Methods for Hybrid Aeroelastic Analysis of Structures

细长结构（如大跨度桥梁）的气动弹性行为需要准确预测，作为其设计的一部分。 当暴露于大气风流时，它们会产生显著的振动，因此需要用于预测风-结构相互作用现象（如涡激振动、抖振和颤振）的实际模型。空气动力学分析的典型方法是经验模型、实验（风洞）模型和最近的数值（计算流体动力学，CFD）模型。对于结构的分析，这些可以结合起来，需要耦合到结构动力学模型。通过敏感性分析，可以量化各个模型组件对全局模型预测及其质量的影响。控制总体预测质量的有前途的概念是自适应或混合建模方法，其中选择模型组件以保持预测质量和效率之间的平衡。该项目的目的是开发一个通用模型框架，使用混合建模方法计算风引起的动态结构响应，该方法结合了可以灵活选择和交换的不同模型，以平衡预测精度和计算速度。这应在元建模技术的帮助下完成，元建模技术用于部分或全部替换具有数学表达式的计算成本高的旋涡粒子CFD模拟，从而在不显著影响预测精度的情况下显著减少计算工作量。神经网络是机器学习技术的一种，在此将网络用作风-结构相互作用中不同强迫项的元模型。应采用敏感性分析来量化单个输入参数对模型预测的影响，以指导模型选择过程。开发的元模型将根据适当的参考模型（如风洞试验）进行验证。这些都采用了伪三维结构建模，它允许同时考虑气动导纳，运动诱导力和旋涡脱落进行分析。因此，可以基于布置在结构元件处的气动弹性模型的灵活混合组合在时域中执行3D结构的气动分析，其中元模型提供高计算效率。建模框架一般适用于任何线状结构，如大跨度桥梁，塔架和桅杆，并具有显着的潜力，为未来的应用风工程问题。

The aeroelastic behavior of slender structures such as long-span bridges needs to be accurately predicted as part of their design. They can develop significant vibrations when exposed to atmospheric wind flow and therefore realistic models for the prediction of wind-structure interaction phenomena such as vortex-induced vibrations, buffeting and flutter are required. Typical methods for aerodynamic analysis are empirical models, experimental (wind tunnel) models and, more recently, numerical (Computational Fluid Dynamics, CFD) models. For the analysis of a structure these can be combined and need to be coupled to structural dynamics models. The influence of the individual model components on the global model prediction and its quality can be quantified by means of sensitivity analyses. Promising concepts to control the overall prediction quality are adaptive or hybrid modelling approaches, where the model components are chosen such as to keep a balance between prediction quality and efficiency. The aim of this project is to develop a general model framework to compute the dynamic structural response due to wind using a hybrid modelling approach which combines different models that can be flexibly selected and exchanged to balance prediction accuracy and computational speed. This shall be done with the help of meta-modelling techniques employed to replace, partially or fully, computationally expensive Vortex Particle CFD simulations with mathematical expressions, thus dramatically reducing computational effort without significantly compromising prediction accuracy. Neural networks are a category of machine learning techniques whereby the network is here used as a meta-model for the different forcing terms in the wind-structure interaction. Sensitivity analyses shall be adopted to quantify the influence of individual input parameters on the model prediction to guide the model selection process. Developed meta-models will be validated against suitable reference models such as wind tunnel tests. These are employed in a pseudo-3D structural modelling, which allows the analysis to be performed by simultaneously considering aerodynamic admittance, motion-induced forces and vortex shedding. The aerodynamic analysis of a 3D structure can thus be performed in time domain based on a flexible hybrid combination of aeroelastic models arranged at the structural elements, with the meta-models providing high computational efficiency. The modelling framework is generally applicable to any line-like structure such as long-span bridges, towers and masts and has significant potential for future applications to wind engineering problems.