CN-121997780-A - Unmanned aerial vehicle electromagnetic scattering characteristic solving method and device based on graph neural network

CN121997780ACN 121997780 ACN121997780 ACN 121997780ACN-121997780-A

Abstract

The invention relates to a method and a device for solving unmanned aerial vehicle electromagnetic scattering characteristics based on a graph neural network, which belong to the technical fields of computational electromagnetics and radar target identification, and specifically comprise the steps of initializing a system and discretizing unmanned aerial vehicle grids, constructing graph structure data based on RWG basis functions, establishing a mapping relation between unstructured grids of a physical space and graph data which can be processed by the graph neural network, constructing a graph neural network solver GraphSolver model to form an end-to-end deep neural network model, training and optimizing the model, and utilizing the trained GraphSolver model to conduct rapid prediction.

Inventors

YANG ZHILIANG
MAO YURU
WANG XINGWANG
Du Changgao
SUN XINGLI
BAI JIANSHENG
YAO JINJIE
AN JIANPING

Assignees

中北大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (8)

1. The unmanned aerial vehicle electric scattering characteristic solving method based on the graph neural network is characterized by comprising the following steps of, S1, initializing a system and discretizing an unmanned aerial vehicle grid; s2, constructing graph structure data based on RWG basis functions, and establishing a mapping relation between unstructured grids of a physical space and graph data which can be processed by a graph neural network; S3, constructing a graph neural network solver GraphSolver model to form an end-to-end deep neural network model; S4, training and optimizing the model; s5, performing quick prediction by using the trained GraphSolver model.
2. The method for solving the electro-mechanical scattering characteristics of the unmanned aerial vehicle based on the graphic neural network according to claim 1, wherein the step S1 comprises the following steps, S101, acquiring a geometric model of three-dimensional computer aided design of the unmanned aerial vehicle to be solved and parameters of radar incident electromagnetic waves, wherein the parameters comprise working frequency Angle of incidence A polarization mode; S102, non-uniform subdivision, namely discretizing the surface of the unmanned aerial vehicle by adopting a self-adaptive grid generation algorithm according to curvature change characteristics of the surface of the unmanned aerial vehicle to generate a non-uniform triangular grid, wherein the grid is automatically encrypted in areas with severe curvature changes or edge diffraction generation at the front edge, the rear edge, the lip mouth of an air inlet channel, the tip end of a tail wing and the like of the wing, and sparse grids are adopted in flat areas such as the belly of a fuselage so as to control the scale of unknown numbers on the premise of ensuring geometric approximation precision; S103, defining a Rao-Wilton-Glisson (RWG) vector basis function for expanding induction current density of the surface of the unmanned aerial vehicle, wherein the first is Individual RWG basis functions Definition on shared first Two adjacent triangular surface elements of a common edge And To ensure normal component continuity when the current crosses the common edge.
3. The method for solving the electro-mechanical scattering characteristics of the unmanned aerial vehicle based on the graphic neural network according to claim 1, wherein the step S2 comprises the following steps, S201, defining a graph node set Mapping each RWG basis function in the unmanned plane grid into one section in the graph neural network by adopting a dual graph construction strategy taking edges as nodes, and if the two sections are shared in the discretized grid Common edges, then the constructed graph contains A plurality of nodes; s202, defining a graph edge set Establishing graph connection according to the spatial adjacent relation of RWG base functions, if triangular surface elements supported by two RWG base functions have shared edges or shared vertexes in space, judging that the two have electromagnetic coupling relation, and establishing edge connection between two corresponding nodes in the graph, so that a topological structure capable of simulating current flowing and diffusing paths on the surface of the unmanned plane conductor is formed; s203, extracting node feature vectors, and calculating input feature vectors for each node The feature vector fuses physical priori information and geometric position information, and a specific calculation formula is as follows: , Wherein, the Is the first The RWG basis functions correspond to the center position coordinates of the common edge, For the physical optical approximation induced current excited by the incident electromagnetic field at the center position, the calculation formula is: , Wherein the method comprises the steps of As the normal vector of the surface at this point, As the impedance of the free-space wave, And The physical characteristics are introduced for the incident electric field vector and the magnetic field vector respectively, so as to provide an initial physical state conforming to the boundary condition of the Maxwell equation set for the neural network and accelerate the convergence of the model.
4. The unmanned aerial vehicle electro-mechanical scattering characteristic solving method based on the graph neural network of claim 3, wherein the GraphSolver model in the step S3 adopts an up-sampling-graph convolution-down-sampling architecture, and specifically comprises the following steps: s301, constructing an up-sampling full-connection network, wherein the up-sampling full-connection network is used for mapping the low-dimensional input feature vector obtained in the step S203 to a high-dimensional potential feature space so as to enhance the nonlinear expression capability of the features; s302, constructing a graph convolution network, namely performing convolution operation on the up-sampled feature graph by utilizing a continuous kernel function, aggregating neighborhood node information to simulate an electromagnetic coupling effect, and regarding the first step The individual nodes, characterized by an update formula defined as: , Wherein, the In order for the weight matrix to be trainable, Is a node Is used to determine the neighbor set of a neighbor, As a relative position vector of the neighboring node with respect to the center node, Is a trainable kernel function network based on relative position Dynamically generating convolution weights so as to adapt to the scale change of the non-uniform grid on the surface of the unmanned aerial vehicle; S303, constructing a down-sampling full-connection network, wherein the down-sampling full-connection network comprises six parallel independent branches which are respectively used for predicting the real part and the imaginary part of three components of the surface current density in the x direction, the y direction and the z direction under a Cartesian coordinate system, and improving the vector regression accuracy through decoupling output.
5. The unmanned aerial vehicle electro-mechanical scattering characteristic solving method based on the graph neural network of claim 4, wherein in the step S4, the method specifically comprises the following steps: S401, preparing data, namely using a group of surface current density coefficients under different incident angles calculated by a moment method as tag data of supervision training; s402, defining a loss function as a Mean Square Error (MSE) between the predicted current density and the tag data: , Wherein the method comprises the steps of To simulate the resulting average current density using the MoM method, Average current density predicted for use with deep learning; s403, training a strategy, namely inputting the graph data generated in the step S2 into a GraphSolver model, and updating network parameters by using a back propagation algorithm and an Adam optimizer until the loss function converges.
6. The unmanned aerial vehicle electromagnetic scattering characteristic solving method based on the graphic neural network is characterized in that when model training optimization is carried out, a transfer learning strategy is adopted for an unmanned aerial vehicle target, firstly, a model is pre-trained on a basic geometric data set comprising a sphere, a round table and a cuboid, so that a universal electromagnetic scattering physical rule is learned, then pre-training parameters are loaded as initial values, and a small amount of unmanned aerial vehicle sample data is utilized for fine adjustment of the model, so that dependence on expensive full-wave simulation data is reduced.
7. The unmanned aerial vehicle electro-mechanical scattering characteristic solving method based on the graph neural network of claim 5, wherein in the step S5, the method specifically comprises the following steps: S501, converting the grid data of the unmanned aerial vehicle design scheme to be tested into a graph structure described in the step S2; s502, reasoning, namely inputting a trained GraphSolver model, and directly outputting a surface current density coefficient corresponding to each RWG basis function in millisecond time; S503, reconstructing and evaluating, namely reconstructing the surface current distribution of the whole machine according to the RWG basis function expansion, and calculating the radar scattering cross section of the unmanned aerial vehicle by using a Stratton-Chu far-field integral formula so as to evaluate the stealth performance of the unmanned aerial vehicle.
8. The method for solving the electromagnetic scattering characteristics of the unmanned aerial vehicle based on the graphic neural network according to any one of claims 1 to 7, wherein the method comprises an unmanned aerial vehicle electromagnetic scattering characteristic solving device, The unmanned aerial vehicle geometric processing module is configured to read an unmanned aerial vehicle geometric model, execute self-adaptive non-uniform triangulation, identify all RWG public edges and construct a sparse adjacency matrix to complete mapping from geometric space to graph space; the physical characteristic encoder is configured as a parallel computing unit and is used for computing the physical optical approximate current characteristics of each graph node according to the input radar parameters and grid normal vectors and providing physical prior for the neural network; GraphSolver the reasoning engine module is configured to be carried on a GPU workstation or a cloud server, run the trained lightweight graph neural network model and execute forward reasoning to output a current coefficient; the stealth characteristic analysis module is configured to map the predicted current coefficient back to the three-dimensional model of the unmanned aerial vehicle, generate a full-plane surface current distribution thermodynamic diagram, calculate and draw an RCS polar coordinate diagram, have the function of automatically identifying a strong scattering source, and are used for assisting the stealth optimization design of the appearance of the unmanned aerial vehicle.

Description

Unmanned aerial vehicle electromagnetic scattering characteristic solving method and device based on graph neural network Technical Field The invention relates to a method and a device for solving unmanned aerial vehicle electromagnetic scattering characteristics based on a graph neural network, and belongs to the technical fields of computational electromagnetics and radar target identification. Background With the evolution of modern air combat modes, unmanned aerial vehicles increasingly play a role in reconnaissance, surveillance, battle and electronic countermeasure. In a strong challenge environment, the viability of a drone depends largely on the size of its radar cross section. Therefore, low detectability (stealth) design is a core indicator of advanced unmanned aerial vehicle development. In order to design an unmanned aerial vehicle with low RCS characteristics, engineers need to simulate electromagnetic scattering characteristics for thousands of different aerodynamic profile solutions, different skin materials, and different radar incidence angles during the design phase. In addition, in mission planning at war time, it is also necessary to evaluate the dynamic RCS of the unmanned aerial vehicle at a specific waypoint relative to the enemy radar in real time to plan an optimal burst path. However, for such targets with complex three-dimensional geometries (including multi-scale components such as fuselage, wings, tail, air intake, etc.) of unmanned aerial vehicles, obtaining accurate electromagnetic scattering properties thereof presents serious challenges: first, conventional full-wave numerical algorithms are computationally inefficient. The classical moment method (Method of Moments, moM) combined with the multi-layer fast multipole sub-algorithm (MLFMA) is high in accuracy but still high in computational complexity. Geometric details such as wing edges, peaks and the like of the unmanned aerial vehicle require extremely high-density triangulation, and the number of unknowns is huge. The single full-wave simulation often requires several hours or even days, consumes a large amount of memory and computing resources, and cannot meet the requirements of 'fast iterative optimization' in stealth design or 'real-time evaluation' in task planning. Second, the high frequency approximation algorithm is not accurate enough. Although the physical optical method (PO) or the bouncing ray method (SBR) has high calculation speed, complex electromagnetic physical mechanisms such as traveling wave, edge diffraction, creeping wave and the like are ignored. The RCS of the unmanned aerial vehicle is often dominated by detail scattering of wing edges, gaps and the like, and prediction errors of the high-frequency approximation algorithm at the key parts are large, so that stealth design failure can be caused. Third, existing deep learning models have poor geometric adaptability. In recent years, deep learning has been introduced into electromagnetic prediction. However, the mainstream Convolutional Neural Network (CNN) can only process regular pixel or voxel data. The surface of the unmanned aerial vehicle is a complex free-form surface, and if the voxelized treatment is adopted, a serious step effect is generated, and key geometric details (such as a sharp trailing edge) are lost, so that RCS prediction distortion is caused. The point cloud network (PointNet), while capable of processing unstructured data, ignores the continuous flow of surface currents (topological connections), making it difficult to predict accurate vector current distribution. Disclosure of Invention The invention provides a method and a device for solving the electromagnetic scattering characteristics of an unmanned aerial vehicle based on a graph neural network, which solve the problems of overlong calculation time and huge memory consumption when analyzing complex targets such as the unmanned aerial vehicle and the like in the existing full-wave numerical calculation method, and the defects that the existing deep learning method is difficult to accurately process the non-uniform triangular grid of the unmanned aerial vehicle, and key geometric details such as wing edges are easy to lose. In order to achieve the above purpose, the technical scheme adopted by the invention is an unmanned aerial vehicle electric scattering characteristic solving method based on a graph neural network, comprising the following steps, S1, initializing a system and discretizing an unmanned aerial vehicle grid; s2, constructing graph structure data based on RWG basis functions, and establishing a mapping relation between unstructured grids of a physical space and graph data which can be processed by a graph neural network; S3, constructing a graph neural network solver GraphSolver model to form an end-to-end deep neural network model; S4, training and optimizing the model; s5, performing quick prediction by using the trained GraphSolver model. Preferably, the s