CN-122020835-A - Vehicle body structure performance prediction method and system based on finite element and graph neural network

CN122020835ACN 122020835 ACN122020835 ACN 122020835ACN-122020835-A

Abstract

The invention provides a vehicle body structure performance prediction method and system based on finite element and a graph neural network, which comprises the steps of obtaining an input file of finite element analysis, extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, each unit is of a geometric structure and comprises a plurality of sides and vertexes, constructing different graphs according to the units, the sides and the vertexes, determining the characteristics of a heterogeneous graph, inputting the characteristics of the heterogeneous graph into a graph neural network proxy model, obtaining a stress component prediction result of a heterogeneous graph node, and determining the vehicle body structure performance according to the stress component prediction result of the heterogeneous graph node. According to the method, after the training of the graph neural network proxy model is completed, only a second level is needed for prediction, and the requirement of rapid iterative optimization in the vehicle body structure design can be met. The constructed heterogram can accurately reserve the hierarchical relation and geometric characteristics of the unit-edge-vertex of the finite element analysis three-dimensional entity network, avoid the loss of structural information and ensure the accuracy of stress component prediction.

Inventors

PAN WENFENG
ZHAO LIBIN
AN YUHANG
SHI BINGYAN

Assignees

河北工业大学
天津文宇元智科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (10)

1. A vehicle body structure performance prediction method based on finite element and graph neural network is characterized by comprising the following steps: The method comprises the steps of obtaining an input file of finite element analysis, and extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, and the units are of a geometric structure and comprise a plurality of edges and vertexes; Constructing an isopgram according to the units, the edges and the vertexes, and determining the characteristics of the isopgram; inputting the heterogeneous graph characteristics into a graph neural network proxy model to obtain a stress component prediction result of heterogeneous graph nodes; And determining the performance of the vehicle body structure according to the stress component prediction result of the heterogeneous graph node.
2. The method of claim 1, wherein the heterogeneous graph features include node features, edge features, and edge index connection relationships; Said constructing an iso-graph from said cells, said edges and said vertices, determining an iso-graph feature comprising: Constructing nodes of an abnormal pattern according to the units, the edges and the vertexes, and determining node characteristics according to the nodes of the abnormal pattern, wherein the nodes of the abnormal pattern at least comprise one of basic graph nodes and additional nodes, the basic graph nodes comprise vertexes, and the additional nodes at least comprise one of edge midpoints, plane barycenters and unit barycenters; Mapping the edges of the units into the edges of the different patterns, and determining edge characteristics; and determining the edge index connection relation according to the connection relation between the edges of the different patterns.
3. The method of claim 2, wherein the node characteristics include at least one of node spatial location coordinates, a node type, and boundary conditions, the node type including an interior node and a boundary node; The edge feature includes at least one of the relative coordinates of the two end points of the edge and the edge length.
4. The method of claim 2, wherein inputting the heterogeneous graph features into a graph neural network proxy model to obtain the stress component prediction result of the heterogeneous graph nodes comprises: Inputting the node characteristics to a node encoder to obtain a node vector; inputting the edge characteristics to an edge encoder to obtain an edge vector; Inputting the node vector, the edge vector and the edge index connection relation into a multi-layer message transmission unit to obtain a node embedded vector; And inputting the node embedded vector to a decoder to obtain a stress component prediction result of the heterogeneous graph node.
5. The method of claim 3, wherein the graph neural network proxy model is trained in the following manner; obtaining an output result file of finite element analysis; determining a strain component true value of a boundary node of the heterogeneous graph according to the output result file; Inputting the stress component prediction result of the heterogeneous graph boundary node and the strain component true value of the heterogeneous graph boundary node into a preset loss function, and training the graph neural network proxy model by adopting an Adam optimizer to minimize the loss function.
6. The method of claim 5, wherein the loss function satisfies physical constraints including stress-strain relationship constraints and energy conservation constraints; for the stress-strain relationship constraint, the online elastic phase satisfies the following equation: ; ; ; ; ; ; Wherein, the 、、 Representing a positive stress; 、、 indicating a positive strain; 、、 representing the shear stress; 、、 represents a shear strain; represents the modulus of elasticity; Representing poisson's ratio; represents the shear modulus.
7. The method of claim 5, wherein the loss function is formulated as: ; Wherein, the Representing the total loss value; Representing the total number of boundary nodes of the iso-graph; Representing the total dimension of the stress component; Represent the first A plurality of boundary nodes; Represent the first A stress component; representing the iso-patterning The first boundary node Predicting results of the individual stress components; representing the iso-patterning The first boundary node True values of the strain components; Representing the super-parameters; Representing a training parameter set of the graph neural network proxy model; Representing the training parameters of the proxy model of the graph neural network.
8. The method of claim 1, wherein the body structure properties include at least one of von Mises equivalent stress, stress concentration area, body structure stiffness, body structure strength, and body structure fatigue life.
9. A vehicle body structure performance prediction system based on finite element and graph neural network, comprising: the finite element extraction module is used for acquiring an input file of finite element analysis and extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, and the units are of a geometric structure and comprise a plurality of edges and vertexes; The heterogeneous graph characteristic determining module is used for constructing a heterogeneous graph according to the unit, the edge and the vertex and determining heterogeneous graph characteristics; the stress component prediction module is used for inputting the heterogeneous graph characteristics into a graph neural network proxy model to obtain a stress component prediction result of the heterogeneous graph nodes; And the vehicle body structure performance determining module is used for determining the vehicle body structure performance according to the stress component prediction result of the heterogeneous graph node.
10. The system of claim 9, wherein the heterogeneous graph features include node features, edge features, and edge index connection relationships; the heterogeneous graph characteristic determining module is used for constructing nodes of an abnormal graph according to the units, the edges and the vertexes, determining node characteristics according to the nodes of the abnormal graph, wherein the nodes of the abnormal graph at least comprise one of basic graph nodes and additional nodes, the basic graph nodes comprise vertexes, and the additional nodes at least comprise one of edge midpoints, plane barycenters and unit barycenters; Mapping the edges of the units into the edges of the different patterns, and determining edge characteristics; and determining the edge index connection relation according to the connection relation between the edges of the different patterns.

Description

Vehicle body structure performance prediction method and system based on finite element and graph neural network Technical Field The disclosure relates to the field of automobile engineering, in particular to a method and a system for predicting the performance of a vehicle body structure based on finite elements and a graph neural network. Background In automotive engineering, body structure performance predictions (e.g., strength, stiffness, fatigue life) typically rely on Finite Element Analysis (FEA). The existing CAE simulation has the following defects although the precision is higher: 1. The calculation time is long, the complete vehicle body finite element model comprises millions of units, and one-time performance prediction usually takes several days to one week, so that the design iteration efficiency is severely restricted; 2. The iteration period is long, when design parameters (such as plate thickness and material properties) change, the simulation is needed again, and the engineering period is greatly prolonged. While some existing fast calculation methods based on the use of empirical formula-based prediction methods or simplified finite element analysis often lack sufficient accuracy and are difficult to capture complex stress distributions. Disclosure of Invention The embodiment of the disclosure provides a vehicle body structure performance prediction method and system based on finite element and graph neural network, which are used for solving the problems of long time consumption, slow design iteration and lack of physical rationality in the existing simulation calculation. Based on the above-mentioned problems, in a first aspect, a method for predicting vehicle body structural performance based on finite element and graph neural network according to an embodiment of the present disclosure includes: The method comprises the steps of obtaining an input file of finite element analysis, and extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, and the units are of a geometric structure and comprise a plurality of edges and vertexes; Constructing an isopgram according to the units, the edges and the vertexes, and determining the characteristics of the isopgram; inputting the heterogeneous graph characteristics into a graph neural network proxy model to obtain a stress component prediction result of heterogeneous graph nodes; And determining the performance of the vehicle body structure according to the stress component prediction result of the heterogeneous graph node. In a second aspect, a vehicle body structure performance prediction system based on finite element and graph neural network is provided, including: the finite element extraction module is used for acquiring an input file of finite element analysis and extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, and the units are of a geometric structure and comprise a plurality of edges and vertexes; The heterogeneous graph characteristic determining module is used for constructing a heterogeneous graph according to the unit, the edge and the vertex and determining heterogeneous graph characteristics; the stress component prediction module is used for inputting the heterogeneous graph characteristics into a graph neural network proxy model to obtain a stress component prediction result of the heterogeneous graph nodes; And the vehicle body structure performance determining module is used for determining the vehicle body structure performance according to the stress component prediction result of the heterogeneous graph node. The beneficial effects of the embodiment of the disclosure include: The method and the system for predicting the vehicle body structure performance based on the finite element and the graph neural network comprise the steps of obtaining an input file of finite element analysis, extracting a three-dimensional entity network, wherein the three-dimensional entity network comprises a plurality of units, each unit is of a geometric structure and comprises a plurality of sides and vertexes, constructing different graphs according to the units, the sides and the vertexes, determining the characteristics of the different graphs, inputting the characteristics of the different graphs into a graph neural network proxy model, obtaining a stress component prediction result of a different graph node, and determining the vehicle body structure performance according to the stress component prediction result of the different graph node. According to the vehicle body structure performance prediction method based on the finite element and the graph neural network, after the graph neural network agent model is trained, only a second level is needed for prediction, and the requirement of rapid iterative optimization in vehicle body structure design can be met. The constructed abnormal pattern ca