechnical engineering and geotechnical mechanics.The main conclusions are as follows:In terms of statistical description and preprocessing of data,the box plot method and isolated forest algorithm were used to eliminate the outliers in 6 groups of outlier data and feature variables.At the same time,the missing values left after the outliers in the feature variables are removed are filled by the Miss Forest algorithm.By comparing the modeling results before and after data preprocessing,it is found that the removal of outliers can obviously improve the accuracy and stability of the models.In addition,it was found that the XGBoost and LightGBM models were barely affected before and after the missing value filling,indicating that the accuracy and stability of the two models could be guaranteed under the condition of a small number of missing values.In this paper,a Bayesian optimization method is proposed to adjust the hyperparameters of the two ensemble algorithms.Through the understanding and introduction of the hyperparameters of the two ensemble algorithms,each algorithm determines 7 hyperparameters and their corresponding ranges for the hyperparameter tuning of the Bayesian optimization algorithm,and finally determines the optimal hyperparameter combination of the two models after 500 iterations.By comparing the modeling results of the two models before and after hyperparameter adjustment,it is found that the prediction performance and model stability of the model established after hyper parameter adjustment are greatly improved in each evaluation index.Five XGBoost and LightGBM models were built under the 5-fold cross-validation to show their prediction performance and stability.At the same time,they were compared with the comparison algorithm and empirical formula and found that the proposed two ensemble learning models have obvious advantages.The average fusion of XGBoost and LightGBM models into a fusion model shows that the fusion model well integrates the predictive performance advantages of the two models.Finally,the relative importance order of the features is given to increase the interpretability of the model,which has important guiding significance for exploring the characteristics of undrained shear strength.