Stress field rapid calculation method based on graph network agent model

Abstract

The invention discloses a rapid stress field calculation method based on a graph network proxy model. The method comprises the following steps: firstly, creating a simulation model and constructing a model training data set; then exporting grid information; a graph neural network method based on message aggregation, transmission and updating is adopted, parameters such as the centroid and the volume of a unit are introduced to serve as node features, the area serves as edge features, graph construction is completed, and compared with a traditional agent model creating method, grid information is fully considered, and calculation precision is improved; by introducing global features and utilizing external environment parameters or partial measuring point stress values, quick and real-time calculation of the stress field of the structural body can be completed, the sensor layout requirement and the hardware cost are reduced, the stress distribution of the structure can be accurately predicted, and the calculation efficiency can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of stress state measurement, and specifically provides a stress field fast calculation method based on a graph network agent model. Background Art

[0002] With the complexity of engineering problems, it is of great significance to quickly calculate the stress state (stress field) of structural parts such as bearings according to environmental conditions such as external loads, so as to analyze their reliability and fatigue. In traditional methods, finite element analysis (FEM) simulation method and actual stress sensor detection method are generally used to calculate stress state, but the finite element analysis (FEM) simulation method has a large amount of calculation, takes a long time, and does not have real-time performance; the actual stress sensor detection method is difficult to calculate the entire stress field of the structure due to the limited number of measurement points and installation conditions. Therefore, it is necessary to develop an alternative method for stress field calculation that only requires a small amount of calculation but maintains sufficient accuracy.

[0003] Existing surrogate models include Kriging model (KRG), response surface model (PRS) and radial basis model (RBF). These methods approximate the simulation process by fitting the objective function, thereby achieving efficient calculation. However, when dealing with large-scale data sets, nonlinear relationships or high-dimensional problems, they usually face problems such as high computational complexity and high risk of overfitting. Specifically, although the Kriging method can provide prediction uncertainty, its training and inference process is relatively slow when the amount of data is large; the response surface model performs poorly when dealing with nonlinear relationships, and may cause overfitting as the dimension increases; the radial basis function network, due to its local limitations, often has insufficient generalization ability when dealing with large-scale problems. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a fast calculation method of stress field based on a graph network agent model, which can not only accurately predict the stress distribution of the structure, but also improve the calculation efficiency.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A fast calculation method of stress field based on graph network agent model includes the following steps:

[0007] Step 1: Build a simulation model and create a data set

[0008] Combine the physical problems of the target stress field with the actual environmental conditions to establish a finite element simulation model of the structural parts; batch train and export the simulation results as a data set through automated program scripts;

[0009] Step 2: Obtain grid information based on the simulation model

[0010] By constructing the obtained simulation model, the mesh information is processed and exported in a common format; the mesh information includes node numbers, node coordinates, unit centroid coordinates and unit topology structure information;

[0011] Step 3: Use the grid information of the simulation model to build a graph network structure

[0012] The graph network structure is defined as:

[0013] G=(V,E)

[0014] Where: V is a set of objects called nodes, defining the centroids of all grid cells as graph nodes; E is a set of pairs of vertices called edges, defining that there is a graph edge between grid cells with the same face;

[0015] The graph network structure includes graph node features and graph edge features. Global features are introduced into the graph network structure to enhance feature expression ability and stability.

[0016] Step 4: Train the graph network proxy model

[0017] A graph network proxy model is trained using a data set through an encoder-processor-decoder architecture; the encoder is used to embed node attribute vectors in the graph network structure into a high-dimensional vector space and to generate edge encoding features through non-learning transformation of edge features; the processor updates the hidden state of the node through a recursive message passing mechanism; the decoder is used to predict the paradigm equivalent stress of each node;

[0018] Step 5: Use the proxy model to quickly calculate the stress field according to environmental conditions

[0019] By changing the value of the incoming global feature, the stress values ​​of all mesh nodes are predicted through the graph network proxy model.

[0020] Further, the step 1 includes the following steps:

[0021] 11) Combine the physical problems of the target stress field and the actual environmental conditions to establish a finite element simulation model of the structural parts;

[0022] 12) Mesh the finite element simulation model to ensure the accuracy and efficiency of the calculation results;

[0023] 13) Apply different mechanical loads and boundary conditions to the finite element simulation model to generate simulation data of stress fields under various working conditions; use the measured data of stress measurement points arranged on the structural parts to evaluate the simulation data: adjust the boundary conditions to optimize the finite element simulation model and reduce the difference between the simulation data and the measured data until the simulation data meets the accuracy requirements;

[0024] 14) Batch training is performed through automated program scripts and simulation results are exported as data sets, and the data sets are divided into training set, validation set and test set; the training set is used to train the graph neural network model, the test set is used to evaluate the generalization performance of the graph neural network model, and the validation set is used to improve hyperparameters and evaluate the graph neural network model to prevent overfitting.

[0025] Furthermore, in step 3, two vertices connected by an edge become neighbors, and the adjacency matrix is ​​used express:

[0026]

[0027] Among them: A ij represents the adjacency matrix; v i and v j Represented by edge e i,j Connect two vertices; represents the set of real numbers; n×n represents the dimension.

[0028] Furthermore, in step 3, the method for constructing the graph node feature is: defining the centroid of all grid cells as the graph node v i ∈V; define the attributes of the graph nodes as represents the set of real numbers; 1×d v Represents the dimension of node attributes; the node attribute vector is expressed as:

[0029] h v =[c,t,v vol ]

[0030] Where: h v is the node attribute vector;

[0031] c is the geometric attribute of the graph node, and c = [x e ,y e ,z e ], x e ,y e and z e are the coordinate values ​​of the centroid of the grid cell in the three-dimensional rectangular coordinate system Oxyz; t is the graph node type, and one-hot encoding is used to represent the grid cell type: if there is a surface node in the grid cell, this cell is also defined as a surface grid cell, and the graph node attribute t is encoded as t = [0, 1]; otherwise, it is an internal grid cell, encoded as t = [1, 0]; v vol It is a volume feature attribute that calculates the volume of each grid cell based on the grid nodes.

[0032] Furthermore, in step 3, the method for constructing the graph edge feature is: define the edge as e i,j∈E, where i and j are two connected nodes, and E represents the set of all edges; the graph edge attribute vector is defined as represents the set of real numbers; 1×d e represents the dimension of node attributes; the edge attribute vector is expressed as:

[0033]

[0034] in: is the edge attribute vector; A e is the area of ​​the contact surface between the mesh units, defined as the sum of the areas of the two contact surface units; Δx, Δy, and Δz are the distances between the centroids of the two mesh units, and:

[0035] Δx=x j ―x i , Δy=y j ―y i , Δz=z j ―z i

[0036] Where: x i ,y i and z i is the centroid coordinate of the i-th grid cell; x j ,y j and z j is the centroid coordinate of the jth grid cell.

[0037] Furthermore, in step 3, global features are introduced into the graph network structure, which is expressed as:

[0038]

[0039] Among them: Q is the key global feature, and represents the set of real numbers; 1×(d v +d Q ) represents the dimension of global features; global features include external loads [f x ,f y ,f z ]、Measurement point position [x f ,y f ,z f ] and the stress signal of the measuring point [∈ ef1 ,∈ ef2 ...∈ efn ];

[0040] In each layer of the graph network structure, the global feature Q is directly concatenated to the attribute vector of each node On the top, we get the enhanced node attribute vector:

[0041]

[0042] in: is the enhanced node attribute vector; || represents the concatenation operation.

[0043] Furthermore, in step 4, the encoder is used to embed the node attribute vector in the graph network structure into a high-dimensional vector space to obtain:

[0044]

[0045] Among them: MLP en-v is the multi-layer perceptron in the node encoder; represents the hidden state of the node at step t = 0; Represents the input vector of the node The dimension is 1×(d v +d θ ); represents the set of real numbers; d v Indicates the feature dimension of the node; d θ Represents other feature dimensions related to the node.

[0046] The encoder is used to generate edge encoding features through non-learning transformation of edge features, and obtain:

[0047]

[0048] Among them: MLP en―e is the multi-layer perceptron in the edge encoder; represents the hidden state of the edge at step t = 0; Represents an edge attribute vector.

[0049] Furthermore, in step 4, the method in which the processor updates the hidden state of the node through a recursive message passing mechanism is:

[0050] Calculate each node v i Neighbor nodes The aggregate information transmitted defines the message aggregation function as follows:

[0051]

[0052] in: is node v i In the hidden state of step t, it describes the feature information of the current node; is the neighbor node v j The hidden state of is used to provide contextual information; ij is the edge feature, indicating that node v i and v j The relationship betweenM is the message passing function, defined as a multilayer perceptron:

[0053]

[0054] Introduce attention mechanism to improve message selectivity and model expressiveness:

[0055]

[0056] Among them: ‖ represents the vector concatenation operation; α ij is the attention weight between nodes, and:

[0057]

[0058] in: Represents the attention score function, which is used to calculate the relevance score between nodes.

[0059] The hidden state of the node will be based on the received message To update, residual connection and layer normalization are introduced in the state update process, and the state update formula is:

[0060]

[0061] Among them: LayerNorm represents layer normalization; Φ U It is the state update function.

[0062] Furthermore, in step 4, the decoder is used to predict the paradigm equivalent stress of each node, which is expressed as:

[0063]

[0064] Among them: MLP de The MLP of the node decoder is used to map the processed features to the output features, completing the conversion process of the encoder-processor-decoder model; is the hidden state of the node at step T; is the processed vertex set.

[0065] The beneficial effects of the present invention are:

[0066] The invention is based on the stress field fast calculation method of the graph network proxy model. First, a high-quality simulation model is created and a model training data set is constructed; then high-quality grid information is derived; a graph neural network method based on message aggregation, transmission and update is adopted, and unit centroid, volume and other parameters are introduced as node features, and area is used as edge features to complete the construction of the graph. Compared with the traditional method of creating a proxy model, the grid information is fully considered and the calculation accuracy is improved; by introducing global features and using external environmental parameters, or stress values ​​of some measuring points, the fast and real-time calculation of the stress field of the structure can be completed, the sensor layout requirements and hardware costs can be reduced, and the calculation efficiency can be improved. In addition, the method of the invention also has broad potential in the field of scientific research, and can solve scientific problems involving high computational or time-consuming tasks, such as materials science, structural optimization and fault prediction, etc. It can also be used for the fast calculation of physical fields such as temperature fields. The method of the invention has important engineering value and promotion prospects, and promotes the development of digital engineering technology.

[0067] The present invention is based on a fast calculation method of stress field based on a graph network proxy model. It constructs a new type of graph signal and introduces global features as node attributes to capture geometry and boundary conditions. At the same time, it uses the attention mechanism method to overcome the common over-smoothing problem in graph deep learning. By using this method of the present invention, not only can the stress distribution of the structure be accurately predicted, but the speed can be hundreds of times faster than the simulator based on finite element (FEA), which provides the possibility for real-time response analysis, calculation, inversion, and structural state monitoring and prediction. In other research fields, this method can also become a powerful tool for solving other complex physical problems, such as temperature field calculation, etc., to promote scientific development and improve engineering productivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0069] Figure 1 It is a flow chart of the stress field rapid calculation method based on the graph network agent model of the present invention;

[0070] Figure 2 Schematic diagram of the graph network agent model. DETAILED DESCRIPTION

[0071] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0072] The stress field fast calculation method based on the graph network proxy model of this embodiment uses the simulation model and the code to batch simulate the stress field distribution of structural parts under different environmental conditions to build a training data set; and uses the stress signal monitored by the stress sensor to evaluate and optimize the simulation data and model; builds the basic graph structure based on the grid information in the simulation model; uses the message passing and information exchange mechanism between adjacent vertices to train the graph network proxy model; finally, based on the actual environmental information and a small amount of stress measurement point information, uses the proxy model to realize the fast calculation of the stress field. Specifically, Figure 1 As shown, the stress field fast calculation method based on the graph network agent model of this embodiment includes the following steps.

[0073] Step 1: Build a simulation model and create a data set

[0074] Combined with the physical problems of the target stress field and the actual environmental conditions, a finite element simulation model of the structural parts is established; the simulation results are batch trained and exported as a data set through automated program scripts.

[0075] Specifically, in this embodiment, constructing the simulation model includes the following steps.

[0076] 11) In the simulation model construction stage, the finite element analysis (FEA) software is used to establish a high-precision finite element simulation model of the structural parts in combination with the physical problems of the target stress field and the actual environmental conditions.

[0077] 12) By analyzing the stress distribution characteristics and geometric shapes, a suitable meshing strategy is selected to mesh the finite element simulation model to ensure the accuracy and efficiency of the calculation results. In this embodiment, the mesh generation uses high-quality tetrahedral or hexahedral units, and local mesh encryption rules in key areas are defined to ensure the ability to capture stress gradient change areas.

[0078] 13) After constructing a finite element simulation model that can accurately simulate the target stress field, apply different mechanical loads and boundary conditions to the finite element simulation model to generate simulation data of the stress field under various working conditions. Use the measured data of the stress measurement points arranged on the structural parts to evaluate the simulation data: continuously adjust the boundary conditions to optimize the finite element simulation model and reduce the difference between the simulation data and the measured data until the simulation data meets the set accuracy requirements.

[0079] 14) Through automated program scripts, batch training and export of simulation results as data sets are performed, and the data sets are divided into training set, validation set, and test set; the training set is used to train the graph neural network model, the test set is used to evaluate the generalization performance of the graph neural network model, and the validation set is used to improve hyperparameters and evaluate the graph neural network model to prevent overfitting. This data set will provide strong support for subsequent efficient stress field simulation.

[0080] Step 2: Obtain grid information based on the simulation model

[0081] By constructing the obtained simulation model, the mesh information is processed and exported in a common format (txt or csv). Specifically, the mesh information includes the node number i = (1, 2...N), the node coordinates C = (x, y, z), the unit centroid coordinates c = (x e ,y e ,z e ) and unit topology information M = [m1, m2, ..., m N ], where the node and centroid coordinates are in the three-dimensional rectangular coordinate system Oxyz; M is the node number value that constitutes all grid units, m i =[i1,i2,i3...]∈M. This mesh information will serve as the basis for graph data construction in subsequent steps. The correctness of the mesh quality and boundary condition settings is verified through simulation runs to ensure that the simulation model can accurately reflect the actual working conditions and lay a solid foundation for subsequent data set generation and proxy model training. The careful design and execution of this process is crucial to the construction and performance of the subsequent graph network model.

[0082] Step 3: Use the grid information of the simulation model to build a graph network structure

[0083] In this embodiment, the graph network structure is defined as:

[0084] G=(V,E)

[0085] Where: V is a set of objects called nodes, and the centroid of all grid cells is defined as the graph node v i ∈V; E is a set of pairs of vertices, called edges, defining a graph edge e between grid cells with the same face i,j ∈E.

[0086] Specifically, two edges e i,j Connected vertices v i and v j Become neighbors and use the adjacency matrix express:

[0087]

[0088] Among them: A ij represents the adjacency matrix; v i and v j Represented by edge e i,j Connect two vertices; represents the set of real numbers; n×n represents the dimension.

[0089] In this embodiment, the graph network structure includes graph node features and graph edge features. In the process of constructing the graph network structure, it is necessary to construct the graph node features and the graph edge features respectively.

[0090] (1) Constructing graph node features

[0091] In this embodiment, the method for constructing the graph node feature is: define the centroid of all grid cells as the graph node v i ∈V; define the attributes of the graph nodes as represents the set of real numbers; 1×d v Represents the dimension of node attributes; the node attribute vector is expressed as:

[0092] h v =[c,t,v vol ]

[0093] Where: h v is the node attribute vector; c is the geometric attribute of the graph node, and c = [x e ,y e ,z e ], x e ,y e and z e are the coordinate values ​​of the centroid of the grid cell in the three-dimensional rectangular coordinate system Oxyz; t is the graph node type, and one-hot encoding is used to represent the grid cell type: if there is a surface node in the grid cell, this cell is also defined as a surface grid cell, and the graph node attribute t is encoded as t = [0, 1]; otherwise, it is an internal grid cell, encoded as t = [1, 0]; v vol It is a volume feature attribute that calculates the volume of each grid cell based on the grid nodes.

[0094] (2) Constructing graph edge features

[0095] In this embodiment, the method for constructing the graph edge feature is: define the edge as e i,j ∈E, where i and j are two connected nodes, and E represents the set of all edges; the graph edge attribute vector is defined as represents the set of real numbers; 1×d e represents the dimension of node attributes; the edge attribute vector is expressed as:

[0096]

[0097] in: is the edge attribute vector; A e is the area of ​​the contact surface between the mesh units, defined as the sum of the areas of the two contact surface units; Δx, Δy, and Δz are the distances between the centroids of the two mesh units, and:

[0098] Δx=x j ―x i , Δy=y j ―y i , Δz=z j ―z i

[0099] Where: x i ,y i and z i is the centroid coordinate of the i-th grid cell; x j ,y j and z j is the centroid coordinate of the jth grid cell.

[0100] (3) Introducing global features

[0101] In this embodiment, the graph network structure will also be improved, and global features will be introduced into the graph network structure to enhance feature expression ability and stability. By combining global features with vertex or edge features, each layer of the network can perceive the global information of the entire graph, making training more stable, alleviating gradient vanishing and gradient explosion, and improving prediction performance.

[0102] Specifically, global features are introduced into the graph network structure, expressed as:

[0103]

[0104] Among them: Q is the key global feature, and represents the set of real numbers; 1×(d v +d Q ) represents the dimension of global features; global features include external loads [f x ,f y ,f z ]、Measurement point position [x f ,y f ,z f ] and the stress signal of the measuring point [∈ ef1 ,∈ ef2 ...∈ efn ];

[0105] In this embodiment, in order to simplify the model, in each layer of the graph network structure, the global feature Q is directly concatenated to the attribute vector of each node On the top, we get the enhanced node attribute vector:

[0106]

[0107] in: is the enhanced node attribute vector; || represents the concatenation operation, so that each node can directly refer to the global feature Q without having to obtain global information through edge propagation.

[0108] Step 4: Train the graph network proxy model

[0109] In this embodiment, a data set is used to train a graph network proxy model through an encoder-processor-decoder architecture. Specifically, Figure 2 As shown, the encoder is used to embed the node attribute vector in the graph network structure into the high-dimensional vector space, and to generate edge encoding features through non-learning transformation of the edge features; the processor updates the hidden state of the node through a recursive message passing mechanism; the decoder is used to predict the paradigm equivalent stress of each node. In this embodiment, the capabilities of graph neural networks (GNNs) are used to process grid-based data. The model uses the Adam optimizer and the Cosine Annealing learning rate scheduler, with a hidden layer size of 128 and an initial learning rate of 0.001. The architecture is divided into three key components, namely the encoder, processor, and decoder, each of which plays a different role in the process of converting the input features of the mesh vertices into output features. Specifically, the proxy model can be expressed as:

[0110]

[0111] in: is the prediction result; Encoder represents the encoder; Processor represents the processor; Decoder represents the decoder.

[0112] (1) Encoder

[0113] In this embodiment, the encoder is used to embed the node attribute vector in the graph network structure into a high-dimensional vector space to obtain:

[0114]

[0115] Among them: MLP en-v is the multi-layer perceptron in the node encoder; represents the hidden state of the node at step t = 0; Represents the input vector of the node The dimension is 1×(d v +d θ ); represents the set of real numbers; d v Indicates the feature dimension of the node; d θ Represents other feature dimensions related to the node.

[0116] For edge features, the encoder is used to generate edge encoding features through non-learning transformation of edge features, and obtain:

[0117]

[0118] Among them: MLP en―e is the multi-layer perceptron in the edge encoder; represents the hidden state of the edge at step t = 0; } represents the edge attribute vector.

[0119] The encoder module implements the function of enhancing the expressiveness of nodes.

[0120] (2) Processor

[0121] The processor is the core component of the model architecture and updates the hidden state of the node through a recursive message passing mechanism. To improve the performance of the model, this embodiment adopts an improved graph message passing mechanism, which can optimize the convergence speed and significantly reduce the error accumulation. The processor gradually updates the hidden state of each node within t = 1, 2, ..., T time steps. The message passing and status update mechanism is described in detail below.

[0122] In this embodiment, the method in which the processor updates the hidden state of a node through a recursive message passing mechanism is as follows.

[0123] The core of message passing is to calculate each node v i Neighbor nodes The aggregate information transmitted defines the message aggregation function as follows:

[0124]

[0125] in: is node v i In the hidden state of step t, it describes the feature information of the current node; is the neighbor node v j The hidden state of is used to provide contextual information; ij is the edge feature, indicating that node v i and v j The relationship between M is the message passing function, defined as a multilayer perceptron:

[0126]

[0127] Introduce attention mechanism to improve message selectivity and model expressiveness:

[0128]

[0129] Among them: ‖ represents the vector concatenation operation; α ij is the attention weight between nodes, and:

[0130]

[0131] in: Represents the attention score function, which is used to calculate the relevance score between nodes.

[0132] After the message passing is completed, the hidden state of the node will be based on the received message To update, residual connection and layer normalization are introduced in the state update process, and the state update formula is:

[0133]

[0134] Where: Φ U is the state update function, defined as:

[0135]

[0136] In order to accelerate the convergence of the model and improve the stability, residual connection and layer normalization are introduced in the state update process. The final improved state update formula is:

[0137]

[0138] Among them: LayerNorm represents layer normalization.

[0139] (3) Decoder

[0140] The decoder is used to predict the paradigm equivalent stress of each node, expressed as:

[0141]

[0142] Among them: MLP de The MLP of the node decoder is used to map the processed features to the output features, completing the conversion process of the encoder-processor-decoder model; is the hidden state of the node at step T; is the processed vertex set.

[0143] Step 5: Use the proxy model to quickly calculate the stress field according to environmental conditions

[0144] By changing the value of the incoming global feature, the stress values ​​of all mesh nodes are predicted through the graph network proxy model.

[0145] After the model training is completed, by changing the incoming global features The stress values ​​of all grid nodes can be obtained by using the value of . If the global feature Q is defined as an external environmental parameter, or the stress value of some measuring points, the stress field can be quickly calculated by adjusting the model input Q. If the global feature Q is defined as the stress value of the measuring point, the real-time calculation of the stress field can be realized by using the real-time input of the stress signals of a small number of measuring points. At the same time, if combined with a three-dimensional visualization program, the display of a three-dimensional cloud map can also be realized.

[0146] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for rapid calculation of stress field based on graph network agent model, characterized by: The steps include: Step 1: Build a simulation model and create a data set Combine the physical problems of the target stress field with the actual environmental conditions to establish a finite element simulation model of the structural parts; batch train and export the simulation results as a data set through automated program scripts; Step 2: Obtain grid information based on the simulation model By constructing the obtained simulation model, the mesh information is processed and exported in a common format; the mesh information includes node numbers, node coordinates, unit centroid coordinates and unit topology structure information; Step 3: Use the grid information of the simulation model to build a graph network structure The graph network structure is defined as: G=(V,E) Where: V is a set of objects called nodes, defining the centroids of all grid cells as graph nodes; E is a set of pairs of vertices called edges, defining that there is a graph edge between grid cells with the same face; The graph network structure includes graph node features and graph edge features; global features are introduced into the graph network structure to enhance feature expression capability and stability; the global features are external environment features or stress values ​​of some measuring points; Step 4: Train the graph network proxy model A graph network proxy model is trained using a data set through an encoder-processor-decoder architecture; the encoder is used to embed node attribute vectors in the graph network structure into a high-dimensional vector space and to generate edge encoding features through non-learning transformation of edge features; the processor updates the hidden state of the node through a recursive message passing mechanism; the decoder is used to predict the paradigm equivalent stress of each node; Step 5: Use the proxy model to quickly calculate the stress field according to environmental conditions By changing the value of the global feature, the stress values ​​of all mesh nodes are predicted through the graph network proxy model.

2. The method for rapid calculation of stress field based on graph network agent model according to claim 1 is characterized in that: The step 1 includes the following steps: 11) Combine the physical problems of the target stress field and the actual environmental conditions to establish a finite element simulation model of the structural parts; 12) Mesh the finite element simulation model to ensure the accuracy and efficiency of the calculation results; 13) Apply different mechanical loads and boundary conditions to the finite element simulation model to generate simulation data of stress fields under various working conditions; Use measured data from stress measurement points on the structural parts to evaluate the simulation data: adjust boundary conditions to optimize the finite element simulation model and reduce the difference between the simulation data and the measured data until the simulation data meets the accuracy requirements; 14) Batch training is performed through automated program scripts and simulation results are exported as data sets, and the data sets are divided into training set, validation set and test set; the training set is used to train the graph neural network model, the test set is used to evaluate the generalization performance of the graph neural network model, and the validation set is used to improve hyperparameters and evaluate the graph neural network model to prevent overfitting.

3. The stress field rapid calculation method based on the graph network agent model according to claim 1 is characterized by: In step 3, two vertices connected by an edge become neighbors, and the adjacency matrix is ​​used express: Among them: A ij represents the adjacency matrix; v i and v j Represented by edge e i,j Connect two vertices; represents the set of real numbers; n×n represents the dimension.

4. The stress field rapid calculation method based on the graph network agent model according to claim 1 is characterized by: In step 3, the method for constructing the graph node feature is: define the centroid of all grid cells as the graph node v i ∈V; define the attributes of the graph nodes as represents the set of real numbers; 1×d v Represents the dimension of node attributes; the node attribute vector is expressed as: h v =[c,t,v vol ] Where: h v is the node attribute vector; c is the geometric attribute of the graph node, and c = [x e ,y e ,z e ], x e ,y e and z e are the coordinate values ​​of the centroid of the grid cell in the three-dimensional rectangular coordinate system Oxyz; t is the graph node type, and one-hot encoding is used to represent the grid cell type: if there is a surface node in the grid cell, this cell is also defined as a surface grid cell, and the graph node attribute t is encoded as t = [0, 1]; otherwise, it is an internal grid cell, encoded as t = [1, 0]; v vol It is a volume feature attribute that calculates the volume of each grid cell based on the grid nodes.

5. The stress field rapid calculation method based on the graph network agent model according to claim 1 is characterized by: In step 3, the method for constructing the graph edge feature is: define the edge as e i,j ∈E, where i and j are two connected nodes, and E represents the set of all edges; the graph edge attribute vector is defined as represents the set of real numbers; 1×d e represents the dimension of node attributes; the edge attribute vector is expressed as: in: is the edge attribute vector; A e is the area of ​​the contact surface between the mesh units, defined as the sum of the areas of the two contact surface units; Δx, Δy, and Δz are the distances between the centroids of the two mesh units, and: Δx=x j ―x i ,Δy=y j ―y i ,Δz=z j ―z i Where: x i ,y i and z i is the centroid coordinate of the i-th grid cell; x j ,y j and z j is the centroid coordinate of the jth grid cell.

6. The stress field rapid calculation method based on the graph network agent model according to claim 1 is characterized by: In step 3, global features are introduced into the graph network structure, which is expressed as: Among them: Q is the key global feature, and represents the set of real numbers; 1×(d v +d Q ) represents the dimension of global features; global features include external loads [f x ,f y ,f z ]、Measurement point position [x f ,y f ,z f ] and the stress signal of the measuring point [∈e f1 ,∈e f2 ...∈e fn ]; In each layer of the graph network structure, the global feature Q is directly concatenated to the attribute vector of each node On the top, we get the enhanced node attribute vector: in: is the enhanced node attribute vector; || represents the concatenation operation.

7. The stress field rapid calculation method based on the graph network agent model according to claim 1 is characterized by: In step 4, the encoder is used to embed the node attribute vector in the graph network structure into a high-dimensional vector space to obtain: Among them: MLP en-v is the multi-layer perceptron in the node encoder; represents the hidden state of the node at step t = 0; Represents the input vector of the node The dimension is 1×(d v +d θ ); represents the set of real numbers; d v Indicates the feature dimension of the node; d θ Represents other feature dimensions related to the node; The encoder is used to generate edge encoding features through non-learning transformation of edge features, and obtain: Among them: MLP en―e is the multi-layer perceptron in the edge encoder; represents the hidden state of the edge at step t = 0; } represents the edge attribute vector.

8. The method for rapid calculation of stress field based on graph network agent model according to claim 1 is characterized in that: In step 4, the method by which the processor updates the hidden state of the node through a recursive message passing mechanism is: Calculate each node v i Neighbor nodes The aggregate information transmitted defines the message aggregation function as follows: in: is node v i In the hidden state of step t, it describes the feature information of the current node; is the neighbor node v j The hidden state of is used to provide contextual information; ij is the edge feature, indicating that node v i and v j The relationship between M is the message passing function, defined as a multilayer perceptron: Introduce attention mechanism to improve message selectivity and model expressiveness: Among them: || represents the vector concatenation operation; α ij is the attention weight between nodes, and: in: represents the attention score function, which is used to calculate the correlation score between nodes; The hidden state of the node will be based on the received message Update, introduce residual connection and layer normalization in the state update process, the state update formula is: Among them: LayerNorm represents layer normalization; Φ U It is the state update function.

9. The method for rapid calculation of stress field based on graph network agent model according to claim 1 is characterized by: In step 4, the decoder is used to predict the paradigm equivalent stress of each node, which is expressed as: Among them: MLP de The MLP of the node decoder is used to map the processed features to the output features, completing the conversion process of the encoder-processor-decoder model; is the hidden state of the node at step T; is the processed vertex set.

Images