基于深度学习的桥梁非线性气动力模型研究  

Nonlinear Aerodynamic Force Model of Bridge Based on Deep Learning

作　　者：张文明[1] 冯丹典 葛耀君[2] ZHANG Wen-ming;FENG Dan-dian;GE Yao-jun(Key Laboratory of C&PC Structures of Ministry of Education,Southeast University,Nanjing 210096,China;State Key Laboratory for Disaster Reduction in Civil Engineering,Department of Bridge Engineering,Tongji University,Shanghai 200092,China)

机构地区：[1]东南大学混凝土及预应力混凝土结构教育部重点实验室,江苏南京210096 [2]同济大学土木工程防灾国家重点实验室,上海200092

出　　处：《桥梁建设》2023年第3期16-24,共9页Bridge Construction

基　　金：国家重点研发计划项目(2022YFB3706703);国家自然科学基金项目(52078134)。

摘　　要：为准确模拟桥梁断面气动力的非线性特性和流体记忆效应,构建了基于深度学习的非线性气动力降阶模型。引入前馈神经网络(FNN)和长短时记忆(LSTM)网络2种深度学习框架,利用CFD强迫振动数值模拟获取非线性气动力数据,采用谐波叠加法合成强迫振动位移信号;结合2种框架的结构特征,以断面位移为输入、气动力为输出,针对性构建了用于网络训练、验证和测试的数据集。以某三塔悬索桥钢箱梁断面为例,分别建立基于FNN和基于LSTM网络的气动力降阶模型,并针对不同频率、自由度的强迫振动和自由振动等工况,评估对比了模型的性能。结果表明:2种降阶模型均可较好地模拟任意合理振动工况下断面非线性气动力,计算效率较数值模拟有极大提升,其中,基于LSTM网络的降阶模型具备更优的非线性气动力模拟性能。Nonlinear aerodynamic reduction models based on deep learning were developed to accurately simulate the nonlinear cross-sectional aerodynamic characteristics and fluid memory effects of bridge.Two frameworks of deep learning,namely feedforward neural network(FNN)and long short-term memory(LSTM)network,were introduced.Nonlinear aerodynamic data was obtained from computational fluid dynamics(CFD)forced vibration numerical simulation,in which the simulated displacement signals were synthesized by harmonic superposition.The data sets for network training,validation,and testing were established based on different structural characteristics of the two frameworks,taking cross-sectional displacements as inputs and aerodynamic forces as outputs.With the case of the steel box stiffening girder of a three tower suspension bridge,the aerodynamic reduced order models(ROMs)based on FNN and LSTM networks were built up,respectively.The performance of the two models were evaluated and compared under forced vibration conditions with different frequencies and degrees of freedom and free vibration conditions.Results demonstrate that both of the two ROMs based on FNN and LSTM networks can effectively simulate the nonlinear cross-sectional aerodynamic forces of bridge under any reasonable vibration conditions,and the computational efficiency is greatly improved compared to numerical simulation.Among them,the ROM based on LSTM network has better nonlinear aerodynamic simulation ability.

关 键 词：桥梁工程 非线性气动力 深度学习 前馈神经网络 长短时记忆网络 数值模拟 

分 类 号：U441.2[建筑科学—桥梁与隧道工程]