Vehicle body structure performance prediction method based on heterogeneous graph neural network

Abstract

The invention belongs to the technical field of engineering structure analysis and optimization, and discloses a vehicle body structure performance prediction method based on a heterogeneous graph neural network. According to the technical key points, the method comprises the following steps: sampling a structure sample required by a data set in a to-be-predicted design domain, and establishing a heterogeneous graph data set by using a heterogeneous graph representation method and a numerical analysis method; defining an HGNN design space by taking the HGT layer as a core; constructing an optimal HGNN model by using a partial heterogeneous graph data set, a controlled random search method and a grid search method; all the heterogeneous graph data sets and the optimal training configuration parameters are used for completing training of the optimal HGNN model; and using the trained HGNN model to predict the performance of any structure sample in the design domain. The method is not only suitable for a homogeneous structure corresponding to a homogeneous graph, but also suitable for a structure which corresponds to a heterogeneous graph and is more widely distributed, the universality of the method is greatly improved, and the expandability of the method is improved on the premise that the application range is not lost.

Description

Technical Field

[0001] The present invention relates to the field of engineering structure analysis and optimization, and in particular to a method for predicting vehicle body structure performance based on heterogeneous graph neural network. Background Art

[0002] In the past decade, deep learning technology has been widely used in the task of predicting the performance of vehicle body structures. In this field, deep learning models are usually used with a certain amount of vehicle body structure analysis data to build a proxy model to predict the target performance of the vehicle body structure. At present, the types of neural networks used for vehicle body structure performance prediction are mainly multi-layer perceptrons (MLP) and convolutional neural networks (CNN). However, a key drawback of these networks is the lack of scalability. Specifically, the proxy models based on MLP or CNN are limited to structures of a specific topology type and will fail immediately when applied to new topology types.

[0003] Although graph neural networks (GNNs) can significantly improve the scalability of surrogate models by leveraging the neighborhood aggregation mechanism, it sacrifices the scope of application of the method. This is because GNNs can only be applied to homogeneous structures corresponding to homogeneous graphs, but not to heterogeneous structures corresponding to heterogeneous graphs. However, the real world contains a wide variety of substances, each of which usually has unique properties. Therefore, heterogeneous graphs are much more common than homogeneous graphs, and it is very important and necessary to propose a vehicle body structural performance prediction method that can be applied to both homogeneous and heterogeneous graphs. Summary of the invention

[0004] In view of the shortcomings in the prior art, the present invention proposes a vehicle body structure performance prediction method that is both scalable and universal by representing the vehicle body structure as heterogeneous graph data and introducing a heterogeneous graph neural network.

[0005] The technical solution adopted by the present invention is:

[0006] A method for predicting vehicle body structure performance based on heterogeneous graph neural network includes the following steps:

[0007] Step 1: Create a heterogeneous graph dataset corresponding to different topological types of vehicle body structures

[0008] (1) Using sampling methods to collect the structural topology type required by the data set within the design domain to be predicted;

[0009] (2) Using the sampling method to change the design parameters of the connection, the same number of structural samples are generated for each of the above structural topology types;

[0010] (3) Decompose the structure sample into vertex components and edge components;

[0011] a. Decompose the structure sample into vertex components of the same or different types;

[0012] b. Determine edge components (of the same or different types) based on the topological relationship between different vertex components in the structure sample;

[0013] c. Assigning feature vectors with the same dimension and different values ​​to each type of vertex component to distinguish different vertex components in each type of vertex component;

[0014] d. Assign a feature vector with the same dimension but different values ​​to each type of edge component to distinguish different edge components in each type of edge component. This step is optional.

[0015] (4) According to the vertex components and edge components obtained after decomposition of the structural sample, they are converted into heterogeneous graph data:

[0016] a. Determine the vertices in the heterogeneous graph: map each type of vertex component to a type of vertex in the heterogeneous graph; map each vertex component to a vertex in the heterogeneous graph, and use the feature vector of the vertex component as the feature of its corresponding vertex;

[0017] b. Determine the edges in the heterogeneous graph: map each type of edge component to a type of edge in the heterogeneous graph; map each edge component to an edge in the heterogeneous graph, and use the feature vector of the edge component as the feature of its corresponding edge.

[0018] (5) Constructing a heterogeneous graph dataset:

[0019] a. Taking the heterogeneous graph data corresponding to the structural sample as the heterogeneous graph sample in the data set;

[0020] b. Use numerical analysis methods to solve the true performance value of each structural sample as the label of the heterogeneous graph sample.

[0021] Step 2: Define the HGNN design space

[0022] (1) HGT layer: Its computation process includes three main components: heterogeneous mutual attention, heterogeneous message passing, and heterogeneous message aggregation for specific targets.

[0023] a. Heterogeneous mutual attention:

[0024]

[0025] Among them, s, t and e represent the source vertex, the target vertex and the edge between the two vertices respectively; H (l-1)[s] and H (l-1)[t] represents the features of the source vertex and the target vertex before the lth layer, respectively; h represents the number of attention heads, d represents the common feature dimension of all vertices in the heterogeneous graph after being projected into the same space, and μ represents the prior tensor encoding the general meaning of each meta-relation; and Weight matrices corresponding to edge type, source vertex type, and target vertex type respectively.

[0026] b. Heterogeneous message passing:

[0027] in, and Represent the weight matrices associated with the source vertex type and edge type respectively;

[0028] c. Aggregation of heterogeneous messages for specific targets:

[0029]

[0030] in, Represents the updated vertex features of the HGT layer.

[0031] (2) Define the intra-layer, inter-layer, and training configuration design dimensions: Define the intra-layer design dimensions based on the variable operations of the above HGT layers and their optional values; define the inter-layer design dimensions based on the variable operations between HGT layers and their optional values; define the training configuration design dimensions based on the training hyperparameters of the network model and their optional values.

[0032] Step 3: HGNN Design Space Search

[0033] (1) Use controlled random search method to search for design dimensions within and between layers:

[0034] a. Select the i-th design dimension dim (i) (i=1, 2, ..., I), I is the total number of intra-layer and inter-layer design dimensions;

[0035] i) Select the first optional value of the design dimension Randomly sample N satisfying Design combination composition collection N is the number of sampled design combinations; use these N design combinations to build HGNN models respectively, and then use some graph datasets and common values ​​of training configuration design dimensions to complete the training, and record the highest prediction accuracy of each model on the validation set As the final score of this design portfolio;

[0036] ii) Select the next optional value for the design dimension $J_i$ is the number of optional values for the current design dimension; the following modifications are made to the $i$-th design dimension values of all $N$ design combinations in i): While keeping the values of other design dimensions unchanged, $N$ new design combinations are generated to form a set Use these $N$ new design combinations to construct HGNN models respectively. Then use part of the graph dataset and the common values of the training configuration design dimensions to complete the training, and record the highest prediction accuracy of each model on the validation set As the final score of this design combination;

[0037] iii) If there are still optional values for this design dimension that have not been taken, return to ii); otherwise, proceed to the next step;

[0038] iv) Take out all the $n$-th ($n = 1, 2,.., N$) design combinations from the set and sort them according to their final scores The design combination with the highest score is marked as 1, the second highest is marked as 2, and so on.

[0039] v) If $n < N$ method, then set $n=n + 1$ and return to iv); otherwise, proceed to the next step;

[0040] vi) Aggregate the above $N$ groups of rankings for all optional values of the current design dimension and determine the optimal design choice $optimal\_val$ of the design dimension $dim$ (i) by analyzing the mean and distribution of the rankings; the smaller the average value of the rankings and the more concentrated the distribution, the better the design choice; (i)

[0041]

[0042] b) If there are still design dimensions in the HGNN design space that have not been taken, then set $i = i + 1$ and return to a); otherwise, proceed to the next step;

[0043] c) Aggregate the optimal design choices $optimal\_val$ of all intra-layer and inter-layer design dimensions above (i) ($i = 1, 2,..., I$) as the input for the next step.

[0044] (2) Construct an optimal HGNN model based on the optimal design choices of intra-layer and inter-layer design dimensions, use the grid search method to search for the training configuration design dimensions, and output the optimal training configuration parameters.

[0044] Step 4: Use the entire graph dataset and the above optimal training configuration parameters to complete the training of the optimal HGNN model, and output the model checkpoint with the highest prediction accuracy on the validation set as the input for the next step.

[0045] Step 5: Randomly sample new structure samples from the design domain to be predicted, and use the above model checkpoints to directly give the predicted value of its performance.

[0046] The present invention overcomes the problem of poor scalability of previous engineering structure performance prediction methods by introducing heterogeneous graph data and heterogeneous graph neural networks. A single model can be used to complete performance prediction tasks for structures of various topological types without sacrificing the application scope of the method. By defining and searching the HGNN design space, the optimal HGNN model and training configuration parameters for the current performance prediction task are found, overcoming the suboptimality of the model in the existing methods and improving the prediction accuracy of the model. Compared with GNN, the versatility of the method is greatly improved, and it can be applied to both homogeneous graph data and heterogeneous graph data, and has broader application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 It is a description of three types of body frames with plates and analysis conditions;

[0049] Figure 3 The process of updating vertex features using the HGT layer;

[0050] Figure 4 It is the prediction accuracy iteration curve of the optimal HGNN model on the validation set;

[0051] Figure 5 Comparison of the prediction results of the radial basis function, MLP and HGNN methods on the test data set samples. DETAILED DESCRIPTION

[0052] The embodiments of the present invention are described in detail below in combination with the technical solutions and the accompanying drawings.

[0053] The structural performance prediction method of the present invention has the following process: Figure 1 The embodiment is shown in Figure 2 Taking the three types of vehicle body frames with plates as examples, the specific implementation process of this method is explained in conjunction with the accompanying drawings.

[0054] 1. Establish a heterogeneous graph dataset corresponding to three topological types of vehicle body frames.

[0055] The three body frames with plates are in the same torsional stiffness analysis condition. Specifically, a concentrated force F = 1980N of equal magnitude and opposite direction is applied at the front suspension support point, while all degrees of freedom are constrained at the rear suspension support point. The body frame consists of a thin-walled rectangular beam and a plate of equal thickness. The beam is characterized by four parameters: a, b, t1 and t2, while the plate is characterized by only one thickness parameter T. Here, we use the z-direction displacement at point V1 as a performance indicator of the body torsional stiffness. The larger the displacement value, the smaller the torsional stiffness.

[0056] In order to represent the body frame with plate as a heterogeneous graph, we first decompose it into two types of vertex components: beam components and plate components. Each beam component corresponds to a beam in the original structure, and each plate component corresponds to a plate in the original structure. The beam component is characterized by the two end point coordinates (x1, y1, z1), (x2, y2, z2) and cross-sectional parameters (a, b, t1, t2) of the beam, while the plate component is characterized by the four corner point coordinates (x1, y1, z1), (x2, y2, z2), (x3, y3, z3) and (x4, y4, z4) of the plate and the plate thickness dimension T.

[0057] Since the body frame with plates contains two types of vertex components after decomposition, namely beam components and plate components, the body frame with plates will contain four types of edge components after decomposition. They are edge components between vertex components of the same type (beam-beam and plate-plate), and edge components between vertex components of different types (beam-plate, beam component is source; plate-beam, plate component is source). When two beams share an endpoint, there is a beam-beam edge component. Similarly, when two plates share an edge, there is a plate-plate edge component. For edge components between vertex components of different types, when the endpoint of the beam is located on an edge of the plate, there are beam-plate and plate-beam edge components. In this embodiment, the edge component does not contain any features.

[0058] After decomposing the original structure, the vertex components and edge components are used as input to the next step to generate a heterogeneous graph representation corresponding to the original structure. We map each type of vertex component to a type of vertex in the heterogeneous graph, and use the feature vector of the vertex component as the feature of the corresponding vertex. Specifically, a beam component will be mapped to a beam vertex in the heterogeneous graph, and the feature vector of the beam component [x1, y1, z1, x2, y2, z2, a, b, t1, t2] will be used as the feature vector of the beam component. T As the feature of the vertex; a board component will be mapped to a board vertex in the heterogeneous graph, and the feature vector of the board component [x1,y1,z1,x2,y2,z2,x3,y3,z3,x4,y4,z4,T] TAs the feature of the vertex of the board. At the same time, each type of edge component is mapped to a type of edge in the heterogeneous graph to represent the topological relationship between different vertices in the heterogeneous graph.

[0059] After obtaining the heterogeneous graph representation of the body frame with plate, the finite element method is used to obtain the true value of the torsional stiffness of different body frames, thereby generating the labels of the heterogeneous graph samples. After that, different heterogeneous graph samples can be generated by changing the cross-sectional size parameters of the beam and the thickness parameters of the plate through the sampling method, completing the collection of the heterogeneous graph dataset. The final heterogeneous graph dataset includes three main components: the feature matrix of all vertex types, the adjacency matrix of all edge types, and the label matrix of all heterogeneous graphs. We collected 5,000 samples for each topology type, thus obtaining a body frame heterogeneous graph dataset containing 15,000 samples. The dataset was shuffled during the training process and divided into training set / validation set / test set in a ratio of 80%, 10%, and 10%.

[0060] 2. Define the HGNN design space.

[0061] The process of the HGT layer using the neighborhood aggregation mechanism and the multi-head attention mechanism to update the vertex features in the heterogeneous graph is as follows: Figure 3 According to the HGT layer, the intra-layer design dimension, inter-layer design dimension and training configuration design dimension of the HGNN design space are defined respectively, as shown in Table 1.

[0062] Table 1 HGNN design space

[0063]

[0064] 3. Construct the optimal HGNN model for the current body frame torsional stiffness prediction task.

[0065] The controlled random search method is used to search the intra-layer and inter-layer design dimensions; the grid search method is used to search the training configuration design dimensions, and the optimal parameters screened are shown in Table 2. The optimal HGNN model is constructed based on the optimal values ​​of the intra-layer and inter-layer design dimensions.

[0066] Table 2 The optimal parameters of HGNN screened out

[0067]

[0068] 4. The optimal HGNN model is trained using the entire heterogeneous graph dataset of the vehicle body frame and the above optimal training configuration parameters. The prediction accuracy iteration curve of the model on the validation set is shown in the figure below: Figure 4 As shown, the final prediction accuracy is 96.04%.

[0069] 5. The trained HGNN model is used to give a fast and accurate prediction of torsional stiffness for the heterogeneous graph samples in the test set. The prediction results of the radial basis function (RBF), MLP and HGNN methods for the test data set samples are compared as shown in the figure. Figure 5 As shown, the closer the distribution of the blocks is to the dotted line, the higher the prediction accuracy of the model and the better the performance.

Claims

1. A vehicle body structure performance prediction method based on heterogeneous graph neural network, characterized in that: The following steps are involved: Step 1: Establish a heterogeneous graph dataset corresponding to vehicle body structures of different topological types: Step 2: Define the HGNN design space; Step 3: HGNN design space search: Step 4: Use all graph datasets and the above optimal training configuration parameters to complete the training of the optimal HGNN model, and output the model checkpoint with the highest prediction accuracy on the validation set as the input for the next step; Step 5: Randomly sample new structure samples from the design domain to be predicted, and use the above model checkpoints to directly give the predicted value of its performance.

2. The method for predicting vehicle body structure performance based on heterogeneous graph neural network according to claim 1 is characterized in that: Step 1: Establish a heterogeneous graph dataset corresponding to vehicle body structures of different topological types, including: (1) Using sampling methods to collect the structural topology type required by the data set within the design domain to be predicted; (2) Using the sampling method to change the design parameters of the connection, the same number of structural samples are generated for each of the above structural topology types; (3) Decompose the structure sample into vertex components and edge components; a. Decompose the structure sample into vertex components of the same or different types; b. Determine edge components (of the same or different types) based on the topological relationship between different vertex components in the structure sample; c. Assigning feature vectors with the same dimension and different values ​​to each type of vertex component to distinguish different vertex components in each type of vertex component; d. Assign a feature vector with the same dimension and different values ​​to each type of edge component to distinguish different edge components in each type of edge component. This step is optional. (4) According to the vertex components and edge components obtained after decomposition of the structural sample, they are converted into heterogeneous graph data: a. Determine the vertices in the heterogeneous graph: map each type of vertex component to a type of vertex in the heterogeneous graph; map each vertex component to a vertex in the heterogeneous graph, and use the feature vector of the vertex component as the feature of its corresponding vertex; b. Determine the edges in the heterogeneous graph: map each type of edge component to a type of edge in the heterogeneous graph; map each edge component to an edge in the heterogeneous graph, and use the feature vector of the edge component as the feature of its corresponding edge; (5) Constructing a heterogeneous graph dataset: a. Taking the heterogeneous graph data corresponding to the structural sample as the heterogeneous graph sample in the data set; b. Use numerical analysis methods to solve the true performance value of each structural sample as the label of the heterogeneous graph sample.

3. The method for predicting vehicle body structure performance based on heterogeneous graph neural network according to claim 1 is characterized in that: Step 2: Define the HGNN design space; including: (1) HGT layer: Its computation process includes three main components: heterogeneous mutual attention, heterogeneous message passing, and heterogeneous message aggregation for specific targets; a. Heterogeneous mutual attention: Among them, s, t and e represent the source vertex, the target vertex and the edge between the two vertices respectively; H (l-1) [s] and H (l-1) [t] represents the features of the source vertex and the target vertex before the lth layer, respectively; h represents the number of attention heads, d represents the common feature dimension of all vertices in the heterogeneous graph after being projected into the same space, and μ represents the prior tensor encoding the general meaning of each meta-relation; and The weight matrices corresponding to edge type, source vertex type, and target vertex type respectively; b. Heterogeneous message passing: in, and Represent the weight matrices associated with the source vertex type and edge type respectively; c. Aggregation of heterogeneous messages for specific targets: in, Represents the updated vertex features of the HGT layer; (2) Define the intra-layer, inter-layer, and training configuration design dimensions: Define the intra-layer design dimensions based on the variable operations of the above HGT layers and their optional values; define the inter-layer design dimensions based on the variable operations between HGT layers and their optional values; define the training configuration design dimensions based on the training hyperparameters of the network model and their optional values.

4. The method for predicting vehicle body structure performance based on heterogeneous graph neural network according to claim 1 is characterized in that: Step 3: HGNN design space search, including: (1) Use controlled random search method to search for design dimensions within and between layers: a. Select the i-th design dimension dim (i) (i=1, 2, ..., I), I is the total number of intra-layer and inter-layer design dimensions; i) Select the first optional value of the design dimension Randomly sample N satisfying Design combination composition collection N is the number of sampled design combinations; use these N design combinations to build HGNN models respectively, and then use some graph datasets and common values ​​of training configuration design dimensions to complete the training, and record the highest prediction accuracy of each model on the validation set As the final score of this design portfolio; ii) Select the next optional value for the design dimension J i is the number of optional values ​​for the current design dimension; the following changes are made to the values ​​of the i-th design dimension of all N design combinations in i). At the same time, keep the values ​​of other design dimensions unchanged, so as to generate N new design combinations to form a set Use these N new design combinations to build HGNN models respectively, and then use some graph datasets and common values ​​of training configuration design dimensions to complete training, and record the highest prediction accuracy of each model on the validation set. As the final score of this design portfolio; iii) If there are still optional values not retrieved for this design dimension, return to ii); otherwise, proceed to the next step; iv) From the collection Take out all nth (n=1,2,..,N) design combinations And according to their final score Sort by number; the design combination with the highest score is labeled 1, the design combination with the second highest score is labeled 2, and so on; v) If n < N method, set n = n + 1 and return to iv); otherwise, proceed to the next step; vi) Summarize all optional values ​​of the above N groups for the current design dimension The design dimension dim is determined by analyzing the mean and distribution of the sorting. (i) The optimal design choice optimal_val (i) ; The smaller the average value of the sort and the more concentrated the distribution, the better the design choice; b. If there are still design dimensions in the HGNN design space that have not been retrieved, set i = i + 1 and return to a); otherwise, proceed to the next step; c. Summarize the optimal design choices for all the above intra-layer and inter-layer design dimensions optimal_val (i) (i=1, 2, ..., I), as the input for the next step; (2) Construct the optimal HGNN model based on the optimal design selection of the intra-layer and inter-layer design dimensions, use the grid search method to search for the training configuration design dimensions, and output the optimal training configuration parameters.

Images