GNN-based blast load space-time distribution prediction method

Abstract

The invention discloses a GNN-based blast load spatial and temporal distribution prediction method, which comprises the following steps of: 1, acquiring blast-related test and numerical simulation results, preprocessing the test and numerical simulation results, and establishing a standard data set; step 2, constructing a GNN-based explosion pressure space-time distribution prediction model, and inputting the data set into the model for training to obtain a training model; and step 3, inputting a to-be-predicted explosion initial state into the training model to obtain a prediction result. According to the method, data processing is carried out by utilizing a grid division technology, and an artificial intelligence model is constructed and trained through a graph neural network, so that spatial and temporal distribution prediction of the explosion pressure load is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of explosion load prediction, and in particular to a method for predicting the spatiotemporal distribution of explosion load based on GNN. Background Art

[0002] In recent years, international hotspot conflicts have continued, and military strikes have caused a lot of damage to infrastructure and buildings. Among them, the explosion shock wave is one of the important damage elements. Accurate and rapid prediction of explosion shock loads is not only of great significance to the resistance of buildings to explosion shock, but also can provide a reference for the effective attack of weapons on targets.

[0003] The existing method for calculating explosion loads is mainly to perform computational fluid dynamics simulation calculations using finite element software. This method is limited by computing power, often has limited calculation scale, and consumes a lot of time. At present, there are also studies that attempt to use artificial intelligence methods to predict explosion loads. However, in these studies, the prediction objects are all some main characteristic parameters of the explosion load (such as peak pressure, peak impulse, etc.), which have the following limitations: (1) It is impossible to fully predict the pressure field of the explosion impact in all time and space, so as to fully describe the explosion process. (2) For a type of explosion scene, some methods are often used to abstract the geometric characteristics of the spatial environment where the explosion occurs into a small number of characteristic parameters, which makes it difficult to accurately describe more complex spatial environments, limiting the use scenarios of artificial intelligence models. Therefore, it is of great significance to develop a general explosion load prediction method that can quickly predict explosion loads and cover both time and space dimensions. Summary of the invention

[0004] Purpose of the invention: The present invention provides a method for predicting the spatiotemporal distribution of explosion loads based on GNN. By establishing an explosion physical flow field dataset, combining grid division technology and graph neural network to build and train an artificial intelligence model, the spatiotemporal distribution prediction of explosion pressure loads can be achieved.

[0005] Technical solution: The method for predicting the spatiotemporal distribution of explosion load based on GNN described in the present invention comprises the following steps:

[0006] Step 1: Obtain explosion-related test and numerical simulation results, perform data preprocessing on the test and numerical simulation results, and establish a standard data set;

[0007] Step 2: construct a GNN-based prediction model for the spatiotemporal distribution of explosion pressure, input the data set into the model for training, and obtain a training model;

[0008] Step 3: input the initial state of the explosion to be predicted into the training model to obtain a prediction result.

[0009] Furthermore, in step 1, the explosion-related test and numerical simulation results are obtained, and data preprocessing is performed on the test and numerical simulation results to establish a standard data set, which specifically includes the following steps:

[0010] Step 11: Collection of explosion load data: Perform CFD calculations using finite element software, collect explosion examples under different environmental conditions, and verify the accuracy of the simulation results of the finite element software using the experimental results of relevant tests;

[0011] Step 12, data preprocessing: The explosion space is divided into a number of spatial hexahedral grid units by meshing technology. The size of the grid unit is specified according to the flow field resolution requirement, and the finite element simulation results in step 11 are mapped to the grid system. The explosion physical information is extracted in the following order: (a) the grid topology relationship and model boundary information are collected, (b) the physical information of the pressure, flow velocity, position and density of each node is collected, and (c) the node type is encoded according to the model boundary information and density information. The encoding follows the following rules: the node air type is 0, the explosive / detonation product type is 1, the open boundary type is 3, and the wall boundary type is 4. After the above steps, the process of associating the relevant physical information and the node-edge adjacency relationship with the nodes and edges of the grid unit is completed;

[0012] Step 13, construction of standard data set: the preprocessed data is stored in the group of corresponding examples in the form of several three-dimensional tensor data tables. The contents of the data tables include: the node composition of each grid unit used to record the adjacency relationship, and the corresponding data table size is: number of grids × number of time steps × 8, the pressure value corresponding to the node, and the corresponding data table size is: number of nodes × number of time steps × 1, node coordinate information, and the corresponding data table size is: number of nodes × number of time steps × 3, node type information, and the corresponding data table size is: number of nodes × number of time steps × 1, gas flow rate information at the node, and the corresponding data table size is: number of nodes × number of time steps × 3, node type information, and the corresponding data table size is: number of nodes × number of time steps × 1. The three dimensions of the data table correspond to the spatial dimension, time dimension and physical quantity dimension. According to the above process, the preprocessed data of each example are recorded to construct a standardized data set.

[0013] Furthermore, the explosion physics information data used to construct the data set includes explosion spaces of different sizes and shapes, different types of spatial boundary conditions, explosives with different charges, different types of explosives, and different forms of charge shapes.

[0014] Furthermore, in step 2, a GNN-based explosion pressure spatiotemporal distribution prediction model is constructed, including an encoding module, a processing module including a GNN module, and a decoding module; the encoding module includes two three-layer multilayer perceptrons MLP, with a hidden layer dimension of 128 layers, which serve as a node encoder and an edge encoder respectively. The node encoder is used to encode the data into graph structure data, read the physical information of pressure, flow rate, coordinates and type from the data table generated in step 13 and input it, so as to realize the encoding of the physical information of pressure, flow rate, coordinates and type into the node features, and the edge encoder reads the node composition of each grid unit in the data table generated in step 13, and combines the input in the order of bottom surface -> top surface -> side surface to form an edge feature for recording the adjacency relationship between nodes; the processing module needs to repeatedly call the GNN module therein to perform data fusion and update on the graph structure data. The GNN module contains a node information transmission method and an edge information transmission method. The information transmission methods are all multi-layer perceptrons (MLPs). First, the information in all node feature vectors in the graph data is transmitted to the edge feature vectors connected to it through the node information transmission method to obtain updated edge feature vectors. Then, the information in the updated edge feature vectors is transmitted to the nodes at both ends of the corresponding edges through the edge information transmission method to obtain updated node vectors. The decoding module includes a three-layer multi-layer perceptron MLP with a hidden layer dimension of 128 layers. After the obtained graph structure data is input, four-dimensional output data are obtained, which correspond to the node velocity (3D) and pressure (1D) data of each node, i.e., the physical information prediction result at the next moment. Among them, the pressure field information is the main prediction target, and the velocity field information is used to calculate the loss in the training process. The model training is carried out in coordination with the pressure field prediction loss, and the pressure field prediction information is compared and verified with each other to prevent overfitting of the model during the training process and improve the efficiency of model training.

[0015] Furthermore, the node encoder is used to encode the data into graph structure data, encode the physical information of pressure, flow rate, coordinates and type into the node features, and encode the adjacency relationship between nodes into the edge features. The specific working method is as follows:

[0016] X'=[x'1,x'2,...,x' N ]=MLP node (X)

[0017] E'=[e'1,e'2,...,e' N ]=MLP edge (E)

[0018] After encoding is completed, the graph structure data returned is as follows:

[0019] data(x=X',edge_attr=E',edge_index=Eindex )

[0020] Among them, X' is the encoded node feature vector, E' is the encoded edge feature vector, and E index It is the node-edge connection information obtained after data preprocessing.

[0021] Furthermore, the information in all node feature vectors in the data obtained in the graph is transferred to the edge feature vectors connected thereto through the node information transfer method to obtain an updated edge feature vector, and then the information in the updated edge feature vector is transferred to the nodes at both ends of the corresponding edge through the edge information transfer method to obtain an updated node vector. The specific process is shown in the following formula:

[0022] e' ij =f E (e ij ,v i ,v j )

[0023]

[0024] where e' ij Represents the connection node v i and v j The edge of the multi-layer perceptron f E Will e' ij 、v i and v j The information at a certain moment is fused and updated to the edge feature e' at the next moment ij , and then through the multi-layer perceptron f V The node v i and all updated edges e' connected to it ij The information is fused and updated to the node feature v' at the next moment i , repeatedly call GNNBlock to perform L information fusion to realize the process of updating various feature information in the graph to the next time step.

[0025] Furthermore, the error used in the model training process is a linear combination of the root mean square error of the velocity data and the pressure data. When performing back propagation for gradient descent, the prediction information of the velocity field and the prediction information of the pressure field can be taken into account at the same time to ensure that the prediction results conform to the basic physical laws and effectively prevent overfitting during training. By adjusting the ratio of linear combination coefficients, the model training effect is improved for different explosion condition data sets. At the same time, masks are added to nodes with rigid boundary types and noise is added to nodes with air types for training to prevent overfitting and improve training efficiency.

[0026] Furthermore, the selection of relevant hyperparameters is as follows: when the velocity error and the pressure error are linearly combined, the ratio of the combination coefficient is 1 to 4, the optimizer is Adam optimizer, and the learning rate is 10 -5 ~10 -4 , Batch_size: 0~20; training 0~100 rounds.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the explosion load prediction method of the present invention can cover the time and space dimensions, and can directly predict the distribution of the explosion load in the spatial dimension and the evolution process in the time dimension, and has a direct and clear description of the prediction result of the explosion process; the present invention uses GNN for feature extraction and information fusion on the basis of grid division technology, has a high feature extraction capability, improves the accuracy of the prediction, can realize parallel computing, and has higher reliability and operating efficiency; the explosion load prediction method of the present invention can independently train different explosion load data sets according to the user's selectivity, and can accurately predict the explosion load under different types of explosives, explosion spaces of different geometric shapes, different boundary conditions, etc., which enhances the generalization ability of the model and is more flexible and accurate in dealing with diverse explosion conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0029] Figure 2 Schematic diagram of the method flow for constructing a data set in the present invention.

[0030] Figure 3 It is a comparison diagram between the prediction results of the present invention and the finite element calculation results. DETAILED DESCRIPTION

[0031] like Figure 1 As shown, a method for predicting the spatiotemporal distribution of explosion loads based on GNN includes the following steps:

[0032] S1: Obtain explosion-related test and numerical simulation results, perform data preprocessing on the test and numerical simulation results, and establish a standard data set;

[0033] S2: constructing a prediction model for the spatiotemporal distribution of explosion pressure based on GNN, inputting the data set into the model for training, and obtaining a training model;

[0034] S3: Inputting the initial state of the explosion to be predicted into the training model to obtain a prediction result.

[0035] like Figure 2 As shown, the specific steps of S1 in this embodiment are:

[0036] 1.1 Collection of explosion load data: CFD calculations are performed using finite element software, explosion examples under different environmental conditions are collected, and the accuracy of the simulation results of the finite element software is verified through the experimental results of relevant tests;

[0037] 1.2 Data preprocessing: The explosion space is divided into several spatial hexahedral grid units through meshing technology. According to the flow field resolution requirements, the size of the grid unit is specified, and the finite element simulation results in step 11 are mapped to the grid system. The explosion physical information is extracted in the following order: (a) the grid topology relationship and model boundary information are collected, (b) the physical information of pressure, flow velocity, position and density of each node is collected, (c) the node type is encoded according to the model boundary information and density information, and the encoding follows the following rules: the node air type is 0, the explosive / detonation product type is 1, the open boundary type is 3, and the wall boundary type is 4. After the above steps, the relevant physical information and node-edge adjacency relationship are associated with the nodes and edges of the grid unit;

[0038] 1.3 Construction of standard data set: The preprocessed data are stored in the corresponding example group in the form of several three-dimensional tensor data tables. The contents of the data tables include: the node composition of each grid unit used to record the adjacency relationship, and the corresponding data table size is: number of grids × number of time steps × 8, the pressure value corresponding to the node, and the corresponding data table size is: number of nodes × number of time steps × 1, node coordinate information, and the corresponding data table size is: number of nodes × number of time steps × 3, node type information, and the corresponding data table size is: number of nodes × number of time steps × 1, gas flow rate information at the node, and the corresponding data table size is: number of nodes × number of time steps × 3, node type information, and the corresponding data table size is: number of nodes × number of time steps × 1. The three dimensions of the data table correspond to the spatial dimension, time dimension and physical quantity dimension. According to the above process, the preprocessed data of each example are recorded to construct a standardized data set.

[0039] The data set in this embodiment consists of 10 examples of explosions in a confined space within a cubic computational domain. The explosive equivalent range is 100 to 2000 g, the computational domain side length is 2 m, and the explosives are all located in the center of the computational space. Each example contains 200 time steps, and the total computation time is 5 ms. The prediction results include the change process of the pressure field distribution in the computational space over time.

[0040] The S2 in this example is specifically to design a GNN-based explosion pressure spatiotemporal distribution prediction model, such as Figure 1 , including an encoding module, a processing module and a decoding module.

[0041] The encoding module in this example is used to obtain the graph structure data corresponding to the physical state at a certain moment in the sample, to encode the pressure, coordinates, flow rate, and type information into the node features of the graph structure data, and to encode the adjacency relationship between nodes into the edge features of the graph structure data.

[0042] The processing module in this example includes a GNN module, which is called cyclically. The GNN module updates the node information and the edge information by fusing them. The above process is repeated several times to obtain updated graph structure data after the information fusion is completed.

[0043] The decoding module in this example is used to obtain the position and pressure information associated with the node, which is regarded as a prediction result of the physical state of the pressure field distribution at the next moment.

[0044] The working steps of the basic model of S2 in this example are as follows:

[0045] 2.1 Put the samples in the standardized data set into the processing module and convert them into graph format data through MLP;

[0046] 2.2 The graph format data is put into the processing module, and the GNN module is repeatedly called to perform data fusion and update on the graph structure data. The GNN module contains a node information transmission method and an edge information transmission method. Both information transmission methods are multi-layer perceptrons (MLP). First, the node information transmission method is used to transmit the information in the feature vectors of all nodes in the graph to the edge feature vectors connected to them to obtain an updated edge feature vector. Then, the edge information transmission method is used to transmit the information in the updated edge feature vector to the nodes at both ends of the corresponding edge to obtain an updated node vector, thereby updating the graph format data.

[0047] 2.3 The updated graph format data is input into the decoding module to obtain the predicted velocity field information and pressure field information at the next moment;

[0048] 2.4 Model training: The error used in the model training process is a linear combination of the root mean square error of the velocity data and the pressure data, and masks are added to the nodes whose node type is a rigid boundary for training.

[0049] The data format conversion method in 2.1 is specifically as follows:

[0050] X'=[x'1,x'2,...,x' N ]=MLP node (X)

[0051] E'=[e'1,e'2,...,e' N ]=MLP edge (E)

[0052] The graph format data is as follows:

[0053] data(x=X',edge_attr=E',edge_index=E index )

[0054] Among them, X' is the encoded node feature vector, E' is the encoded edge feature vector, and E index It is the node-edge connection information obtained after data preprocessing;

[0055] Preferably, the above MLP has three layers, and the hidden layer dimension is 128.

[0056] The specific method for updating the graph structure data in 2.2 is:

[0057] e' ij =f E (e ij ,v i ,v j )

[0058]

[0059] where e' ij Represents the connection node v i and v j The edge of the multi-layer perceptron f E Will e' ij 、v i and v j The information at a certain moment is fused and updated to the edge feature e' at the next moment ij . Then, through the multi-layer perceptron f V The node v i and all updated edges e' connected to it ij The information is fused and updated to the node feature v' at the next moment i . Repeatedly call GNNBlock to perform L information fusion to achieve the process of updating various feature information in the graph to the next time step;

[0060] Preferably, the MLP of the decoding module in 2.3 has three layers, wherein the hidden layer dimension is 128.

[0061] Preferably, the relevant hyper parameters for model training in 2.4 of this example are selected as follows:

[0062] When the velocity error and the pressure error are linearly combined, the ratio of the combination coefficient is 2.33, the optimizer is Adam optimizer, and the learning rate is 5*10 -5 , Batch_size: 2; training for 70 rounds.

[0063] The specific steps of S3 in this example are:

[0064] The initial state of the explosion to be predicted is input into the training model to obtain the prediction results, and the pressure flow field distribution prediction results of 0.75ms, 1.5ms, 2.25ms and 3ms are selected for comparison with the CFD calculation results using blastFoam. Figure 3 shown.

[0065] In summary, this application obtains explosion-related test and numerical simulation results, pre-processes the test and numerical simulation results, and establishes a standard data set; secondly, constructs a GNN-based explosion pressure spatiotemporal distribution prediction model, which includes an encoding module, a processing module, and a decoding module to achieve the graph format conversion and information fusion update of the physical state, and then converts the updated graph format data into physical state information; the data set is input into the model for training to obtain a training model; finally, the initial state of the explosion to be predicted is input into the training model to obtain the prediction result of the spatiotemporal distribution of the pressure field. The present invention provides a technical basis for the development of protective engineering and weapon engineering by combining graph neural networks, multi-layer perceptrons, and information transmission mechanisms.

Claims

1. A method for predicting the spatiotemporal distribution of explosion loads based on GNN, characterized in that: The steps include: Step 1: Obtain explosion-related test and numerical simulation results, perform data preprocessing on the test and numerical simulation results, and establish a standard data set; Step 2: construct a GNN-based prediction model for the spatiotemporal distribution of explosion pressure, input the data set into the model for training, and obtain a training model; Step 3: input the initial state of the explosion to be predicted into the training model to obtain a prediction result.

2. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 1, characterized in that: In step 1, the explosion-related test and numerical simulation results are obtained, and data preprocessing is performed on the test and numerical simulation results. The establishment of a standard data set specifically includes the following steps: Step 11: Collection of explosion load data: Perform CFD calculations using finite element software, collect explosion examples under different environmental conditions, and verify the accuracy of the simulation results of the finite element software using the experimental results of relevant tests; Step 12, data preprocessing: the explosion space is divided into a number of spatial hexahedral grid units by using a grid division technique. The size of the grid unit is specified according to the flow field resolution requirement, and the finite element simulation results in step 11 are mapped to the grid system to extract the explosion physical information. After the above steps, the process of associating the relevant physical information and the node-edge adjacency relationship with the nodes and edges of the grid unit is completed; Step 13: Construction of a standard data set: The preprocessed data is stored in the corresponding example group in the form of several three-dimensional tensor data tables. The preprocessed data of each example is recorded according to the above process to construct a standardized data set.

3. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 2, characterized in that: In step 12, the explosion physics information data used to construct the data set includes explosion spaces of different sizes and shapes, different types of spatial boundary conditions, explosives with different charges, different types of explosives, and different forms of charge shapes.

4. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 1, characterized in that: In step 12, the explosion physical information is extracted in the following order: (a) collecting the grid topological relationship and model boundary information, (b) collecting the physical information of pressure, flow velocity, position and density of each node, and (c) encoding the node type according to the model boundary information and density information. The encoding follows the following rules: the node air type is 0, the explosive / detonation product type is 1, the open boundary type is 3, and the wall boundary type is 4.

5. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 1, characterized in that: In step 13, the content of the data table includes: the node composition of each grid unit for recording the adjacency relationship, and its corresponding data table size is: number of grids × number of time steps × 8, the pressure value corresponding to the node, and its corresponding data table size is: number of nodes × number of time steps × 1, node coordinate information, and its corresponding data table size is: number of nodes × number of time steps × 3, node type information, and its corresponding data table size is: number of nodes × number of time steps × 1, gas flow rate information at the node, and its corresponding data table size is: number of nodes × number of time steps × 3, node type information, and its corresponding data table size is: number of nodes × number of time steps × 1. The three dimensions of the data table correspond to the spatial dimension, time dimension and physical quantity dimension.

6. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 1, characterized in that: In step 2, a GNN-based explosion pressure spatiotemporal distribution prediction model is constructed, including an encoding module, a processing module including a GNN module, and a decoding module; the encoding module includes two three-layer multi-layer perceptrons MLP, with a hidden layer dimension of 128 layers, which serve as a node encoder and an edge encoder respectively. The node encoder is used to encode the data into graph structure data, read the pressure, flow rate, coordinate, and type physical information from the data table generated in step 13 and input them to realize the encoding of the pressure, flow rate, coordinate, and type physical information into the node features. The edge encoder reads the node composition of each grid unit in the data table generated in step 13, and combines the input in the order of bottom surface -> top surface -> side surface to form an edge feature for recording the adjacency relationship between nodes; the processing module needs to repeatedly call the GNN module therein to perform data fusion and update on the graph structure data. The GNN module contains a node information transmission method and an edge information transmission method. Both information transmission methods are multi-layer perceptrons (MLPs). First, the information in all node feature vectors in the graph data is transmitted to the edge feature vectors connected to it through the node information transmission method to obtain updated edge feature vectors. Then, the information in the updated edge feature vectors is transmitted to the nodes at both ends of the corresponding edges through the edge information transmission method to obtain updated node vectors. The decoding module includes a three-layer multi-layer perceptron MLP with a hidden layer dimension of 128 layers. After the obtained graph structure data is input, output data of four dimensions are obtained, which correspond to the node flow velocity and pressure data of each node, i.e., the physical information prediction result at the next moment. Among them, the pressure field information is the main prediction target, and the flow velocity field information is used to calculate the loss in the training process. The model training is carried out in coordination with the pressure field prediction loss, and the pressure field prediction information is compared with each other to prevent overfitting of the model during training and improve the efficiency of model training.

7. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 6, characterized in that: The node encoder is used to encode data into graph structure data, encode the physical information of pressure, flow rate, coordinates and type into node features, and encode the adjacency relationship between nodes into edge features. Its specific working method is as follows: X'=[x'1,x'2,...,x' N ]=MLP node (X) E'=[e'1,e'2,...,e' N ]=MLP edge (E) After encoding is completed, the graph structure data returned is as follows: data(x=X',edge_attr=E',edge_index=E index ) Among them, X' is the encoded node feature vector, E' is the encoded edge feature vector, and E index It is the node-edge connection information obtained after data preprocessing.

8. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 6, characterized in that: The node information transmission method is used to transmit the information in all node feature vectors in the data obtained by the graph to the edge feature vectors connected to it, and the updated edge feature vectors are obtained. Then, the edge information transmission method is used to transmit the information in the updated edge feature vectors to the nodes at both ends of the corresponding edges, and the updated node vectors are obtained. The specific process is shown in the following formula: it' ij =f E (e ij ,v i ,v j ) where e' ij Represents the connection node v i and v j The edge of the multi-layer perceptron f E Will e' ij 、v i and v j The information at a certain moment is fused and updated to the edge feature e' at the next moment ij , and then through the multi-layer perceptron f V The node v i and all updated edges e' connected to it ij The information is fused and updated to the node feature v' at the next moment i , repeatedly call GNN Block to perform L information fusions to realize the process of updating various feature information in the graph to the next time step.

9. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 6, characterized in that: The error used in the model training process is a linear combination of the root mean square error of the velocity data and the pressure data. When performing back propagation for gradient descent, it can take into account the prediction information of the velocity field and the prediction information of the pressure field at the same time. By adjusting the ratio of the linear combination coefficients, the model training effect is improved for different explosion condition data sets. At the same time, masks are added to nodes with rigid boundary type and noise is added to nodes with air type for training.

10. The method for predicting the spatiotemporal distribution of explosion loads based on GNN as claimed in claim 1, characterized in that: In step 3, the selection of relevant hyperparameters is as follows: when the velocity error and the pressure error are linearly combined, the ratio of the combination coefficient is 1 to 4, the optimizer is Adam optimizer, and the learning rate is 10 -5 ~10 -4 , Batch_size: 0~20; training 0~100 rounds.

Images