机器学习方法能通过建立以建筑信息、地震动参数为输入,韧性指标为输出的非线性映射关系,对建筑结构进行抗震韧性评估,但当训练数据规模较大时,其训练过程由于涉及求解大规模逆矩阵致使计算效率低下且极其占用计算机内存.为此,提出基向量引导的支持向量机(basis vectors-guided support vector machines for regression,BVLS-SVMR)模型,从大规模训练样本中提取小规模子样本,并将其映射到高维特征空间里作为基向量,替代大规模原基向量用于建立预测模型.为了验证BVLS-SVMR模型的准确性和高效性,基于9 356个钢筋混凝土(RC)框架(教学楼)抗震韧性的数据,分别与支持向量机(least squares support vector machines for regression,LS-SVMR)模型和传统有限元法(FEM)进行对比.结果表明:BVLS-SVMR 模型的测试集预测精度与LS-SVMR模型的测试集预测精度(决定系数R2)相差0.011,但计算时间是LS-SVMR模型的1/10,是传统FEM的1/21 709;BVLS-SVMR模型能准确且快速地预测教学楼的抗震韧性指标.
Basis vectors-guided support vector machines for seismic resilience assessment of RC frames
Machine learning methods can evaluate the seismic resilience of buildings by establishing a nonlinear mapping relationship between inputs related to building information and seismic parameters and outputs representing resilience indicators.However,large training datasets pose challenges due to the computation of large-scale inverse matrices,leading in low computational efficiency and high memory usage.To address this issue,we propose a basis vector guided support vector machine regression(BVLS-SVMR)model.Replacing the original large-scale basis vectors for building predictive models,this model extracts small-scale sub-samples from large training datasets and maps them into a high-dimensional feature space as basis vectors.To validate its accuracy and efficiency,seismic resilience data from 9 356 reinforced concrete(RC)frame buildings(school buildings)were used.Compared with the support vector machine(LS-SVMR)model and the traditional finite element method(FEM),these results demonstrate the BVLS-SVMR model exhibited a test set prediction accuracy difference of only 0.011 compared to the LS-SVMR model and its computation time was only 1/10 of the LS-SVMR model and 1/21 709 of the traditional FEM.This proves BVLS-SVMR model's ability to accurately and rapidly predict seismic resilience indicators for school buildings.
国家科技期刊平台
NSTL
万方数据
