再生骨料混凝土氯离子扩散系数机器学习预测模型对比研究
Comparative study on machine learning prediction models for chloride diffusion coefficients of recycled aggregate concrete
关键词：
再生骨料混凝土；氯离子扩散系数；机器学习；BP神经网络；高斯过程回归；广义加性模型；支持向量机；随机森林
Key words：
Recycled aggregate concrete (RAC); Chloride diffusion coefficient; Machine learning; Back propagation (BP) neural network; Gaussian process regression; Generalized additive model; Support vector machine; Random forest
刊期：
2026年第9期
DOI：
10.19761/j.1000-4637.25070319
文章编号：
基金项目：
铁四院重大科研课题项目（KY2023047S）。
作者：
胡 健
单位：
中铁第四勘察设计院集团有限公司，湖北 武汉 430063
摘要
摘 要：再生骨料混凝土（Recycled aggregate concrete，RAC）渗透性能预测是评估其耐久性能、推广固废建材工程应用的关键。传统经验预测模型参数有限、非线性表征能力不足，难以适配复杂配合比与工程工况下RAC氯离子扩散特性的预测需求。为实现RAC氯离子扩散系数的高效预测，本文基于72组数据，筛选水泥用量、用水量、再生骨料品质、水胶比、再生骨料替代率和28 d抗压强度6项核心影响参数，构建预测数据集。分别建立反向传播（Backpropagation，BP）神经网络、高斯过程回归（Gaussian process regression，GPR）、广义加性模型（Generalized additive model，GAM）、支持向量机（Support vector machine，SVM）、随机森林回归（Random forest regression，RFR）五种模型，结合R2、MMAPE、MMAE、RRMSE四项评价指标，从训练集、验证集、测试集多维度系统对比各模型的预测精度、稳定性与泛化性能。结果表明：BP神经网络模型与GPR模型训练拟合效果优异，GAM泛化稳定性较差，RFR模型存在明显过拟合问题，SVM模型整体预测精度最低；BP神经网络模型综合性能最优，多数据集预测误差小、离散度低，五折交叉验证相关系数r均值达0.935 2，鲁棒性与泛化能力优异，可有效表征多因素耦合作用下氯离子的非线性扩散规律。研究成果可为RAC耐久性能精准预测与绿色建材工程应用提供技术参考。
Abstract： Prediction of permeability performance of recycled aggregate concrete (RAC) is critical for evaluating its durability and promoting the engineering application of solid waste building materials. Traditional empirical prediction models incorporate limited parameters and lack sufficient capacity to characterize nonlinear relationships, so they fail to predict the chloride diffusion characteristics of RAC under complex mix proportions and engineering service conditions. To realize accurate and efficient prediction of the chloride diffusion coefficient of RAC, a prediction dataset was established based on 72 groups of data, with six core influencing parameters screened out, namely cement dosage, water consumption, recycled aggregate quality, water-binder ratio, recycled coarse aggregate replacement rate and 28-day compressive strength. Five prediction models are constructed respectively, including Backpropagation (BP) neural network, Gaussian process regression (GPR), generalized additive model (GAM), support vector machine (SVM) and random forest regression (RFR). Four evaluation indicators, namely R2, MMAPE, MMAE and RRMSE, were adopted to systematically compare the prediction accuracy, stability and generalization performance of each model from the dimensions of training set, validation set and test set. The results reveal that BP neural network model and GPR model exhibit excellent fitting performance on the training set, GAM suffers poor generalization stability, RFR model has obvious overfitting defects, and SVM model delivers the lowest overall prediction accuracy. Among all algorithms, BP neural network achieves the optimal comprehensive performance with minor prediction errors and low dispersion across multiple datasets. The mean correlation coefficient r of 5-fold cross-validation reaches 0.935 2, demonstrating outstanding robustness and generalization capacity, which can accurately describe the nonlinear chloride diffusion law under the coupling effect of multiple factors. The research findings can provide technical references for the precise durability prediction of RAC and the engineering popularization of green building materials.
结论
（1）不同算法的预测表现存在显著差异：GPR模型与BP神经网络模型训练拟合能力优异，训练误差低；GAM具备一定非线性拟合能力，但泛化稳定性差，数据集间性能波动明显；RFR模型存在过拟合问题，训练集表现良好但测试集精度大幅衰减；SVM模型整体预测性能最差，难以捕捉多因素耦合下氯离子扩散的复杂非线性规律，难以适配本研究预测场景。
（2）BP神经网络模型在训练集、验证集、测试集中均保持最优的拟合精度与稳定性，预测误差集中且异常值少，可精准描述输入参数与氯离子扩散系数之间的非线性耦合关系。五折交叉验证结果进一步验证了该模型的鲁棒性，测试集相关系数平均值达0.935 2，标准差仅为0.042 2，离散程度低，泛化性能可靠，可优先应用于RAC渗透性能的预测工作。
（3）BP神经网络模型具备明确的工程适用场景，仅适用于水胶比0.4~0.6、再生骨料替代率0~100%、标准常温养护条件下RAC氯离子扩散系数的预测。模型存在一定应用局限性，未考虑冻融、硫酸盐侵蚀等极端服役环境的耦合作用，且仅针对氯离子扩散系数单一耐久性能指标开展预测，不适用于非常规配合比、多侵蚀环境及其他耐久性能预测场景。后续可通过扩充多工况、多参数样本数据集，优化模型结构与超参数体系，进一步提升模型的通用性与工程适配性。
参考文献
[1] 刘玉涛,高艺榕,马必聪.内养生混凝土力学行为及抗氯离子渗透性能研究[J].硅酸盐通报,2023,42(6):2027-2036.
[2] 李凯,杨璐璐,史才军.基于不规则骨料堆积结构的混凝土水渗透性的研究[J].材料导报,2024,38(12):103-110.
[3] 杜修力,金浏,张仁波.力学荷载对混凝土中氯离子渗透扩散行为影响述评[J].建筑结构学报,2016(1):107-125.
[4] 金立兵,余化龙,王振清,等.再生混凝土抗氯离子渗透的五相细观数值模拟[J].郑州大学学报(工学版),2022,43(1):83-89.
[5] 孙国文,孙伟,王彩辉.现代混凝土传输行为与其微结构之间关系的研究方法及其进展[J].材料导报,2018,32(17):3010-3022.
[6] 肖梦霞.基于BP神经网络高温高湿高压条件下混凝土气体渗透性预测[D].哈尔滨:哈尔滨工程大学,2023.
[7] 吴贤国,刘茜,王洪涛,等.基于随机森林和支持向量机的高性能混凝土抗渗性预测研究[J].硅酸盐通报,2021,40(3):829-835,844.
[8] ?魻ztas A,Pala M,?魻zbay E,et al.Predicting the compressive strength and slump of high strength concrete using neural network[J].Construction and Building Materials,2006,20(9):769-775.
[9] 姚小俊,吴迪.基于贝叶斯优化高斯过程回归法的再生混凝土力学性能预测[J].科学技术与工程,2023,23(7):2968-2975.
[10] 卢俊齐,蒋明强,汪金满.基于多参数分析的再生骨料混凝土的抗冻性研究[J].技术与市场,2025,32(5):32-36.
[11] 张欣怡,戴成元,李微雨,等.基于TPE-XGBoost算法的再生粗骨料混凝土抗压强度预测模型[J].建筑科学与工程学报,2024,41(6):100-110.
[12] 张志宇,甘楠林,刘伟,等.基于随机森林的自密实再生混凝土抗压强度预测研究[J].河南科学,2024,42(12):1773-1783.
[13] 鲍玖文,王云伟,牟新宇,等.持压荷载与干湿循环作用下再生混凝土氯盐侵蚀行为[J].复合材料学报,2023,40(2):1015-1024.
[14] 袁化强.基于胶凝材料颗粒间相互作用力及骨料特性的自密实混凝土流变性及机理研究[D].济南:山东大学,2023.
[15] Vázquez E,Barra M,Aponte D,et al.Improvement of the durability of concrete with recycled aggregates in chloride exposed environment[J].Construction and Building Materials,2014,67:61-67.
[16] Obla K H,Lobo C L,Kim H.Tests and criteria for concrete resistant to chloride ion penetration[J].ACI Materials Journal,2016,113(5):621-631.
[17] 胡波,柳炳康,张李黎.再生混凝土氯离子渗透性能测试与分析[J].合肥工业大学学报(自然科学版),2009,32(8):1240-1243.
[18] Tuyan M,Mardani-Aghabaglou A,Ramyar K.Freeze-thaw resistance, mechanical and transport properties of self-consolidating concrete incorporating coarse recycled concrete aggregate[J].Materials and Design,2014,53:983-991.
[19] 高延红,郭宝玲,汪蒙,等.自然潮差环境下混凝土渗透性稳定时间及机理[J].水力发电学报,2022,41(1):50-62.
[20] 王晓明.水泥基渗透结晶型防水剂对不同水灰比混凝土的影响研究[J].中国建筑防水,2023(4):13-16.
[21] 张鹏鑫.冻融环境下矸石基混凝土毛细吸水及抗氯离子渗透性研究[D].西安:西安科技大学,2023.
[22] 田爽.水饱和度对再生混凝土力学性能影响的试验研究[D].西安:西安理工大学,2023.
[23] 张彬,杨硕,周军霞.多维荷载对不同含水率混凝土渗透性的影响[J].长江科学院院报,2019,36(4):109-112.
[24] Bao Jiuwen,Li Shuguo,Zhang Peng,et al.Influence of the incorporation of recycled coarse aggregate on water absorption and chloride penetration into concrete[J].Construction and Building Materials,2020,239:117845.
[25] 毛高峰.再生粗骨料混凝土试验研究[D].青岛:青岛理工大学,2008.
[26] 覃荷瑛.再生混凝土氯离子渗透性试验研究及细观数值模拟[D].南宁:广西大学,2012.
[27] 韩帅.再生粗骨料品质和替代率对再生混凝土耐久性能的影响[D].青岛:青岛理工大学,2015.
[28] 张超,任鹏鹏,邓鹏,等.再生骨料透水混凝土力学性能和耐久性研究[J].湖南大学学报(自然科学版),2023,50(3):185-194.
[29] Jones M R,Dhir R K,Gill J P.Concrete surface treatment: Effect of exposure temperature on chloride diffusion resistance[J].Cement and Concrete Research,1995,25(1):197-208.
[30] Al-Amoudi O S B,Al-Kutti W A,Ahmad S,et al.Correlation between compressive strength and certain durability indices of plain and blended cement concretes[J].Cement and Concrete Composites,2009,31(9):672-676.
[31] 王晓飞.再生粗骨料品质和替代率对再生混凝土力学性能的影响[D].青岛:青岛理工大学,2015.
[32] Dunlop P,Smith S.Estimating key characteristics of the concrete delivery and placement process using linear regression analysis[J].Civil Engineering and Environmental Systems,2003,20(4):273-290.
[33] Lawrence L K.Probability, Statistics, and Decision for Civil Engineers[J].Technometrics,1971,13(1):211.
[34] 张品.基于机器学习算法的盾构掘进地表沉降预测及控制[D].长沙:湖南大学,2019.
[35] 彭龙,韩国庆,邬书豪,等.基于机器学习算法的CO2腐蚀速率预测[J].西安石油大学学报(自然科学版),2023,38(2):113-121.
[36] 田霞,徐瑞民,孙倩主编.统计计算[M].北京:电子工业出版社,2023.
引用本文
胡健.再生骨料混凝土氯离子扩散系数机器学习预测模型对比研究[J].混凝土与水泥制品,2026,53(9):112-119. Hu Jian.Comparative study on machine learning prediction models for chloride diffusion coefficients of recycled aggregate concrete[J].China Concrete and Cement Products,2026,53(9):112-119 (in Chinese).
上一个： 粉煤灰掺量对氯氧镁水泥凝结时间、体积 稳定性和微观结构的影响
下一个： 水灰比与云母含量耦合作用下花岗岩机制砂混凝土力学、抗侵蚀性能及体积稳定性研究
相关文件