Building penetration trajectory prediction method based on graph neural network

Abstract

The invention discloses a building penetration trajectory prediction method based on a graph neural network, and the method comprises the steps: building a three-dimensional refined finite element model based on key factors, carrying out the numerical simulation through LS-DYNA software, obtaining building penetration trajectory data, processing the data, and obtaining a building target penetration trajectory database. Dividing the database into a training set, a verification set and a test set in proportion; building a building penetration trajectory prediction model based on the graph neural network architecture, training the building penetration trajectory prediction model by using the training set, performing performance evaluation on the trained model by using the verification set, and stopping training when a prediction error converges to obtain a trained building penetration trajectory prediction model; and inputting the building penetration trajectory data into the trained building penetration trajectory prediction model to obtain a final predicted building penetration trajectory. The method has the advantages of high prediction precision, wide application range and high calculation efficiency.

Description

Technical Field

[0001] This invention relates to the field of terminal ballistics technology, specifically to a method for predicting building penetration trajectories based on graph neural networks. Background Technology

[0002] In existing technologies, the prediction of the penetration trajectory of building targets mainly relies on empirical formulas, theoretical analytical methods, or numerical simulation. Empirical formulas are established based on fitting experimental data, are simple in form, computationally efficient, and reliable under specific projectile-target conditions; however, they are difficult to reproduce the details of the damage process and are limited by the experimental conditions on which they are based, making them unsuitable for complex conditions such as oblique penetration and multi-layered targets, thus limiting their applicability. Theoretical analytical methods have a strong physical foundation and are more universal than empirical formulas; however, their core lies in solving for penetration resistance and depth, and they cannot accurately describe key local damage phenomena such as impact cratering and collapse, resulting in insufficient completeness of calculation results; furthermore, this method still relies heavily on simplification and correction when dealing with multi-layered targets and oblique penetration, limiting its applicability. Numerical simulation has the highest result completeness and scenario adaptability among these three methods, capable of accurately reproducing the entire penetration process, structural response, and local damage characteristics; however, its computational cost is extremely high and its efficiency is low, especially when simulating the penetration of large-spacing multi-layered targets, where the computational load is enormous and cannot meet the urgent need for real-time trajectory prediction.

[0003] Leveraging the powerful nonlinear mapping capabilities of machine learning, various machine learning models have been successfully applied in recent years to rapidly predict projectile penetration effects. These algorithms, through learning from extensive experimental data, effectively capture the complex nonlinear relationships between factors such as initial penetration velocity, projectile geometry, and structural material properties, and projectile penetration performance, thereby establishing highly accurate prediction models. However, existing machine learning prediction models are often developed based on structured tabular data, outputting feature scalars such as projectile penetration depth and residual velocity, failing to capture the distribution characteristics of damage effects, such as equivalent stress and strain. Furthermore, existing models rely solely on nonlinear mapping capabilities, directly predicting performance indicators based on artificially set initial conditions, neglecting the description of key damage processes and failing to learn the intrinsic physical mechanisms underlying damage phenomena, thus limiting their generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting building penetration trajectories based on graph neural networks, which has the advantages of high accuracy, strong generalization ability, and high computational efficiency.

[0005] The present invention adopts the following technical solution:

[0006] A method for predicting building penetration trajectories based on graph neural networks includes the following steps:

[0007] S1. Based on key factors, a three-dimensional refined finite element model is constructed. Numerical simulation is carried out using LS-DYNA software to obtain building penetration trajectory data. The data is then processed to obtain a building target penetration trajectory database, which is then divided into training set, validation set, and test set according to the proportions.

[0008] S2. Build a building penetration trajectory prediction model based on graph neural network architecture. Train the building penetration trajectory prediction model using the training set and evaluate the performance of the trained model using the validation set. Stop training when the prediction error converges to obtain the trained building penetration trajectory prediction model.

[0009] S3. Input the building penetration trajectory data into the trained building penetration trajectory prediction model to obtain the final predicted building penetration trajectory.

[0010] Furthermore, in step S1, key factors include impact velocity, projectile tilt angle, projectile angle of attack, projectile diameter, projectile length-to-diameter ratio, curvature-to-diameter ratio, and floor thickness.

[0011] A three-dimensional refined finite element model was constructed using the adaptive smooth particle hydrodynamics method;

[0012] The building penetration trajectory data includes the coordinates, velocity, equivalent stress, and equivalent plastic strain of the projectile and the building nodes; the projectile is assumed to remain rigid during penetration, the rotation of the projectile about its axis is ignored, the projectile motion is limited to a selected two-dimensional plane, and the deflection of the projectile in directions outside the plane is ignored.

[0013] The time history data of the projectile and building nodes on the symmetry plane in the building penetration trajectory data are extracted and global normalized to obtain the building target penetration trajectory database.

[0014] Furthermore, in step S2, the building intrusion trajectory prediction model includes an encoder module, a processor module, and a decoder module; wherein the encoder module includes node encoding functions and edge encoding functions; the processor module includes ten cascaded graph network modules, and the graph network modules include node update functions and edge update functions; the decoder module includes node decoding functions.

[0015] Furthermore, in step S2, the training set is input into the building penetration trajectory prediction model. Dynamic parameters are obtained through forward propagation, and the mean squared error is used as the loss function. The Adam optimizer is used for parameter updates, and the learning rate employs an exponential decay strategy, with the specific formula as follows:

[0016] ;

[0017] in, This represents the updated learning rate. This represents the initial learning rate. Indicates the attenuation coefficient. Indicates the current iteration step. Indicates the total number of decay steps;

[0018] The performance of the building penetration trajectory prediction model is evaluated using a validation set, with the root mean square error as the evaluation metric. Training is stopped when the loss on the validation set converges, and the trained building penetration trajectory prediction model is obtained.

[0019] The trained building penetration trajectory prediction model was evaluated using a test set, with the root mean square error (RMSE) as the final performance metric.

[0020] Furthermore, in step S3, graph data is constructed based on the building penetration trajectory data. The graph data includes node features and edge features. The node features include the velocity components of each node in the first five time steps and the node type embedding. The edge features include the relative displacement vectors between neighboring nodes and their scalar distances.

[0021] The graph data is input into the building intrusion trajectory prediction model. The node features are processed by the node encoding function in the encoder module to obtain the node embedding vector, and the edge features are processed by the edge encoding function in the encoder module to obtain the edge embedding vector.

[0022] The processor module is used to update the node embedding vector and edge embedding vector in a multi-layer graph network. The neighboring node embedding vector and edge embedding vector are aggregated and processed by the edge update function to obtain the updated edge embedding vector. The target node embedding vector and the updated neighboring edge embedding vector are aggregated and processed by the node update function to obtain the updated node embedding vector.

[0023] After the node embedding vector is processed by the node decoding function in the decoder module, the target dynamic parameters are obtained, which include the node acceleration components and damage.

[0024] Based on the nodal acceleration components, the velocity components of the nodes are updated using the Euler integral method, with the specific formula as follows:

[0025] ;

[0026] in, For the first Each node at time... Quantity, For the first Each node at time... velocity components, For the first Each node at time... acceleration components, For time step;

[0027] Based on the velocity components of the nodes, the node coordinates are updated using the Euler integral method, with the specific formula as follows:

[0028] ;

[0029] in, For the first Each node at time... , For the first Each node at time... The coordinates;

[0030] Based on the updated velocity components and coordinates of the nodes, the updated building penetration trajectory is obtained.

[0031] The updated building penetration trajectory is corrected using a rigidity correction module through rigid transformation. The specific steps are as follows:

[0032] Step 1: Based on the initial coordinates of the node (time) ) and time The node coordinates are used to calculate the initial geometric center coordinates and time. The geometric center coordinates are given by the following formula:

[0033] ;

[0034] ;

[0035] in, Represents the initial geometric center coordinates of the node. Indicates at time The geometric center coordinates of the node Indicates the number of nodes. Indicates the first The initial coordinates of each node;

[0036] Step 2: Calculate the rigid displacement of the nodes The specific formula is as follows:

[0037] ;

[0038] Step 3: Based on the initial and current principal inertial axis directions of the node, obtain the overall rigid rotation angle of the node. ;

[0039] Step 4: Based on the rigid rotation angle Constructing a two-dimensional rigid rotation matrix The specific formula is as follows:

[0040] ;

[0041] Step 5: Update the first step based on rigid translation and rotation transformations. Each node at time... The coordinates are given by the following formula:

[0042] ;

[0043] in, For the first Each node at time... Corrected coordinates;

[0044] Based on the corrected coordinates of the nodes, the final predicted building intrusion trajectory is obtained.

[0045] Furthermore, the R-radius method is used to establish the connection relationship between neighboring nodes in the graph data, and the Euclidean distance between the target node and other nodes is calculated. When the distance does not exceed the connected radius, the other node is regarded as a neighboring node of the target node, and an undirected edge is established between the two.

[0046] The selection criteria for the connectivity radius are as follows:

[0047] The dynamic response between nodes is driven by wave propagation behavior; at each time step Within the processor, spatial information flows through message passing, and the maximum propagation range of message passing... for:

[0048] ;

[0049] in, The range of a single message transmission determined by the connectivity radius; For the number of message passes; For waves in a real physical system at time steps The propagation distance within, , The elastic wave velocity of the target material, , The Young's modulus of the material. Density;

[0050] exist Under fixed conditions, the critical radius The calculation formula is:

[0051] ;

[0052] The value of the connectivity radius is not less than .

[0053] Furthermore, both the node encoding function and the edge encoding function include a multilayer perceptron and a layer normalization layer. The multilayer perceptron includes an input layer, a hidden layer, and an output layer, with a LeakyReLU activation function connected after the hidden layer.

[0054] The node features are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node encoding function to obtain the hidden representation. The hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. The activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node encoding function to map the hidden representation to the desired embedding dimension, resulting in the output result. This result is then processed by normalization and learnable affine transformation of the normalization layer in the node encoding function to obtain the node embedding vector.

[0055] The process of obtaining edge embedding vectors from edge features through edge encoding functions is the same as the process of obtaining node embedding vectors based on node features.

[0056] Furthermore, both the node update function and the edge update function include a multilayer perceptron and a layer normalization layer; the multilayer perceptron includes an input layer, a hidden layer, and an output layer, with a LeakyReLU activation function connected after the hidden layer;

[0057] The node embedding vectors of the aggregated neighboring nodes and the edge embedding vectors of the corresponding edges are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the edge update function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the edge update function to map the hidden representation to the desired embedding dimension, resulting in the output result. This result is then processed by normalization and learnable affine transformation of the normalization layer in the edge update function. This result is then residually concatenated with the initial edge embedding vector to obtain the updated edge embedding vector.

[0058] The node embedding vector of the aggregate node and the updated edge embedding vector of its adjacent edges are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node update function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node update function to map the hidden representation to the desired embedding dimension, resulting in the output result. This output result is then processed by normalization and learnable affine transformation of the normalization layer in the node update function. This result is then residually concatenated with the initial node embedding vector to obtain the updated node embedding vector.

[0059] Furthermore, the node decoding function includes a multilayer perceptron; the multilayer perceptron includes an input layer, a hidden layer, and an output layer, with a LeakyReLU activation function connected after the hidden layer;

[0060] The updated node embedding is processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node decoding function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node decoding function to map the hidden representation to the desired embedding dimension, thus obtaining the target dynamic parameters.

[0061] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program, which is executed by a processor to perform the aforementioned building penetration trajectory prediction method based on graph neural networks.

[0062] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0063] 1. High Prediction Accuracy: This invention introduces a penetration trajectory prediction model based on graph neural networks, which can simultaneously output spatiotemporal distribution information of multiple physical quantities such as node coordinates, velocities, and damage, providing complete data dimensions. Through the proposed feature design and integral update mechanism, the penetration trajectory prediction model can dynamically predict the dynamic and mechanical response quantities during penetration based on the node velocity sequence, accurately predicting the deformation and damage evolution process of building targets under penetration. Simultaneously, the added rigidity correction module is used to perform rigid transformation constraint correction on the projectile prediction results, ensuring that the prediction results meet the rigid body kinematic constraints, thereby further improving prediction accuracy and physical consistency.

[0064] 2. High computational efficiency: This invention employs neural network matrix operations, significantly reducing computational complexity and time costs. Furthermore, this invention simplifies the three-dimensional penetration problem into a two-dimensional planar problem and extracts symmetry plane data based on refined three-dimensional numerical simulation results to construct a high-quality training set. While meeting the physical accuracy requirements of the penetration problem, this significantly improves computational efficiency and training speed.

[0065] 3. Strong generalization ability: The message passing operation of the processor module of this invention can effectively learn the underlying physical interaction laws during building penetration, and has a strong physical inductive bias ability, thus maintaining stable prediction performance under different impact conditions. At the same time, the selection of the K value in this invention is based on the wave propagation characteristics, ensuring that the propagation range at the information transmission level meets or exceeds the wave propagation distance requirements of the real physical system, thereby significantly improving the generalization ability and applicability. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0067] Figure 2 This is a schematic diagram of the encoder module of the present invention.

[0068] Figure 3 This is a schematic diagram of the processor module of the present invention.

[0069] Figure 4 This is a schematic diagram of the decoder module of the present invention.

[0070] Figure 5 This is a diagram showing the evaluation results of the trained building invasion trajectory prediction model in an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0072] To achieve the above objectives, this invention proposes a method for predicting building penetration trajectories based on graph neural networks, such as... Figure 1 As shown, the specific steps are as follows:

[0073] S1. Based on key factors, a three-dimensional refined finite element model is constructed. Numerical simulation is conducted using LS-DYNA software to obtain building penetration trajectory data. This data is then processed to obtain a building target penetration trajectory database, which is proportionally divided into training, validation, and test sets. Specifically:

[0074] Key factors include impact velocity, projectile tilt angle, projectile angle of attack, projectile diameter, projectile length-to-diameter ratio, curvature-to-diameter ratio, and floor thickness;

[0075] A semi-symmetric, refined three-dimensional finite element model was constructed using an adaptive smooth particle hydrodynamics method.

[0076] The building penetration trajectory data includes the coordinates, velocity, equivalent stress, and equivalent plastic strain of the projectile and the building nodes; the projectile is assumed to remain rigid during penetration, the rotation of the projectile about its axis is ignored, the projectile motion is limited to a selected two-dimensional plane, and the deflection of the projectile in directions outside the plane is ignored.

[0077] The time history data of the projectile and building nodes on the symmetry plane in the building penetration trajectory data are extracted and global normalized to obtain the building target penetration trajectory database.

[0078] S2. A building penetration trajectory prediction model is built based on a graph neural network architecture, implemented using PyTorch and PyTorch-Geometric. The model is trained using a training set and its performance is evaluated using a validation set. Training stops when the prediction error converges, resulting in a fully trained building penetration trajectory prediction model. Specifically:

[0079] The building intrusion trajectory prediction model includes an encoder module, a processor module, and a decoder module; among which, for example... Figure 2 As shown, the encoder module includes node encoding functions and edge encoding functions; as Figure 3 As shown, the processor module includes ten cascaded graph network modules, each containing node update functions and edge update functions; as... Figure 4 As shown, the decoder module includes node decoding functions;

[0080] The training set is input into the building penetration trajectory prediction model, and dynamic parameters are obtained through forward propagation. The mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates. The learning rate adopts an exponential decay strategy, and the specific formula is as follows:

[0081] ;

[0082] in, This represents the updated learning rate. This represents the initial learning rate. Indicates the attenuation coefficient. Indicates the current iteration step. Indicates the total number of decay steps;

[0083] The performance of the building penetration trajectory prediction model is evaluated using a validation set, with the root mean square error as the evaluation metric. Training is stopped when the loss on the validation set converges, and the trained building penetration trajectory prediction model is obtained.

[0084] The trained building penetration trajectory prediction model was evaluated using a test set, with the root mean square error (RMSE) as the final performance metric.

[0085] S3. Input the building penetration trajectory data into the trained building penetration trajectory prediction model to obtain the final predicted building penetration trajectory; specifically:

[0086] Graph data is constructed based on building penetration trajectory data. The graph data includes node features and edge features. The node features include the velocity components of each node in the first five time steps and the node type embedding. The velocity components are used to characterize the evolution trend of the node's motion state, and the node type is represented by a trainable embedding vector to capture the node's material category information. The edge features include the relative displacement vectors between neighboring nodes and their scalar distances, which are used to describe the spatial geometric relationships.

[0087] like Figure 2 As shown, the graph data is input into the building intrusion trajectory prediction model. The node features are processed by the node encoding function in the encoder module to obtain the node embedding vector, and the edge features are processed by the edge encoding function in the encoder module to obtain the edge embedding vector. Specifically, both the node encoding function and the edge encoding function include a multilayer perceptron and a layer normalization layer. The multilayer perceptron includes an input layer, a hidden layer and an output layer, and the hidden layer is connected to the LeakyReLU activation function.

[0088] The node features are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node encoding function to obtain the hidden representation. The hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. The activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node encoding function to map the hidden representation to the desired embedding dimension, resulting in the output result. This result is then processed by normalization and learnable affine transformation of the normalization layer in the node encoding function to obtain the node embedding vector.

[0089] The process of obtaining edge embedding vectors from edge features through edge encoding functions is the same as the process of obtaining node embedding vectors based on node features.

[0090] like Figure 3 As shown, a processor module is used to update the node embedding vector and edge embedding vector in a multi-layer graph network. The neighboring node embedding vector and edge embedding vector are aggregated, and after processing by the edge update function, the updated edge embedding vector is obtained. Similarly, the target node embedding vector and the updated neighboring edge embedding vector are aggregated, and after processing by the node update function, the updated node embedding vector is obtained. Specifically, both the node update function and the edge update function include a multi-layer perceptron and a layer normalization layer. The multi-layer perceptron includes an input layer, a hidden layer, and an output layer, with a LeakyReLU activation function connected after the hidden layer.

[0091] The node embedding vectors of the aggregated neighboring nodes and the edge embedding vectors of the corresponding edges are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the edge update function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the edge update function to map the hidden representation to the desired embedding dimension, resulting in the output result. This result is then processed by normalization and learnable affine transformation of the normalization layer in the edge update function. This result is then residually concatenated with the initial edge embedding vector to obtain the updated edge embedding vector.

[0092] The node embedding vector of the aggregate node and the updated edge embedding vector of its adjacent edges are processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node update function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node update function to map the hidden representation to the desired embedding dimension, resulting in the output result. This output result is then processed by normalization and learnable affine transformation of the normalization layer in the node update function. This result is then residually concatenated with the initial node embedding vector to obtain the updated node embedding vector.

[0093] like Figure 4 As shown, the updated node embedding vector is processed by the node decoding function in the decoder module to obtain the target dynamic parameters, which include node acceleration components and damage, used to describe the dynamic response and material damage characteristics of each node during the penetration process; specifically, the node decoding function includes a multilayer perceptron; the multilayer perceptron includes an input layer, a hidden layer and an output layer, and the hidden layer is connected to the LeakyReLU activation function;

[0094] The updated node embedding is processed by linear affine transformation and element-wise linear combination of the hidden layer of the multilayer perceptron in the node decoding function to obtain the hidden representation. This hidden representation is then processed by nonlinear mapping of the LeakyReLU activation function to obtain the activation representation. This activation representation is then processed by linear affine mapping of the output layer of the multilayer perceptron in the node decoding function to map the hidden representation to the desired embedding dimension, thus obtaining the target dynamic parameters.

[0095] Based on the nodal acceleration components, the velocity components of the nodes are updated using the Euler integral method, with the specific formula as follows:

[0096] ;

[0097] in, For the first Each node at time... Quantity, For the first Each node at time... velocity components, For the first Each node at time... acceleration components, For time step;

[0098] Based on the velocity components of the nodes, the node coordinates are updated using the Euler integral method, with the specific formula as follows:

[0099] ;

[0100] in, For the first Each node at time... , For the first Each node at time... The coordinates;

[0101] Based on the updated velocity components and coordinates of the nodes, the updated building penetration trajectory is obtained.

[0102] The updated building penetration trajectory is corrected using a rigidity correction module through rigid transformation. The specific steps are as follows:

[0103] Step 1: Based on the initial coordinates of the node (time) ) and time The node coordinates are used to calculate the initial geometric center coordinates and time. The geometric center coordinates are given by the following formula:

[0104] ;

[0105] ;

[0106] in, Represents the initial geometric center coordinates of the node. Indicates at time The geometric center coordinates of the node Indicates the number of nodes. Indicates the first The initial coordinates of each node;

[0107] Step 2: Calculate the rigid displacement of the nodes The specific formula is as follows:

[0108] ;

[0109] Step 3: Based on the initial and current principal inertial axis directions of the node, obtain the overall rigid rotation angle of the node. ;

[0110] Step 4: Based on the rigid rotation angle Constructing a two-dimensional rigid rotation matrix The specific formula is as follows:

[0111] ;

[0112] Step 5: Update the first step based on rigid translation and rotation transformations. Each node at time... The coordinates are given by the following formula:

[0113] ;

[0114] in, For the first Each node at time... Corrected coordinates;

[0115] Based on the corrected coordinates of the nodes, the final predicted building intrusion trajectory is obtained.

[0116] Among them, the R-radius method is used to establish the connection relationship of neighboring nodes in the graph data, calculate the Euclidean distance between the target node and other nodes, and when the distance does not exceed the connected radius, the other node is regarded as the neighboring node of the target node, and an undirected edge is established between the two.

[0117] The selection of the connectivity radius is based on wave propagation characteristics to ensure that the message transmission range in the diagram matches the wave propagation capability of the physical system. The specific selection criteria are as follows:

[0118] During message passing in the graph neural network, the dynamic responses (including acceleration, stress, and strain) between nodes are driven by wave propagation behavior; at each time step Internally, the processor achieves spatial information flow through message passing. Each message passing can only aggregate information from one neighboring node, ensuring that within the time step... The information propagation range within the system corresponds to the wave propagation distance in a real physical system, defining the maximum propagation range of message transmission. for:

[0119] ;

[0120] in, The range of a single message transmission determined by the connectivity radius; The number of message passes (i.e., the number of graph network modules); For waves in a real physical system at time steps The propagation distance within, , The elastic wave velocity of the target material, , The Young's modulus of the material. Density;

[0121] exist Under fixed conditions, the critical radius The calculation formula is:

[0122] ;

[0123] The value of the connectivity radius is not less than This ensures that the message transmission speed of the graph data is not lower than the material wave propagation speed, thereby achieving a physical consistency simulation of the wave propagation behavior during the penetration of building targets.

[0124] Example:

[0125] In this embodiment, the relevant information of the key factors is shown in Table 1. Based on the value range of the key factors, the LHS (Latin Hypercube Sampling) method is used for parameter space sampling and working condition design to ensure uniform sample coverage in the high-dimensional parameter space. A total of 108 sets of simulated working conditions were designed.

[0126] Table 1. Relevant information on key factors

[0127]

[0128] A semi-symmetric, refined 3D finite element model was imported into HyperMesh for mesh generation. Appropriate finite element meshes were generated, and mesh quality was checked. After meshing, the model was exported as an LS-DYNA keyword format file (.k file). This .k file was then opened using LS-PrePost, and pre-solution settings were completed, including boundary conditions, contact definitions, initial conditions, material models, and solution control parameters. Specific settings are as follows: In boundary conditions, symmetric boundary conditions were applied to the model's symmetry planes to utilize symmetry; the contact algorithm was selected as ERODING_SURFACE_TO_SURFACE; initial velocity components were applied to the projectile as initial conditions to define the impact condition. Regarding the material model, the projectile was defined as a rigid body, using LS-DYNA material model 020-RIGID, and the structural components were defined using a concrete constitutive model, specifically LS-DYNA material model 111-JOHNSON_HOLMQUIST_CONCRETE. For solving the control parameters, the projectile is modeled using the finite element method, while the building structure is modeled using the adaptive smooth particle hydrodynamic method. Solid elements that fail due to large deformation are automatically converted into SPH particles. The failure criterion is the maximum principal strain. When the maximum principal strain of an element exceeds the threshold of 0.3, the solid element is deleted and converted into an SPH particle. The total simulation time is set to 100ms, and the historical output results are saved every 0.5ms.

[0129] Based on the numerical simulation output, time-history data such as coordinates, velocity, and damage of each node within the symmetry plane are extracted. To improve the numerical stability and convergence performance of subsequent neural network model training, the extracted data is normalized using the MAX-MIN normalization method to obtain a database of building target penetration trajectories. The maximum and minimum parameters used in the normalization are statistically determined globally across all data types (including coordinates, velocity, and damage) in all simulation results.

[0130] The obtained database of building target penetration trajectories will be randomly divided into training, validation, and test sets at a ratio of 80%, 10%, and 10% respectively to ensure the generalization of model training and the objectivity of evaluation. The divided data will be stored in .npz format for subsequent model training and performance evaluation.

[0131] Node features include the velocity components (dimension 10) of each node in the first five time steps and the node type embedding (dimension 16), so the input dimension of the node features is 26. The hidden dimension and output dimension of the hidden layers in the multilayer perceptron are both set to 128. Edge features include the relative displacement vectors (dimension 2) and their scalar distances (dimension 1) between neighboring nodes, so the input dimension of the edge features is 3.

[0132] The target dynamics parameters include nodal acceleration components (dimension 2) and damage (dimension 1), so the input dimension of the target dynamics parameters is 3.

[0133] Set the number of message passes , , , ,but The connectivity radius is set to 0.14.

[0134] The initial learning rate is set to 1×10. -4 The decay coefficient is 0.1, causing the learning rate to gradually decrease from 1×10⁻⁶ during training. -4 Decay to approximately 1×10 -6 Furthermore, to mitigate the error accumulation problem caused by long-term iterations in autoregressive prediction scenarios, random walk noise perturbation is introduced during the training phase, with a noise mean of 0 and a standard deviation of 1×10⁻⁶. -3 The training process was performed on a single NVIDIA RTX 4090 GPU with a training batch size of 1.

[0135] The trained building penetration trajectory prediction model was evaluated using a test set, with root mean square error (RMSE) as the final performance metric. The evaluation results are as follows: Figure 5 As shown. From Figure 5 As can be seen, the trained building penetration trajectory prediction model can accurately predict the projectile's trajectory. The predicted residual velocity is highly consistent with the baseline numerical simulation results, effectively predicting the impact of velocity changes on the penetration process. Furthermore, the model can stably predict the influence of changes in the projectile's angle of impact on its trajectory deflection characteristics, demonstrating good reliability. Simultaneously, compared to traditional numerical simulation methods, this model exhibits a significant computational efficiency advantage in the inference phase; actual tests show an approximately 10% speed improvement in inference. 6 This can meet the real-time requirements of engineering projects.

[0136] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that when the computer program is executed by a processor, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0137] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A building penetration trajectory prediction method based on a graph neural network, characterized in that, The method comprises the following steps: S1. Based on key factors, a three-dimensional refined finite element model is constructed, numerical simulation is carried out through LS-DYNA software, building penetration trajectory data is obtained, the data is processed, a building target penetration trajectory database is obtained, and the database is divided into a training set, a verification set and a test set according to a proportion; S2. A building penetration trajectory prediction model is built based on a graph neural network architecture, the building penetration trajectory prediction model is trained using the training set, the performance of the model in the training is evaluated using the verification set, the training is stopped when the prediction error converges, and the trained building penetration trajectory prediction model is obtained; S3. The building penetration trajectory data is input into the trained building penetration trajectory prediction model, and the final predicted building penetration trajectory is obtained.

2. The building penetration trajectory prediction method based on a graph neural network according to claim 1, characterized in that, In step S1, the key factors include impact velocity, projectile inclination angle, projectile angle of attack, projectile diameter, projectile length-diameter ratio, projectile curvature ratio and floor thickness; A three-dimensional refined finite element model is constructed using an adaptive smoothed particle hydrodynamics method; The building penetration trajectory data includes the coordinates, velocities, equivalent stresses and equivalent plastic strains of the projectile and the building nodes; the projectile is set to remain rigid during penetration, the self-rotation of the projectile around the projectile axis is ignored, the movement of the projectile is limited to a selected two-dimensional plane, and the deflection of the projectile in directions other than the plane is ignored; The time history data of the projectile and the building nodes on the symmetry plane in the building penetration trajectory data is extracted and globally normalized to obtain the building target penetration trajectory database.

3. The graph neural network-based building penetration trajectory prediction method according to claim 1, wherein, In step S2, the building penetration trajectory prediction model comprises an encoder module, a processor module and a decoder module; the encoder module comprises a node encoding function and an edge encoding function; the processor module comprises ten cascaded graph network modules, and the graph network module comprises a node update function and an edge update function; the decoder module comprises a node decoding function.

4. The graph neural network-based building penetration trajectory prediction method according to claim 1, wherein, In step S2, the training set is input into the building penetration trajectory prediction model, the kinetic parameters are obtained through forward propagation, the mean square error is used as the loss function, the Adam optimizer is used for parameter updating, the learning rate adopts an exponential decay strategy, and the specific formula is: ； wherein, denotes the updated learning rate, denotes the initial learning rate, denotes the decay coefficient, denotes the current iteration step number, denotes the total decay step number; The performance of the building penetration trajectory prediction model is evaluated using the verification set, the root mean square error is used as the evaluation index, the training is stopped when the loss on the verification set reaches convergence, and the trained building penetration trajectory prediction model is obtained; The trained building penetration trajectory prediction model is evaluated using the test set, and the root mean square error is used as the final performance index.

5. The building penetration trajectory prediction method based on a graph neural network according to claim 3, characterized in that, In step S3, graph data is constructed based on the building penetration trajectory data, and the graph data comprises node features and edge features; The node features comprise velocity components of each node at the first five time steps and node type embedding; The edge features comprise relative displacement vectors between adjacent nodes and scalar distances; The graph data is input into the building penetration trajectory prediction model, the node features are processed through the node encoding function in the encoder module to obtain node embedding vectors, and the edge features are processed through the edge encoding function in the encoder module to obtain edge embedding vectors; The node embedding vector and the edge embedding vector are updated by a multi-layer graph network using a processor module, the neighborhood node embedding vector and the edge embedding vector are aggregated, and the updated edge embedding vector is obtained by processing through an edge update function; the target node embedding vector and the updated neighborhood edge embedding vector are aggregated, and the updated node embedding vector is obtained by processing through a node update function; The updated node embedding vector is processed by a node decoding function in the decoder module to obtain a target dynamic parameter, which includes a node acceleration component and a damage; Based on the node acceleration component, the velocity component of the node is updated using Euler integration method, and the specific formula is: ； wherein, is the velocity component of the jth node at time step t, is the velocity component of the jth node at time step t, is the acceleration component of the jth node at time step t, is the time step;​​​​​​ Based on the velocity component of the node, the coordinates of the node are updated using Euler integration method, and the specific formula is: ； wherein, is the coordinate of the th node at time is the coordinate of the th node at time ;​ Based on the updated velocity component of the node and the coordinates of the node, the updated building penetration trajectory is obtained; The updated building penetration trajectory is corrected by a rigid correction module, and the specific steps are: Step 1: Based on the initial coordinates and time of the node The node coordinates are used to calculate the initial geometric center coordinates and time. The geometric center coordinates are given by the following formula: ； ； wherein, denotes the initial geometric center coordinates of the nodes, denotes the geometric center coordinates of the nodes at time denotes the geometric center coordinates of the nodes, denotes the number of nodes, denotes the initial coordinates of the th node; Step 2: Calculate the rigid displacement of the node The specific formula is: ； Step 3: Obtain the rigid rotation angle of the node as a whole according to the principal inertia axis direction of the initial state and the current state of the node ; Step 4: According to the rigid rotation angle Constructing the two-dimensional rigid rotation matrix The specific formula is: ； Step 5: Update the coordinates of the i-th node at time t based on the rigid translation and rotation transformation, specifically: Step 5: Update the coordinates of the i-th node at time t based on the rigid translation and rotation transformation, specifically:​ ； wherein, is the modified coordinate of the th node at time t; Based on the corrected coordinates of the node, the final predicted building penetration trajectory is obtained.

6. The building penetration trajectory prediction method based on a graph neural network according to claim 5, characterized in that, The connection relationship of the neighborhood nodes in the graph data is established by using the R-radius method, the Euclidean distance between the target node and other nodes is calculated, and when the distance does not exceed the connection radius, the other node is taken as the neighborhood node of the target node, and an undirected edge is established between the two nodes. The selection standard of the connection radius is: The dynamic response among the nodes is driven by the wave propagation behavior; at each time step the processor implements the spatial information flow by message passing, the maximum propagation range of the message passing is : ； wherein, a range of single message passing for a connected radius; a number of message passes; a propagation distance of a wave within a time step in a real physical system, , , a speed of an elastic wave of a target material, , a Young's modulus of a material, a density; In The critical radius is calculated by the formula ； The value of the communication radius is not less than .

7. The building penetration trajectory prediction method based on a graph neural network according to claim 5, characterized in that, The node encoding function and the edge encoding function both include a multi-layer perception and a layer normalization layer, the multi-layer perception includes an input layer, a hidden layer and an output layer, and the hidden layer is connected with a LeakyReLU activation function; The node feature is processed by the linear affine transformation and the element-wise linear combination of the hidden layer of the multi-layer perception in the node encoding function to obtain a hidden representation, the hidden representation is processed by the non-linear mapping of the LeakyReLU activation function to obtain an active representation, the active representation is processed by the linear affine mapping of the output layer of the multi-layer perception in the node encoding function to map the hidden representation to the expected embedding dimension to obtain an output result, and the output result is processed by the standardization and the learnable affine transformation of the layer normalization layer in the node encoding function to obtain a node embedding vector; The process of obtaining the edge embedding vector by processing the edge feature by the edge encoding function is the same as the process of obtaining the node embedding vector based on the node feature.

8. The graph neural network-based building penetration trajectory prediction method according to claim 7, characterized in that, The node update function and the edge update function both include a multi-layer perception and a layer normalization layer; the multi-layer perception includes an input layer, a hidden layer and an output layer, and the hidden layer is connected with a LeakyReLU activation function; The node embedding vectors of the aggregated neighborhood nodes and the edge embedding vectors of the corresponding edges are processed by linear affine transformation and element-wise linear combination of the hidden layers of the multi-layer perceptron in the edge update function, to obtain hidden representations, the hidden representations are processed by non-linear mapping of the LeakyReLU activation function, to obtain activated representations, the activated representations are processed by linear affine mapping of the output layer of the multi-layer perceptron in the edge update function, to map the hidden representations to the desired embedding dimension, to obtain output results, the results are processed by standardization and learnable affine transformation of the layer normalization layer in the edge update function, the results are residual connected with the initial edge embedding vectors, to obtain updated edge embedding vectors; The node embedding vectors of the aggregated nodes and the updated edge embedding vectors of the adjacent edges are processed by linear affine transformation and element-wise linear combination of the hidden layers of the multi-layer perceptron in the node update function, to obtain hidden representations, the hidden representations are processed by non-linear mapping of the LeakyReLU activation function, to obtain activated representations, the activated representations are processed by linear affine mapping of the output layer of the multi-layer perceptron in the node update function, to map the hidden representations to the desired embedding dimension, to obtain output results, the results are processed by standardization and learnable affine transformation of the layer normalization layer in the node update function, the results are residual connected with the initial node embedding vectors, to obtain updated node embedding vectors.

9. The building penetration trajectory prediction method based on a graph neural network according to claim 5, characterized in that, The node decoding function comprises a multi-layer perceptron; the multi-layer perceptron comprises an input layer, a hidden layer and an output layer, and the hidden layer is connected with a LeakyReLU activation function; The updated node embedding is processed by linear affine transformation and element-wise linear combination of the hidden layers of the multi-layer perceptron in the node decoding function, to obtain hidden representations, the hidden representations are processed by non-linear mapping of the LeakyReLU activation function, to obtain activated representations, the activated representations are processed by linear affine mapping of the output layer of the multi-layer perceptron in the node decoding function, to map the hidden representations to the desired embedding dimension, to obtain target kinetic parameters.

10. A computer-readable storage medium storing a computer program, the computer-readable storage medium being characterized by, The computer program is run by the processor to perform the building penetration trajectory prediction method based on the graph neural network in any one of claims 1 to 9.

Images