基于特征向量信息支持向量机的RC框架易损性曲线预测  

Eigenvectors-informed Support Vector Machines for Fragility Curve Predictions of RC Frames

作　　者：周宇 骆欢 ZHOU Yu;LUO Huan(Hubei Geological Disaster Prevention and Control Engineering Technology Research Center,Yichang 443002,Hubei,China;College of Civil Engineering&Architecture,China Three Gorges University,Yichang 443002,Hubei,China)

机构地区：[1]湖北省地质灾害防治工程技术研究中心,湖北宜昌443002 [2]三峡大学土木与建筑学院,湖北宜昌443002

出　　处：《地震研究》2024年第3期359-368,共10页Journal of Seismological Research

基　　金：湖北省自然科学基金面上项目(2022CFB294);国家自然科学基金青年科学项目(52208485).

摘　　要：易损性曲线将结构破坏等级与地震动强度相关联,能够直观地反映结构破坏的概率,但在建立易损性曲线的过程中需要大量的结构非线性时程分析结果,因而计算效率不高。机器学习方法已被证明能较好地解决这一问题,但当训练数据的规模较大时,由于训练过程涉及求解大规模逆矩阵致使计算效率依然低下。为此,本文提出了一种特征向量信息支持向量机(EILS-SVM)的新方法克服此类方法的不足。在大规模数据集下,EILS-SVM能够筛选小规模子样本建立低秩核矩阵。这使得其训练过程只需求解小规模低秩矩阵的逆矩阵,进而极大提高计算效率。为了验证EILS-SVM的准确性和高效性,基于16500个钢筋混凝土(RC)框架在地震作用下的破坏数据,分别与支持向量机(LS-SVM)、随机森林、神经网络、线性判别分析(LDA)、贝叶斯作对比。结果表明,EILS-SVM能准确预测RC框架的易损性曲线,其计算效率最高能提升近27倍。Fragility curves establish a correlation between structural damage levels and seismic intensity,offering an intuitive depiction of the probability of structural failure.However,the generation of these curves necessitates a substantial amount of structural nonlinear time-history analysis results,thereby rendering the computational process inefficient.Machine learning techniques have been proven to effectively address this issue,yet their efficacy diminishes with the increase in the scale of training data due to the computational demands of solving large-scale inverse matrices during the training phase.In response,this paper proposes a novel methodology,the Eigenvector Information-supported Support Vector Machine(EILS-SVM),which surmounts the limitations associated with these techniques.By employing a selective subsample to construct a low-rank kernel matrix in the context of large-scale datasets,the EILS-SVM method requires only the inversion of small-scale,low-rank matrices,significantly enhancing computational efficiency.To validate the accuracy and efficiency of the EILS-SVM,it is benchmarked against conventional models such as the Least Squares Support Vector Machine(LS-SVM),Random Forest,Neural Networks,Linear Discriminant Analysis(LDA),and Bayesian methods,using a dataset comprised of 16500 instances of damage in Reinforced Concrete(RC)frames subjected to seismic activities.The results indicate that the EILS-SVM is capable of accurately predicting the fragility curves of RC frames,with a computational efficiency improvement of up to 27 times compared to existing methodologies.

关 键 词：钢筋混凝土框架 易损性曲线 特征向量 支持向量机 机器学习 

分 类 号：TU973.2[建筑科学—结构工程]