基于神经网络的非线性结构有限元模型修正研究

现有的动态有限元模型修正方法几乎都是建立在线性假设基础之上，修正中利用固有频率等线性系统特征量。工程中，真实的结构振动系统都是非线性的。虽然在许多情况下，线性化假设获得的结果能够较为准确地反映真实系统的特性。但是，在结构的非线性特征较为明显时，必须考虑非线性因素，这时，现有的模型修正方法将不再适用。现以非线性梁为研究对象，采用基于神经网络的修正方法探索了非线性结构的有限元模型修正问题。仿真研究中利用有限元分析的响应数据训练神经网络。修正结果表明，包括非线性弹簧刚度系数在内的三个设计参数修正后误差均在1％以内。说明基于神经网络的有限元模型修正方法适用于解决非线性结构的有限元模型修正问题。

Study on Finite Element Model Updating of Nonlinear Structures Using Neural Network

Most of the dynamic finite element (FE) model updating methods is proposed with the assumption that structures are linear. In fact, non-linearity can often be observed when investigating the dynamic behavior of real structures. Under most circumstance, linear models of such structures are able to describe the dynamic behavior with certain accuracy. However, when the non-linear effects are not neglectable, non-linearity must be considered. Present updating methods wouldn't be applicative any longer. This paper presents the application of the newly proposed updating method on the updating of a beam with nonlinear component. FE model of the beam is updated using simulated experimental response. Frequency Responses are used as training data for neural network. The error of the stiffness coefficient of non-linear spring is less than 1%. This provides a feasible solution for FE model updating of nonlinear structures.