数据驱动的UHPC与钢筋界面黏结强度研究

Study on Interface Bond Strength Between UHPC and Rebar Using Data Driven Method

超高性能混凝土(Ultra-high Performance Concrete,UHPC)与钢筋界面的良好黏结是保证结构安全和正常使用的关键,直接决定了结构设计与性能评估。然而,传统基于试验或力学推导的界面黏结性能评估方法难以反映众多因素的影响,预测精度低、方差大。近年来,基于人工智能的数据驱动技术发展迅猛,为解决上述问题提供了新思路。建立了样本容量为670的钢筋与UHPC界面黏结试验数据库,分析了特征相关性和主要因素影响规律。剔除数据缺失、数据噪声等无效数据后,得到了557组有效数据子库。基于机器学习方法训练并生成了9种黏结强度预测模型,包括4个线性模型和5个非线性模型。采用决定系数(R^(2))、均方根误差(RMSE)和平均绝对误差(MAE)指标开展了模型预测精度评价,对比了机器学习模型和传统模型的预测精度。结果表明:钢纤维掺量2%的试验样本占75.8%,钢筋直径16 mm的试验样本占64.6%,缺乏其他纤维掺量与钢筋直径的研究样本,尤其是低配纤、高配纤以及大直径钢筋的试验数据。5次随机抽样训练预测结果中,人工神经网络模型对试验结果预测最好,R^(2)、RMSE和MAE分为0.918、4.22和2.51,支持向量机模型对试验结果预测最差。对全集预测结果中,规范和学者提出的传统模型精度不足且过于保守,精度最高的模型预测结果R^(2)为0.474、RMSE为11.5、MAE为9.5。9种机器学习方法中,人工神经网络模型和树类模型对试验结果预测最好,精度最高模型R^(2)为0.966、RMSE为2.9、MAE为1.6;相较精度最高的传统模型,其R^(2)提高了103.8%,RMSE和MAE分别降低了74.8%和83.2%,表明机器学习模型可大幅提升预测精度,为UHPC与钢筋界面黏结强度计算提供了新的思路。

High bond strength between ultra-high-performance concrete(UHPC)and steel reinforcing bars is critical for ensuring structural safety and serviceability;thus,it dominates the structural design and performance assessment.However,conventional bond strength prediction methods based on experimental tests and theoretical approaches do not consider the effects of all the design parameters,resulting in inaccurate results.Recently,the development of artificial intelligence has provided new insights into predicting bond strength.This study developed a database of 670 tests on bond strength to investigate the relevance and effects of the primary design parameters.A sub-database consisting of 557 tests was established,eliminating invalid and noisy data.Nine estimation models for the bond strength,including four linear and five nonlinear models,were trained and generated using machine learning methods.The results were evaluated using the determination coefficient(R^(2)),root mean square error(RMSE),and mean absolute error(MAE).The following conclusions were drawn:In total,75.8%of the data points are for specimens with 2%steel fibers,and 64.6%of the data points are for steel bars with the diameter of 16 mm.Data for other steel fiber volumes and rebar diameters,particularly with low steel fiber volumes and large-diameter rebars,are lacking.For the results from the five random samplings,the artificial neural network(ANN)model yields the best estimation results,with R^(2),RMSE,and MAE values of 0.918,4.22,and 2.51,respectively.The support vector machine(SVM)produces results with the lowest accuracy.Considering the prediction results for the entire database,the current codes and available theoretical approaches yield inaccurate and conservative results,with R^(2),RMSE,and MAE values of 0.474,11.5,and 9.5,respectively.Among the nine machine learning methods,the ANN model provides the best prediction with R^(2),RMSE,and MAE values of 0.996,2.9,and 1.6,respectively;thus,R^(2) increased by 103.8%,while the RMSE and MAE reduced by 74.8%and 83.2%,respectively.Therefore,machine learning methods provide new insights for estimating the bond strength in UHPC,exhibiting improved prediction results.