Strain-displacement reconstruction method of finite element model structure

Abstract

The invention discloses a strain-displacement reconstruction method of a finite element model structure, particularly relates to a method for constructing a strain-displacement relationship by adopting a NeuroDynaStract neural network based on a finite element structure, and belongs to the field of finite element displacement reconstruction. The method is used for performing strain-displacement reconstruction on a finite element model based on a NeuroDynaStructure neural network architecture, and comprises the following steps: fitting measured strain and measured displacement through an established NeuroDynaStructure neural finite element algorithm by adopting a finite element method based on basic measured data, constructing a training set, and performing training by adopting a radial basis network algorithm, in the training process, a principal component analysis (PCA) method is used for carrying out dimension reduction processing, the operand is reduced, and a correction strain-correction displacement relation is obtained. The method is used for reconstructing deformation of the real-time nonlinear finite element simulation model, the finite element model and the neural network are combined together, a novel method for obtaining the strain-displacement relation based on the finite element model and test data is provided, and the structural reconstruction precision can be improved while the neural network architecture is simplified.

Description

technical field [0001] The invention relates to a strain-displacement reconstruction method of a finite element model structure, in particular to the construction of a strain-displacement relationship based on a finite element structure using a NeuroDynaStruct neural network, and belongs to the field of finite element displacement reconstruction. Background technique [0002] As an alternative to mechanical modeling, neural networks have been used to solve some engineering problems, such as structural vibration and structural stability. Neural networks reduce computation time and are trained only with experimental data, without the need to identify material parameters. For example, a neural network-based beam element combines neural network methods with classical finite element methods to develop intelligent finite elements, which have lower computational costs than classical methods. The neural network is combined with the finite element method for strain-displacement reco...

AI Technical Summary

Problems solved by technology

[0009] 1. The data requirements of the training set are relatively large, and when the amount of data is large, it will have a significant impact on the calculation speed

[0010] 2. The training algorithm adopted is BP neural network. The training accuracy of this algorithm increases with the increase of the training set. However, as the training data increases, over-fitting phenomenon is prone to occur, resulting in a decrease in training accuracy.

[0011] 3. The inverse finite element method and neural network are not combined to solve the strain-displacement relationship. 4. The training data set is not processed by PCA, and the training data volume is large, which affects the calculation speed

Images