基于物理驱动支持向量机方法的地震作用下结构动力响应求解  

Solving structural dynamic response under earthquake based on physics-driven SVM method

作　　者：杜轲[1,2] 吴文贤 林志鹏 骆欢 DU Ke;WU Wenxian;LIN Zhipeng;LUO Huan(Key Laboratory of Earthquake Engineering and Engineering Vibration,Institute of Engineering Mechanics,China Earthquake Administration,Harbin 150080,China;Key Laboratory of Earthquake Disaster Mitigation,Ministry of Emergency Management,Harbin 150080,China;College of Civil Engineering&Architecture,China Three Gorges University,Yichang 443002,China)

机构地区：[1]中国地震局工程力学研究所地震工程与工程振动重点实验室,哈尔滨150080 [2]地震灾害防治应急管理部重点实验室,哈尔滨150080 [3]三峡大学土木与建筑学院,湖北宜昌443002

出　　处：《振动与冲击》2025年第3期284-290,共7页Journal of Vibration and Shock

基　　金：中国地震局工程力学研究所基本科研业务费专项项目(2023B07);国家重点研发计划(2023YFC3805203)。

摘　　要：物理驱动机器学习是一种将物理原理融入机器学习框架的前沿方法。通过引入物理知识,该方法旨在使模型更为贴合实际世界的物理规律和约束,以提高模型在学习过程中对数据本质特征的准确捕捉。该研究使用了一种以支持向量机为基础的物理驱动方法,用于精确计算结构的动力响应。该算法通过最小化多输出最小二乘支持向量机的目标函数,实现了对回归模型参数的精准拟合。同时,通过在特征空间中引入系统动态平衡方程和初始条件的物理约束,无需事先训练数据即可有效计算结构的动力响应。随后开展在地震动荷载作用下的单自由度体系和二层剪切框架多自由度体系的动力响应,并将所用方法与传统方法的结果进行了对比。分析结果表明,提出的物理驱动机器学习方法在精度和大时间步长性能方面均显著优于传统方法。Physics-driven machine learning is a cutting-edge method integrating physical principles into machine learning framework.By introducing physics knowledge,this method aims to make a model more fit physical laws and constraints of real world,and improve the model’s correct capture of data essential features in learning process.Here,a physics-driven method based on support vector machine(SVM)was proposed to accurately calculate structural dynamic response.This algorithm could realize accurate fitting of a regression model’s parameters by minimizing the objective function of a multi-output least squares SVM.Meanwhile,by introducing physical constraints of a system dynamic equilibrium equation and initial conditions in feature space,structural dynamic response could be effectively calculated without the need for prior training data.Subsequently,dynamic responses of a single-DOF system and a two-layer shear frame multi-DOF system under seismic load were calculated,the results of the proposed method were compared with those of traditional methods.The analysis results showed that the proposed physics-driven machine learning method is significantly superior to traditional methods in terms of accuracy and large time step performance.

关 键 词：机器学习 支持向量机 物理驱动 无标记数据 结构动力响应分析 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3