融合XGBoost和SHAP的混凝土抗压强度预测分析模型

Prediction and analysis of concrete compressive strength based on XGBoost and SHAP

摘要

【目的】为精准预测混凝土抗压强度、突出XGBoost模型的预测优势和实现XGBoost模型的可解释功能,【方法】构建以水泥、龄期、水等8种影响因素为输入特征和以抗压强度为目标特征的1 030个样本数据集,建立支持向量回归(SVR)、随机森林(RF)和极端梯度提升树(XGBoost)机器学习算法模型,开展混凝土抗压强度预测研究,将XGBoost模型和ACI209公式的预测结果对比,同时,引入SHAP模型对XGBoost模型进行解释和分析。【结果】结果表明:XGBoost模型的预测精度最高,R²为0.952, MAE为2.48, MAPE为9.16, RMSE为3.58,但XGBoost模型对小于30 MPa的低抗压强度样本的预测误差较大,随着抗压强度的增大,XGBoost模型的预测精度提高,超限样本比例从25%下降到2.7%;与ACI209公式的预测结果相比,XGBoost模型在龄期56 d和100 d样本的预测值绝对误差率均值为4.10%,3.64%,而ACI209公式则为11.27%,17.96%。【结论】XGBoost模型适用于混凝土强度大于30 MPa的样本的预测;SHAP模型不仅可以定量地给出特征重要性排序,还能定性地给出每个特征参数是对抗压强度的影响规律,能为混凝土相关研究及其他需要对机器学习模型进行解释的研究提供参考。

Abstract

[Objective] To accurately predict the compressive strength of concrete, highlight the predictive advantages of the XGBoost model, and realize the interpretable function of the XGBoost model, [Methods] a data set of 1030 samples with eight factors such as cement, age, water and others as input features and compressive strength as target features is constructed, and machine learning algorithm models of Support Vector Regression(SVR), Random Forest(RF) and Extreme Gradient Boosting Tree(XGBoost) to research on concrete compressive strength prediction, comparing the prediction result of the XGBoost model and the ACI209 formula, and meanwhile, introducing the SHAP model to explain and analyze the XGBoost model. [Results] The result show that the XGBoost model has the highest prediction accuracy with R² of 0. 952, MAE of 2. 48, MAPE of 9. 16, and RMSE of 3. 58; however, the prediction error of the XGBoost model for low compressive strength samples less than 30 MPa is larger, and the prediction accuracy of the XGBoost model improves as the compressive strength increases, and the proportion of exceeding the limit samples decreases from 25% to 2. 7%; compared with the prediction result of ACI209 formula, the mean absolute error rate of the XGBoost model's prediction values for samples of age 56 d and 100 d are 4. 10% and 3. 64%, compared with 11. 27% and 17. 96% for ACI209 formula. [Conclusion] The XGBoost model is suitable for the prediction of samples with concrete strength greater than 30 MPa; The SHAP model can not only quantitatively give the ranking of feature importance, but also qualitatively give the influence of each feature parameter on compressive strength, which can provide a reference for concreterelated research and other studies that need to explain machine learning models.

关键词

机器学习 / XGBoost / SHAP / 抗压强度预测 / 混凝土 / 力学性能

Key words

machine learning / XGBoost / SHAP / compressive strength prediction / concrete / mechanical properties

基金资助

宁夏回族自治区重点研发计划项目(2022BEG02056)