Unmanned aerial vehicle electromagnetic scattering characteristic solving method and device based on graph neural network

Abstract

The invention relates to an unmanned aerial vehicle electromagnetic scattering characteristic solving method and device based on a graph neural network, and belongs to the technical field of computational electromagnetism and radar target recognition, and the method specifically comprises the steps: carrying out the initialization of a system and the discretization of an unmanned aerial vehicle grid; constructing graph structure data based on an RWG primary function, and establishing a mapping relation between an unstructured grid of a physical space and graph data which can be processed by a graph neural network; a GraphSolver model of a graph neural network solver is constructed, and an end-to-end deep neural network model is formed; training and optimizing the model; according to the method, the problems of too long calculation time and huge memory consumption when an existing full-wave numerical calculation method is used for analyzing complex targets such as an unmanned aerial vehicle are solved; and the existing deep learning method is difficult to accurately process the non-uniform triangular mesh of the unmanned aerial vehicle and is easy to lose key geometric details such as wing edges.

Description

Technical Field

[0001] This invention relates to a method and apparatus for solving the electromagnetic scattering characteristics of unmanned aerial vehicles (UAVs) based on graph neural networks, belonging to the fields of computational electromagnetics and radar target recognition technology. Background Technology

[0002] With the evolution of modern air combat, unmanned aerial vehicles (UAVs) are playing an increasingly prominent role in reconnaissance, surveillance, strike, and electronic warfare. In highly contested environments, the survivability of UAVs largely depends on the size of their radar cross-section. Therefore, low observability (stealth) design has become a core indicator in the development of advanced UAVs.

[0003] To design UAVs with low RCS characteristics, engineers need to simulate the electromagnetic scattering characteristics of thousands of different aerodynamic shapes, skin materials, and radar incident angles during the design phase. Furthermore, in wartime mission planning, it is also necessary to evaluate the dynamic RCS of the UAV relative to enemy radar at specific waypoints in real time to plan the optimal penetration path.

[0004] However, obtaining accurate electromagnetic scattering characteristics of unmanned aerial vehicles (UAVs), which have complex three-dimensional geometries (including multi-scale components such as fuselage, wings, tail, and air intakes), presents a severe challenge: First, traditional full-wave numerical algorithms are computationally inefficient. While the classic Method of Moments (MoM) combined with the Multilevel Fast Multipole Algorithm (MLFMA) offers high accuracy, its computational complexity remains high. Geometric details such as wing edges and spires of UAVs require extremely high-density triangular meshing, resulting in a massive number of unknowns. A single full-wave simulation often takes hours or even days and consumes significant amounts of memory and computing resources, failing to meet the demands of "rapid iterative optimization" in stealth design or "real-time evaluation" in mission planning.

[0005] Second, the accuracy of high-frequency approximation algorithms is insufficient. While the Physical Optics (PO) or Shooting Ray Rendering (SBR) methods are computationally fast, they neglect complex electromagnetic physical mechanisms such as traveling waves, edge diffraction, and creeping waves. The RCS of UAVs is often dominated by detailed scattering from wing edges and gaps, and high-frequency approximation algorithms have significant prediction errors in these critical areas, potentially leading to the failure of stealth design.

[0006] Third, existing deep learning models suffer from poor geometric adaptability. In recent years, deep learning has been introduced into electromagnetic prediction. However, mainstream convolutional neural networks (CNNs) can only process regular pixel or voxel data. The surface of a drone is a complex free-form surface; voxelization would produce a severe "staircase effect," losing crucial geometric details (such as sharp trailing edges) and leading to distorted RCS predictions. While PointNet can handle unstructured data, it ignores the continuous flow of surface currents (topological connectivity), making it difficult to predict accurate vector current distributions. Summary of the Invention

[0007] To address the technical problems existing in the prior art, this invention provides a method and apparatus for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks. It solves the problems of excessive computation time and huge memory consumption of existing full-wave numerical calculation methods when analyzing complex targets such as UAVs, as well as the shortcomings of existing deep learning methods in accurately handling non-uniform triangular meshes of UAVs and easily losing key geometric details such as wing edges.

[0008] To achieve the above objectives, the technical solution adopted in this invention is a method for solving the electromagnetic scattering characteristics of unmanned aerial vehicles (UAVs) based on graph neural networks, comprising the following steps: S1. Initialize the system and discretize the UAV grid; S2. Construct graph structured data based on RWG basis functions and establish a mapping relationship between unstructured grids in physical space and graph data that can be processed by graph neural networks; S3. Construct the GraphSolver model to form an end-to-end deep neural network model; S4. Train and optimize the model; S5. Use the trained GraphSolver model for fast prediction.

[0009] Preferably, step S1 includes the following steps: S101. Parameter Acquisition: Acquire the geometric model of the three-dimensional computer-aided design of the UAV to be solved, as well as the parameters of the incident electromagnetic wave from the radar. The parameters include the operating frequency. Incident direction angle and polarization mode; S102. Non-uniform meshing: In response to the curvature variation characteristics of the UAV surface, an adaptive mesh generation algorithm is used to discretize the UAV surface and generate a non-uniform triangular mesh. In particular, the mesh is automatically densified in areas with drastic curvature changes or edge diffraction, such as the leading edge and trailing edge of the wing, the air intake lip and the tail tip, while a sparse mesh is used in flat areas such as the fuselage belly, in order to control the scale of unknowns while ensuring geometric approximation accuracy. S103. Basis Function Definition: Define Rao-Wilton-Glisson (RWG) vector basis functions to expand the induced current density on the surface of the UAV, where the ... RWG basis functions Defined in the shared first Two adjacent triangular elements sharing a common edge and This is to ensure the continuity of the normal component when the current crosses the common edge.

[0010] Preferably, step S2 includes the following steps: S201, Define the set of graph nodes The strategy of constructing a dual graph using edges as nodes is adopted, mapping each RWG basis function in the UAV mesh to a node in the graph neural network; if the discretized mesh has a total of If there are 10 common edges, then the constructed graph contains 10 ... One node; S202, Define the graph edge set Based on the spatial adjacency relationship of RWG basis functions, a graph connection is established. If the triangular facets supported by two RWG basis functions share an edge or a vertex in space, it is determined that there is an electromagnetic coupling relationship between the two, and an edge connection is established between the corresponding two nodes in the graph, thereby forming a topology that can simulate the flow and diffusion path of current on the surface of the UAV conductor. S203. Extract node feature vectors: Calculate the input feature vector for each node. The feature vector integrates prior physical information and geometric location information, and the specific calculation formula is as follows: , in, For the first The coordinates of the center position of the common edge correspond to the RWG basis functions. The physical optical approximation induced current generated at this central location by the incident electromagnetic field is calculated using the following formula: , in This is the surface normal vector at that location. For free space wave impedance, and These are the incident electric field vector and the magnetic field vector, respectively. Introducing physical features aims to provide the neural network with an initial physical state that conforms to the boundary conditions of Maxwell's equations, thereby accelerating model convergence.

[0011] Preferably, the GraphSolver model in step S3 adopts an "upsampling-graph convolution-downsampling" architecture, specifically including the following steps: S301. Construct an upsampling fully connected network: used to map the low-dimensional input feature vector obtained in S203 to a high-dimensional latent feature space to enhance the nonlinear expressive power of the features; S302. Construct a graph convolutional network: Utilize continuous kernel functions to perform convolution operations on the upsampled feature map, aggregating neighborhood node information to simulate electromagnetic coupling effects. For the ... The feature update formula for each node is defined as follows: , in For trainable weight matrix, For nodes The neighborhood group, Let be the relative position vector of the neighboring nodes with respect to the center node. It is a trainable kernel function network that determines the relative position of kernels. Dynamically generate convolution weights to adapt to scale changes in the non-uniform mesh on the UAV surface; S303. Construct a downsampling fully connected network: It contains six parallel independent branches, which are used to predict the real and imaginary parts of the surface current density in the x, y, and z directions in the Cartesian coordinate system, respectively. The accuracy of vector regression is improved by decoupling the output.

[0012] Preferably, step S4 specifically includes the following steps: S401. Data preparation: A set of surface current density coefficients at different incident angles, calculated using the method of moments, are used as label data for supervised training. S402. Loss Function Definition: The loss function is defined as the mean squared error (MSE) between the predicted current density and the tag data. , in The average current density obtained using the MoM method simulation, The average current density is predicted using deep learning; S403, Training Strategy: Input the graph data generated in S2 into the GraphSolver model, and use the backpropagation algorithm and Adam optimizer to update the network parameters until the loss function converges.

[0013] Preferably, when optimizing the model training, a transfer learning strategy is adopted for the UAV target. First, the model is pre-trained on a dataset of basic geometric shapes including spheres, frustums, and cuboids to learn the general laws of electromagnetic scattering. Then, the pre-trained parameters are loaded as initial values, and the model is fine-tuned using a small amount of UAV sample data to reduce the dependence on expensive full-wave simulation data.

[0014] Preferably, step S5 specifically includes the following steps: S501, Graph Conversion: Convert the grid data of the UAV design scheme under test into the graph structure described in step S2; S502, Inference: Input the trained GraphSolver model and output the surface current density coefficient corresponding to each RWG basis function directly in milliseconds; S503. Reconstruction and Evaluation: The surface current distribution of the entire aircraft is reconstructed based on the RWG basis function expansion, and the radar cross section of the UAV is calculated using the Stratton-Chu far-field integral formula to evaluate its stealth performance.

[0015] Preferably, the feature is that it includes a device for solving the electromagnetic scattering characteristics of a UAV, the device comprising, UAV geometry processing module: configured to read the UAV geometric model, perform adaptive non-uniform triangulation, identify all RWG common edges, and construct a sparse adjacency matrix to complete the mapping from geometric space to graph space; Physical feature encoder: configured as a parallel computing unit, it calculates the physical optical approximation current features of each graph node based on the input radar parameters and grid normal vector, providing physical priors for the neural network; GraphSolver inference engine module: Configured to run the above-trained lightweight graph neural network model on a GPU workstation or cloud server, and perform forward inference to output current coefficients; Stealth Characteristics Analysis Module: Configured to map the predicted current coefficient back to the UAV's 3D model, generate a heat map of the current distribution on the entire surface of the UAV, calculate and draw the RCS polar coordinate map, and have the function of automatically identifying strong scattering sources, which is used to assist in the stealth optimization design of the UAV's shape.

[0016] Compared with the prior art, the present invention has the following technical effects: 1. The computation speed is improved by orders of magnitude. Compared with the traditional MoM method, the present invention does not require the construction and solution of dense matrices in the inference stage, and the computation time is reduced from several hours to milliseconds, making it possible to quickly screen thousands of UAV aerodynamic schemes in the early stage of design.

[0017] 2. Perfectly adaptable to complex aerodynamic shapes: By mapping RWG basis functions to graph nodes, this invention overcomes the limitation of CNNs that can only process regular grids. Combined with continuous kernel convolution, the model can accurately handle the electromagnetic response of extremely densely meshed regions such as the trailing edge and wingtip of the UAV wing, without the geometric accuracy loss caused by voxelization.

[0018] 3. Reduce data acquisition costs. The proposed "basic shape → complex UAV" migration strategy effectively utilizes low-cost data to learn physical laws. Only a small amount of expensive UAV full-wave simulation data is needed to obtain a high-precision model, which greatly reduces the threshold and cost of building AI models.

[0019] 4. Supports real-time mission planning. Due to its extremely fast computation speed, the system can be deployed on ground stations or even onboard computers to assess the probability of detection of UAVs under their current attitude and flight path in real time, and assist in dynamic trajectory planning. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the mapping relationship between the RWG basis functions on the surface of the UAV and the graph nodes, as well as the principle of graph structure construction in this invention.

[0022] Figure 3 This is a detailed schematic diagram of the architecture of the GraphSolver neural network in this invention.

[0023] Figure 4 This is a flowchart illustrating the pre-training of the basic geometry and the transfer learning of the UAV model in this invention.

[0024] Figure 5 This is a functional block diagram of the hardware architecture of the device for solving the electromagnetic scattering characteristics of UAVs in this invention. Detailed Implementation

[0025] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0026] like Figures 1 to 4 As shown, a method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks is presented. This method specifically includes the following steps: S1. Initialize the system and discretize the UAV grid; S2. Construct graph structured data based on RWG basis functions and establish a mapping relationship between unstructured grids in physical space and graph data that can be processed by graph neural networks; S3. Construct the GraphSolver model to form an end-to-end deep neural network model; S4. Train and optimize the model; S5. Use the trained GraphSolver model for fast prediction.

[0027] In step S1, during system initialization and UAV mesh discretization, S101. Parameter Acquisition: Acquire the three-dimensional computer-aided design geometric model of the UAV to be solved (such as a CAD geometric model) and the parameters of the incident electromagnetic wave from the radar. The parameters include the operating frequency. Incident direction angle And polarization mode (such as horizontal polarization HH or vertical polarization VV); S102. Non-uniform meshing: To address the curvature variation characteristics of the UAV surface, an adaptive mesh generation algorithm is used to discretize the UAV surface, generating a non-uniform triangular mesh. Specifically, the mesh is automatically refined in areas with drastic curvature changes or edge diffraction, such as the wing leading edge, trailing edge, air intake lip, and tail tip, while a sparse mesh is used in flat areas such as the fuselage belly, to control the scale of unknowns while ensuring geometric approximation accuracy. S103. Basis Function Definition: Define Rao-Wilton-Glisson (RWG) vector basis functions to expand the induced current density on the surface of the UAV, where the ... RWG basis functions Defined in the shared first Two adjacent triangular elements sharing a common edge and This is to ensure the continuity of the normal component when the current crosses the common edge.

[0028] In step S2, when constructing graph structured data based on RWG basis functions, a mapping relationship is established between the unstructured grid in physical space and the graph data that the graph neural network can process. Specifically, this includes: S201, Define the set of graph nodes The strategy employs a dual graph construction approach, using edges as nodes, to map each RWG basis function (i.e., each internal common edge) in the UAV mesh to a node in the graph neural network. If the discretized mesh contains a total of... If there are 10 common edges, then the constructed graph contains 10 ... One node; S202, Define the graph edge set : Establish graph connections based on the spatial adjacency relationship of RWG basis functions. If the triangular facets supported by two RWG basis functions share edges or vertices in space, it is determined that there is an electromagnetic coupling relationship between them, and an edge connection is established between the corresponding two nodes in the graph, thereby forming a topology that can simulate the flow and diffusion path of current on the surface of the UAV conductor; S203. Extract node feature vectors: Calculate the input feature vector for each node. The feature vector integrates prior physical information and geometric location information, and the specific calculation formula is as follows: , in, For the first The coordinates of the center position of the common edge correspond to the RWG basis functions. The approximate physical optics (PO) induced current excited at this central location by the incident electromagnetic field is calculated using the following formula: , in This is the surface normal vector at that location. For free space wave impedance, and These are the incident electric field vector and the magnetic field vector, respectively. Introducing physical features aims to provide the neural network with an initial physical state that conforms to the boundary conditions of Maxwell's equations, thereby accelerating model convergence.

[0029] In step S3, when constructing the GraphSolver model, an end-to-end deep neural network model is built. The GraphSolver model adopts an "upsampling-graph convolution-downsampling" architecture, specifically including: S301. Construct an upsampled fully connected network (U-FCN): This network is used to map the low-dimensional input feature vector (e.g., 9-dimensional) obtained in S203 to a high-dimensional latent feature space (e.g., 64-dimensional) to enhance the non-linear expressive power of the features. S302. Construct a Graph Convolutional Network (GCN): Utilize continuous kernel functions to perform convolution operations on the upsampled feature map, aggregating neighborhood node information to simulate electromagnetic coupling effects. For the... The feature update formula for each node is defined as follows: , in For trainable weight matrix, For nodes The neighborhood group, Let be the relative position vector of the neighboring nodes with respect to the center node. It is a trainable kernel network that determines the relative position of kernels. Dynamically generate convolution weights to adapt to scale changes in the non-uniform mesh on the UAV surface; S303. Construct a downsampled fully connected network (D-FCNs): It contains six parallel independent branches, which are used to predict the real and imaginary parts of the surface current density in the x, y, and z directions in the Cartesian coordinate system, respectively. The accuracy of vector regression is improved by decoupling the output.

[0030] In step S4, during model training and optimization... S401. Data preparation: A set of surface current density coefficients at different incident angles, calculated using the method of moments (MoM), are used as label data for supervised training. S402. Loss Function Definition: The loss function is defined as the mean squared error (MSE) between the predicted current density and the tag data. , in The average current density obtained using the MoM method simulation, The average current density is predicted using deep learning; S403, Training Strategy: Input the graph data generated in S2 into the GraphSolver model, and update the network parameters using the backpropagation algorithm and the Adam optimizer until the loss function converges. For UAV targets, a transfer learning strategy is preferred. First, pre-train the model on a dataset containing basic geometric shapes such as spheres, frustums, and cuboids to learn the general laws of electromagnetic scattering physics. Then, load the pre-trained parameters as initial values ​​and fine-tune the model using a small amount of UAV sample data to reduce dependence on expensive full-wave simulation data.

[0031] In step S5, when using the trained GraphSolver model for fast prediction, S501, Graph Conversion: Convert the grid data of the UAV design scheme under test into the graph structure described in S2; S502, Inference: Input the trained GraphSolver model and output the surface current density coefficient corresponding to each RWG basis function directly in milliseconds; S503. Reconstruction and Evaluation: The surface current distribution of the entire aircraft is reconstructed based on the RWG basis function expansion, and the radar cross section (RCS) of the UAV is calculated using the Stratton-Chu far-field integral formula to evaluate its stealth performance.

[0032] like Figure 5 As shown, the present invention also provides a device for solving the electromagnetic scattering characteristics of unmanned aerial vehicles based on graph neural networks, which includes the following four modules: UAV geometry processing module: configured to read the UAV CAD model, perform adaptive non-uniform triangulation, identify all RWG common edges, and construct a sparse adjacency matrix to complete the mapping from geometric space to graph space; Physical feature encoder: configured as a parallel computing unit, it calculates the physical optical approximation current features of each graph node based on the input radar parameters and grid normal vector, providing physical priors for the neural network; GraphSolver inference engine: Configured to run the above-trained lightweight graph neural network model on a GPU workstation or cloud server, and perform forward inference to output current coefficients; Stealth Characteristics Analysis Module: Configured to map the predicted current coefficient back to the UAV's 3D model, generate a heat map of the current distribution on the entire surface of the UAV, calculate and draw the RCS polar coordinate map, and have the function of automatically identifying strong scattering sources (such as specular reflection points and edge diffraction points) to assist in the stealth optimization design of the UAV's shape.

[0033] This invention achieves real-time, high-precision prediction of surface current distribution and radar cross section of UAVs with complex aerodynamic shapes by constructing a graph structure mapping mechanism based on RWG basis functions, combining it with a graph neural network model incorporating physical prior features, and utilizing a transfer learning strategy. Furthermore, by combining the physical rigor of computational electromagnetics with the efficiency of deep learning, a fast graph neural network solver, GraphSolver, specifically designed for complex geometric targets of UAVs, is proposed. This method maps the RWG basis functions describing the UAV surface to graph nodes and introduces a transfer learning mechanism, using a pre-trained model with simple geometry to accelerate the solution of complex UAV models, thus solving the technical bottleneck of the trade-off between "high precision" and "high efficiency" in UAV stealth design.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.

Claims

1. A method for solving the electromagnetic scattering characteristics of unmanned aerial vehicles (UAVs) based on graph neural networks, characterized in that: Includes the following steps, S1. Initialize the system and discretize the UAV grid; S2. Construct graph structured data based on RWG basis functions and establish a mapping relationship between unstructured grids in physical space and graph data that can be processed by graph neural networks; S3. Construct the GraphSolver model to form an end-to-end deep neural network model; S4. Train and optimize the model; S5. Use the trained GraphSolver model for fast prediction.

2. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 1, characterized in that: Step S1 includes the following steps: S101. Parameter Acquisition: Acquire the geometric model of the three-dimensional computer-aided design of the UAV to be solved, as well as the parameters of the incident electromagnetic wave from the radar. The parameters include the operating frequency. Incident direction angle and polarization mode; S102. Non-uniform meshing: In response to the curvature variation characteristics of the UAV surface, an adaptive mesh generation algorithm is used to discretize the UAV surface and generate a non-uniform triangular mesh. In particular, the mesh is automatically densified in areas with drastic curvature changes or edge diffraction, such as the leading edge and trailing edge of the wing, the air intake lip and the tail tip, while a sparse mesh is used in flat areas such as the fuselage belly, in order to control the scale of unknowns while ensuring geometric approximation accuracy. S103. Basis Function Definition: Define Rao-Wilton-Glisson (RWG) vector basis functions to expand the induced current density on the surface of the UAV, where the ... RWG basis functions Defined in the shared first Two adjacent triangular elements sharing a common edge and This is to ensure the continuity of the normal component when the current crosses the common edge.

3. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 1, characterized in that: Step S2 includes the following steps: S201, Define the set of graph nodes The dual graph construction strategy of "edge as node" is adopted to map each RWG basis function in the UAV mesh to a node in the graph neural network; If the discretized grid contains a total of If there are 10 common edges, then the constructed graph contains 10 ... One node; S202, Define the graph edge set Based on the spatial adjacency relationship of RWG basis functions, a graph connection is established. If the triangular facets supported by two RWG basis functions share an edge or a vertex in space, it is determined that there is an electromagnetic coupling relationship between the two, and an edge connection is established between the corresponding two nodes in the graph, thereby forming a topology that can simulate the flow and diffusion path of current on the surface of the UAV conductor. S203. Extract node feature vectors: Calculate the input feature vector for each node. The feature vector integrates prior physical information and geometric location information, and the specific calculation formula is as follows: ， in, For the first The coordinates of the center position of the common edge correspond to the RWG basis functions. The physical optical approximation induced current generated at this central location by the incident electromagnetic field is calculated using the following formula: ， in This is the surface normal vector at that location. For free space wave impedance, and These are the incident electric field vector and the magnetic field vector, respectively. Introducing physical features aims to provide the neural network with an initial physical state that conforms to the boundary conditions of Maxwell's equations, thereby accelerating model convergence.

4. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 3, characterized in that: The GraphSolver model in step S3 adopts an "upsampling-graph convolution-downsampling" architecture, specifically including the following steps: S301. Construct an upsampling fully connected network: used to map the low-dimensional input feature vector obtained in S203 to a high-dimensional latent feature space to enhance the nonlinear expressive power of the features; S302. Construct a graph convolutional network: Utilize continuous kernel functions to perform convolution operations on the upsampled feature map, aggregating neighborhood node information to simulate electromagnetic coupling effects. For the ... The feature update formula for each node is defined as follows: ， in, For trainable weight matrix, For nodes The neighborhood group, Let be the relative position vector of the neighboring nodes with respect to the center node. It is a trainable kernel function network that determines the relative position of kernels. Dynamically generate convolution weights to adapt to scale changes in the non-uniform mesh on the UAV surface; S303. Construct a downsampling fully connected network: It contains six parallel independent branches, which are used to predict the real and imaginary parts of the surface current density in the x, y, and z directions in the Cartesian coordinate system, respectively. The accuracy of vector regression is improved by decoupling the output.

5. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 4, characterized in that: Step S4 specifically includes the following steps: S401. Data preparation: A set of surface current density coefficients at different incident angles, calculated using the method of moments, are used as label data for supervised training. S402. Loss Function Definition: The loss function is defined as the mean squared error (MSE) between the predicted current density and the tag data. ， in The average current density obtained using the MoM method simulation, The average current density is predicted using deep learning; S403, Training Strategy: Input the graph data generated in S2 into the GraphSolver model, and use the backpropagation algorithm and Adam optimizer to update the network parameters until the loss function converges.

6. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 5, characterized in that: When optimizing the model training, a transfer learning strategy is adopted for UAV targets. First, the model is pre-trained on a dataset of basic geometric shapes including spheres, frustums, and cuboids to learn the general laws of electromagnetic scattering. Then, the pre-trained parameters are loaded as initial values, and the model is fine-tuned using a small amount of UAV sample data to reduce the dependence on expensive full-wave simulation data.

7. The method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to claim 5, characterized in that: Step S5 specifically includes the following steps: S501, Graph Conversion: Convert the grid data of the UAV design scheme under test into the graph structure described in step S2; S502, Inference: Input the trained GraphSolver model and output the surface current density coefficient corresponding to each RWG basis function directly in milliseconds; S503. Reconstruction and Evaluation: The surface current distribution of the entire aircraft is reconstructed based on the RWG basis function expansion, and the radar cross section of the UAV is calculated using the Stratton-Chu far-field integral formula to evaluate its stealth performance.

8. A method for solving the electromagnetic scattering characteristics of UAVs based on graph neural networks according to any one of claims 1-7, characterized in that: This includes a device for solving the electromagnetic scattering characteristics of unmanned aerial vehicles (UAVs), the device comprising, UAV geometry processing module: configured to read the UAV geometric model, perform adaptive non-uniform triangulation, identify all RWG common edges, and construct a sparse adjacency matrix to complete the mapping from geometric space to graph space; Physical feature encoder: configured as a parallel computing unit, it calculates the physical optical approximation current features of each graph node based on the input radar parameters and grid normal vector, providing physical priors for the neural network; GraphSolver inference engine module: Configured to run the above-trained lightweight graph neural network model on a GPU workstation or cloud server, and perform forward inference to output current coefficients; Stealth Characteristics Analysis Module: Configured to map the predicted current coefficient back to the UAV's 3D model, generate a heat map of the current distribution on the entire surface of the UAV, calculate and draw the RCS polar coordinate map, and have the function of automatically identifying strong scattering sources, which is used to assist in the stealth optimization design of the UAV's shape.

Images