Vehicle body structure performance prediction method based on graph neural network

Abstract

The invention discloses a vehicle body structure performance prediction method based on a graph neural network, and belongs to the technical field of engineering structure analysis and optimization. The method is characterized by comprising the following steps: collecting structure samples with different topology types and design parameters in a to-be-predicted design domain; converting the structure sample into graph data by using a graph representation method; solving by using a numerical analysis method to obtain a performance true value of the structure sample as a label of the graph sample; using the graph data set to screen out optimal structure parameters and training configuration parameters of the GNN model in the GNN design space; and an optimal GNN model is constructed, and is used for predicting the performance of any structure in a design domain after being trained. According to the method, the defect that a single model in common network types can only predict the performance of one topology type structure can be overcome, so that the performance of all topology type structures in a design domain can be predicted by using the single model, good expandability is achieved, the number of the models can be reduced, the training time can be shortened, and the product development period is shortened.

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 graph neural network. Background Art

[0002] With the revival of artificial intelligence, its representative deep learning technology has not only surpassed the highest level of humans in many computer-related fields (image classification, computer Go, etc.), but also played an important role in more and more traditional research fields, including the performance prediction of vehicle body structures. At present, most deep learning-based vehicle body structure performance prediction methods use one-dimensional vectors or two-dimensional images to represent structural objects. Typically, multi-layer perceptrons (MLPs) or convolutional neural networks (CNNs) are used as universal function approximators to predict the target performance of structures. With the support of the powerful representation capabilities of deep neural networks and the large amount of data generated by numerical analysis methods or engineering experiments, these deep learning-based methods can easily achieve higher approximation accuracy than traditional surrogate models (such as response surface methods, etc.).

[0003] However, MLP and CNN encounter a critical scalability issue when applied to vehicle body structure performance prediction tasks. Specifically, models based on MLP or CNN are limited to structures of a specific topology type and fail immediately once applied to other topologies. This defect greatly limits the scope of application of the model and causes a waste of computing resources. Therefore, it is necessary to establish a performance prediction method with good scalability for structures of different topologies, so that a single model can be used to complete the performance prediction tasks of multiple topology types. 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 with good scalability by representing the vehicle body structure as graph data and introducing graph neural network.

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

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

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

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

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

[0010] 3) Decompose the structure sample into two types of components, namely vertex components and edge components;

[0011] a) assigning feature vectors of the same dimension and different values ​​to all vertex components to distinguish different vertex components;

[0012] b) Assign feature vectors of the same dimension but different values ​​to all edge components to distinguish different edge components.

[0013] 4) According to the components obtained after decomposing the above structural samples, convert them into graph data:

[0014] a. Determine the vertices in the graph: map each vertex component to a vertex in the graph, and use the feature vector of the vertex component as the feature of its corresponding vertex;

[0015] b. Determine the edges in the graph: map each edge component to an edge in the graph, and use the feature vector of the edge component as the feature of its corresponding edge.

[0016] 5) Build a graph dataset:

[0017] a. Use the graph data corresponding to the structure sample as the graph sample in the data set;

[0018] b. Use numerical analysis methods to solve the true performance value of each structural sample as the label of the graph sample.

[0019] Step 2: Define the GNN design space

[0020] 1) Define the general GNN layer: The process of updating the vertex features in the graph using the general GNN layer can be divided into three sub-processes:

[0021] a. Message construction process:

[0022]

[0023] Among them, ρ (l) Represents the message construction function of layer l, and are the features of the central vertex, neighboring vertex, and the connecting edge between them, DROPOUT, ACT, and BN represent Dropout, activation function, and batch normalization operations with multiple optional values, respectively. MLP is a 2-layer fully connected network, and Represents the first-order neighborhood of the central vertex v;

[0024] b. Message aggregation process:

[0025] Among them, (l) represents the message aggregation function of layer l, AGG represents an aggregation function with multiple optional values;

[0026] c. Vertex update process:

[0027] Among them, φ (l) Represents the vertex update function of layer l.

[0028] 2) Define intra-layer, inter-layer and training configuration design dimensions: Define intra-layer design dimensions based on the variable operations of the above general GNN layers and their optional values; define inter-layer design dimensions based on the variable operations between GNN layers and their optional values; define training configuration design dimensions based on the training hyperparameters of the network model and their optional values.

[0029] Step 3: GNN Design Space Search

[0030] 1) Use controlled random search method to search for design dimensions within and between layers:

[0031] 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;

[0032] i) Select the first optional value of the design dimension Randomly sample N satisfying Design combination, forming a collection N is the number of sampled design combinations; use N design combinations to build GNN 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;

[0033] ii) Select the next optional value for the design dimension Ji is the number of optional values ​​of the current design dimension; the values ​​of the i-th design dimension of all design combinations in i) are modified as follows 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 design combinations to build GNN 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;

[0034] iii) If there are still optional values ​​that have not been obtained for the design dimension, return to ii); otherwise, execute the next step;

[0035] iv) From the collection Take out all nth (n=1,2,...,N) design combinations and sort them according to their final scores The design combination with the highest score is marked as 1, the second-highest design combination is marked as 2, and so on.

[0036] v) If n < N methods, then let n = n + 1 and return to iv); otherwise, execute the next step;

[0037] 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)

[0038] b) If there are still unselected design dimensions in the GNN design space, then let i = i + 1 and return to a); otherwise, execute the next step;

[0039] 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.

[0040] 2) Construct an optimal GNN 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.

[0041] Step 4: Use all graph datasets and the above optimal training configuration parameters to complete the training of the optimal GNN model, and output the model checkpoint with the highest prediction accuracy on the validation set as the input for the next step.

[0042] Step 5: Randomly sample new structural samples from the design domain to be predicted, and directly give the predicted values of their performance using the above model checkpoint.

[0043] By introducing graph data and graph neural networks, the present invention overcomes the problem of poor scalability of previous vehicle body structure performance prediction methods, and can complete the performance prediction tasks of various topological type structures using a single model; through the definition and search of the GNN design space, the optimal GNN model and training configuration parameters for the current performance prediction task are found, overcoming the sub-optimality of the models in existing methods and improving the prediction accuracy of the model; the GNN model used has fewer parameters and faster inference efficiency, which can accelerate the development process of topological optimization and product family design. Description of the Drawings

[0044] Figure 1 It is a flow chart of steps S101 to S105 of a structural performance prediction method based on a graph neural network of the present invention;

[0045] Figure 2 Description of the body frame and analysis conditions for the three conceptual design stages;

[0046] Figure 3 The process of updating vertex features in the graph using the neighborhood aggregation mechanism for the general GNN layer;

[0047] Figure 4 Results of the within-stratum and between-stratum design dimensionality search using the controlled random search method (visualized using bar graphs and violin plots);

[0048] Figure 5 It is the prediction accuracy iteration curve of the optimal GNN model on the validation set. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below in combination with the technical solutions and the accompanying drawings.

[0050] The present invention proposes a vehicle body structure performance prediction method based on graph neural network, and its process is as follows: Figure 1 As shown below. Figure 2 Taking the three car body frames in the conceptual design stage as examples, the specific implementation process of this method is explained in conjunction with the accompanying drawings.

[0051] 1. Establish a graph dataset corresponding to three topological types of vehicle body frames.

[0052] The three frames 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 is entirely composed of thin-walled rectangular beams, which are characterized by four parameters: a, b, T and 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.

[0053] In order to represent the vehicle body frame as a graph, we first decompose it into vertex components and edge components. The vertex components correspond to the connection points of the two beams in the original structure and can be represented by the three-dimensional coordinates (x, y, z) of the point; while the edge components correspond to the beams in the original structure and can be represented by the cross-sectional parameters (a, b, T, t) of the beams.

[0054] After completing the decomposition of the original structure, the beam components and edge components are used as input to the next step to generate the final graph representation corresponding to the original structure. Specifically, we map the vertex components to vertices in the graph and map the feature vectors [x, y, z] of the vertex components to the vertices in the graph. TAs the feature of the vertex; at the same time, the edge component is mapped to the edge in the graph, and the feature vector of the edge component [a, b, T, t] T as the feature of the edge.

[0055] After obtaining the graph representation of the body frame structure, the finite element method is used to extract the true value of the torsional stiffness of different body frames, thereby generating the labels of the graph samples. By changing the cross-sectional size parameters of the beam through the sampling method, different graph samples can be generated to complete the collection of the graph dataset. The final graph dataset includes 4 main components: the adjacency matrix of all graph samples, the feature matrix of all vertices, the feature matrix of all edges, and the label matrix of all graphs. We collected 5,000 samples for each topology type, resulting in a body frame graph dataset containing 15,000 samples. The dataset was shuffled during training and divided into training set / validation set / test set in a ratio of 80%, 10%, and 10%.

[0056] 2. Define the GNN design space.

[0057] The process of updating vertex features in the graph using the neighborhood aggregation mechanism is as follows: Figure 3 According to the general GNN layer, the intra-layer design dimension, inter-layer design dimension and training configuration design dimension of the GNN design space are defined respectively, as shown in Table 1.

[0058] Table 1 GNN design space

[0059]

[0060] 3. Construct the optimal GNN model for the current body frame torsional stiffness prediction task.

[0061] The search of design dimensions within and between layers is performed using a controlled random search method, and visualization is performed using bar charts and violin plots, such as Figure 4 The grid search method is used to search for the training configuration design dimensions; the optimal parameters obtained by screening are shown in Table 2. The optimal GNN model is constructed based on the optimal values ​​of the intra-layer and inter-layer design dimensions.

[0062] Table 2 Optimal parameters at screening

[0063]

[0064]

[0065] 4. Use the entire image dataset of the vehicle body frame and the above optimal training configuration parameters to complete the training of the optimal GNN model. The prediction accuracy iteration curve of the model on the validation set is as follows: Figure 5 As shown, the final prediction accuracy is 97.87%.

[0066] 5. Use the trained GNN model to give fast and accurate torsional stiffness predictions for the image samples in the test set.

Claims

1. A method for predicting the performance of a vehicle body structure based on a graph neural network, comprising the following steps: Step 1: Establish a graph data set corresponding to vehicle body structures of different topological types Step 2: Define the GNN design space Step 3: Search the GNN design space Step 4: Use the entire graph data set and the above optimal training configuration parameters to complete the training of the optimal GNN 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 structural samples from the design domain to be predicted, and directly give the predicted values of their performance using the above model checkpoint.

2. The method for predicting the performance of a vehicle body structure based on a graph neural network according to claim 1, wherein: The establishment of the graph data set corresponding to vehicle body structures of different topological types in Step 1 includes: 1) Use a sampling method in the design domain to be predicted to collect the structural topological types required for the data set; 2) Use a sampling method to change the design parameters of the structure, and generate the same number of structural samples for each of the above structural topological types; 3) Decompose the structural samples into two types of components, namely vertex components and edge components; a. Assign feature vectors with the same dimension and different values to all vertex components to distinguish different vertex components; b. Assign feature vectors with the same dimension and different values to all edge components to distinguish different edge components; 4) According to the components obtained after the above decomposition of the structural samples, convert them into graph data: a. Determine the vertices in the graph: Map each vertex component to a vertex in the graph, and use the feature vector of the vertex component as the feature of its corresponding vertex; b. Determine the edges in the graph: Map each edge component to an edge in the graph, and use the feature vector of the edge component as the feature of its corresponding edge; 5) Construct the graph data set: a. Use the graph data corresponding to the structural samples as the graph samples in the data set; b. Use numerical analysis methods to solve the true performance values of each structural sample as the labels of the graph samples.

3. The method for predicting the performance of a vehicle body structure based on a graph neural network according to claim 1, wherein: Step 2: Define the GNN design space, including: 1) Define a general GNN layer: Split the process of updating the vertex features in the graph using the general GNN layer into 3 sub-processes: a. Message construction process: Among them, ρ (l) Represents the message construction function of layer l, and are the features of the central vertex, neighboring vertex, and the connecting edge between them, DROPOUT, ACT, and BN represent Dropout, activation function, and batch normalization operations with multiple optional values, respectively. MLP is a 2-layer fully connected network, and Represents the first-order neighborhood of the central vertex v; b. Message aggregation process: Among them, (l) represents the message aggregation function of layer l, AGG represents an aggregation function with multiple optional values; c. Vertex update process: Among them, φ (l) Represents the vertex update function of layer l; 2) Define the in-layer, inter-layer, and training configuration design dimensions: Define the in-layer design dimension based on the variable operations and their optional values of the above general GNN layer; Define the inter-layer design dimension based on the variable operations and their optional values between GNN layers; Define the training configuration design dimension based on the training hyperparameters of the network model and their optional values.

4. The method for predicting the performance of a vehicle body structure based on a graph neural network according to claim 1, wherein: Step 3: Search the GNN design space, including: 1) Use a controlled random search method to search the in-layer and inter-layer design dimensions: 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, forming a collection N is the number of sampled design combinations; use N design combinations to build GNN 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; ii) Select the next optional value for the design dimension J i is the number of optional values ​​for the current design dimension; the values ​​of the i-th design dimension of all design combinations in i) are modified as follows 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 design combinations to build GNN 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 in this design dimension that have not been taken, 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, then 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 unobtained design dimensions in the GNN design space, 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) Based on the optimal design selection of intra-layer and inter-layer design dimensions, the optimal GNN model is constructed, the grid search method is used to search the training configuration design dimensions, and the optimal training configuration parameters are output.

Images