基于机器学习的钢筋混凝土单偏压柱纵向配筋率预测方法研究  

Research on longitudinal reinforcement ratio prediction method of reinforced concrete single eccentric compression columns based on machine learning

作　　者：肖岩松 杨志勇 乔保娟[1,2] Xiao Yansong;Yang Zhiyong;Qiao Baojuan(China Aecademy of Building Researh,Bejing 100013,China;CABR Technology Co.Ltd,Architectural Structure Nonlinear Simulation Application Research Center,Beijing 100013,China)

机构地区：[1]中国建筑科学研究院有限公司,北京100013 [2]中建研科技股份有限公司建筑结构非线性仿真应用研究中心,北京100013

出　　处：《建筑科学》2024年第5期11-20,共10页Building Science

摘　　要：钢筋混凝土(RC)柱纵向配筋计算方法对于保证钢筋混凝土结构安全具有重要意义,各国规范均提出了简便高效的RC柱纵向配筋计算方法。由于采用了一系列简化假定,现有规范计算方法具有一定离散性,计算准确性和精度也有待提升。本文广泛搜集了国内外RC柱试验数据,采用材料强度实测值对中国规范、美国规范和欧洲规范RC柱纵向配筋计算方法进行了校核,结果表明3种计算方法所得小偏压受力状态下RC柱纵向配筋率离散性较大。本文提出了基于机器学习的RC柱纵向配筋率计算方法,基于530个试验样本,训练了决策树、支持向量机、随机森林和极致梯度提升4个机器学习模型,以模型拟合优度、泛化性能与复杂程度为标准进行特征迭代,采用随机搜索和十折交叉验证方法进行超参数调优,通过在测试集上的综合性能评价,确定了极端梯度提升(eXtreme Gradient Boosting,下简称XGBoost)最优模型。采用沙普利加性解释(SHapley Additive exPlanations,下简称SHAP)方法与部分依赖图对XGBoost模型进行解释,对模型输出的特征参数影响进行了可视化,证明了特征参数影响的规律与力学规律相符,验证了本文机器学习模型的可靠性。相比于规范计算方法,机器学习模型显示了更高的精度和更低的离散性,可作为“物理驱动”RC柱纵向配筋率计算方法的有效补充,为更加准确高效地完成RC柱设计提供了新途径。Longitudinal reinforcement calculation in reinforced concrete(RC)columns is vital for ensuring the safety of RC structures.International standards provide simplified and efficient methods for this purpose.Due to inherent assumptions,these standard methods exhibit variability,necessitating improvements in accuracy and precision.This paper extensively collects global RC column test data,validating the longitudinal reinforcement calculation methods of Chinese,American,and European standards using actual material strength values.The analysis shows significant variability in reinforcement ratios under small eccentric loading across all methods.A machine learning-based approach for calculating RC column longitudinal reinforcement ratios is proposed,utilizing 530 experimental samples to train Decision Tree,Support Vector Machine,Random Forest,and Extreme Gradient Boosting models.Feature iteration based on model fit,generalization,and complexity,along with hyperparameter optimization using random search and ten-fold cross-validation,identifies the XGBoost model as optimal.SHapley Additive exPlanations(SHAP)methods and partial dependence plots elucidate the XGBoost model,visualizing the impact of feature parameters consistent with mechanical principles.Compared to standard methods,the machine learning model demonstrates higher accuracy and reduced variability,serving as an effective complement to"physics-driven"RC column design methods,offering a new avenue for more precise and efficient RC column design.

关 键 词：钢筋混凝土柱 承载力设计 纵向配筋率 机器学习 超参数调优 可解释性 

分 类 号：TU318[建筑科学—结构工程]