【作者】 费庆国； 张令弥；

【Author】 FEI Qing-guo, ZHANG Ling-mi(Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China)

【机构】 南京航空航天大学振动工程研究所；

【摘要】 结构动态有限元模型修正通常归结为反问题求解。采用RBF神经网络进行有限元模型的修正,将特征量视为自变量,设计参数视为因变量,以RBF神经网络逼近两者之间的非线性映射关系,然后利用神经网络的泛化特性直接求解设计参数的目标值,从而模型修正归结为正问题进行研究。该方法避开了现有的基于反问题的有限元模型修正所面临的复杂的非线性优化计算。研究了基于RBF神经网络的模型修正的关键技术问题,包括网络的训练算法、样本点选择策略等。以欧洲航空科技组织的基准模型——GARTEUR飞机模型为例,应用该方法对其有限元模型进行修正。基于仿真数据的修正,模态频率误差在1%以内,参数误差在2%以内。基于测量数据的修正,模态频率误差在3%以内。更多还原

【Abstract】 Finite element model updating used to be concluded as an inverse problem. Features of the structure are looked on as the function of design parameters. Parameters are updated based on first derivative. This paper presents a method which concludes model updating as a positive problem. Features and design parameters are regarded as independent variables and dependent variables respectively. The trained radial basis function neural network is utilized as map function between them. The target value of design parameter can be estimated directly due to the generalization character of neural network. The method avoids solving the complicated optimization problem which is a common question for present methods. Some key problems such as training algorithm and data sampling techniques are also discussed. The benchmark of European academia, GARTEUR, is used in the simulation. Finite element model of GARTEUR is updated using the presented method. Error of physical parameters and modal frequencies decrease to 1% and 2% respectively when the simulated data is used as reference. When experimental data is used, error of modal frequencies is less than 3%.更多还原