欢迎访问《硅酸盐通报》官方网站,今天是

硅酸盐通报 ›› 2026, Vol. 45 ›› Issue (9): 3070-3082.DOI: 10.16552/j.cnki.issn1001-1625.2026.0229

• 水泥混凝土 • 上一篇    下一篇

基于机器学习的混凝土抗压强度全龄期预测与解释性研究

李硕(), 秦艳慧(), 李蒙恩, 马浩原, 尹金帅, 郑璐璐, 邓苏文   

  1. 青岛理工大学土木工程学院,青岛 266520
  • 收稿日期:2026-03-16 修订日期:2026-04-08 出版日期:2026-09-15 发布日期:2026-10-09
  • 通信作者: 秦艳慧,博士,副教授。E-mail:yanhuiqin1919@126.com
  • 作者简介:李 硕(2001—),男,硕士研究生。主要从事混凝土耐久性方面的研究。E-mail:2806438891@qq.com
  • 基金资助:
    山东省自然科学基金资助项目(ZR2023MD036);山东省高等学校青年教师创新团队(2022KJ310)

Full-Age Prediction and Interpretability Study of Concrete Compressive Strength Based on Machine Learning

LI Shuo(), QIN Yanhui(), LI Meng’en, MA Haoyuan, YIN Jinshuai, ZHENG Lulu, DENG Suwen   

  1. School of Civil Engineering,Qingdao University of Technology,Qingdao 266520,China
  • Received:2026-03-16 Revised:2026-04-08 Published:2026-09-15 Online:2026-10-09

摘要:

针对传统混凝土抗压强度测定周期长、成本高、难以实现全过程监控的问题,本研究基于1 250组试验数据,采用随机森林(random forest)、XGBoost(eXtreme Gradient Boosting)、LightGBM(light gradient boosting machine)和CatBoost(categorical boosting)四种算法,构建混凝土抗压强度全龄期预测模型,结合SHAP(SHapley Additive exPlanations)可解释性分析,揭示强度影响机制。结果表明:CatBoost模型综合性能最优,全样本决定系数R2达0.96,均方根误差(RMSE)为3.29 MPa,平均绝对误差(MAE)为2.25 MPa,且在低龄期(≤7 d)、中龄期(14~28 d)、高龄期(>28 d)全阶段保持稳定预测能力;XGBoost在无粉煤灰体系中表现最佳(R2=0.98),CatBoost在无高炉矿渣体系中适应性最强(R2=0.95)。SHAP分析验证了水胶比定律与火山灰效应的核心作用,明确龄期、水泥用量、用水量为关键因素,用水量170~180 kg/m3为强度拐点。研究可为混凝土配比优化、现场快速质量评估提供科学依据,具有重要工程应用价值。

关键词: 混凝土抗压强度, 机器学习模型, CatBoost, SHAP解释, 龄期划分

Abstract:

To address the problems of a long testing period, high cost, and difficulty in achieving full-process monitoring in the traditional determination of concrete compressive strength, this study developed a full-age prediction model for concrete compressive strength based on 1 250 sets of test data using four learning algorithms, namely random forest (RF), eXtreme Gradient Boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). Combined with SHapley Additive exPlanations (SHAP) interpretability analysis, the influencing mechanism of concrete strength was revealed. The results show that the CatBoost model exhibits the optimal comprehensive performance, with a coefficient of determination R2 of 0.96, a root mean square error (RMSE) of 3.29 MPa, and a mean absolute error (MAE) of 2.25 MPa for the full dataset. It also maintains a stable predictive capability across all stages including the early age (≤7 d), medium age (14 d to 28 d), and late age (>28 d). XGBoost achieves the best performance in the fly ash-free system (R2=0.98), while CatBoost has the strongest adaptability in the ground granulated blast furnace slag-free system (R2=0.95). SHAP analysis verifies the core effects of the water-binder ratio law and pozzolanic effect, identifies age, cement content, and water content as the key influencing factors, with the water content of 170 kg/m3 to 180 kg/m3 being the strength inflection point. This study can provide a scientific basis for the optimization of concrete mix proportion and on-site rapid quality assessment, and has important engineering application value.

Key words: concrete compressive strength, machine learning model, CatBoost, SHAP interpretation, age stage division

中图分类号: