基于BP神经网络的有限元应力修匀的研究

Study on finite element stress smothing based on BP neural network

以四边形平面应力等参单元为例,以高斯积分点为样本点,基于BP神经网络分别考察了以下两种训练模型:以高斯积分点的坐标为输入,Mises应力为输出;以高斯积分点的坐标和位移同时为输入,Mises应力为输出。通过与传统的有限元整体应力修匀比较,经过BP神经网络学习得到的结点Mises应力,不论是单独以坐标为输入还是以坐标和位移同时作为输入得到的结点Mises应力精度都比有限元传统的整体应力修匀高,并且第二种BP神经网络训练模型得到的结点Mises应力比第一种的精度要高。

This paper takes the quadrilateral plane stresses isoparametric element as an example and Gaussian integral points as the sample points, based on the BP neural network, there are two training models investigated：The coordinates of the Gaussian integration points are input and the Mises stress is output. The coordinates and displacements of the Gaussian integral points are input at the same time, and the Mises stress is output. Compared with the traditional finite element global stress smoothing, the Mises stress of the nodes is studied by BP neural network, whether it is the coordinate as the input or the coordinate and the displacement as the input at the same time, the accuracy of the Mises stress is higher than that of the traditional finite element global stress smoothing, and the second type of BP neural network training model has the higher accuracy than the first one.