An intelligent calculation method for the electromagnetic characteristics of radar three-dimensional targets

Abstract

This application relates to an intelligent calculation method for the electromagnetic characteristics of radar three-dimensional targets. The method includes: obtaining a radar three-dimensional target; performing an equivalent modeling of the radar three-dimensional target using a discrete dipole array; constructing a moment information neural network model; predicting the current coefficients of the discrete dipole array obtained by modeling according to the moment information neural network model; and calculating the radiation field based on the electromagnetic field superposition principle by combining the current coefficients on the basis of the geometric characteristic parameters of the discrete dipole array, so as to realize the equivalent reconstruction of the electromagnetic characteristics of the target space. Using this method can improve the rapid prediction of the electromagnetic characteristics of specific radar targets.

Description

Technical Field

[0001] This application relates to the technical field of radar target imaging, and particularly to an intelligent calculation method for radar three-dimensional target electromagnetic characteristics. Background Art

[0002] In the field of radar target imaging, accurately obtaining the electromagnetic characteristics of a target is crucial for achieving high-resolution and high-precision imaging. As a computer simulation technology, electromagnetic full-wave numerical calculation plays an indispensable role in this field. By using numerical discretization methods to solve the Maxwell equations and the corresponding boundary conditions satisfied by the target, it can accurately give the electromagnetic reflection and scattering characteristics of radar targets with different materials and shapes under electromagnetic wave irradiation. These characteristic data are the core basis for constructing radar target images and performing target recognition and analysis, and have extremely high engineering application value and academic research significance in many fields such as reconnaissance, aerospace, and civilian security. However, when current radar target imaging uses traditional electromagnetic numerical algorithms, many shortcomings are exposed. The traditional algorithm solves the Maxwell equations based on computer discretization technology, and this solution mode determines that it is inevitably limited by the calculation scale and calculation resources. When performing imaging calculations on long-distance and large-size radar targets, due to the large number of target scatterers and complex calculation grid division, the amount of calculation increases exponentially, not only consuming a large amount of calculation resources, but also significantly increasing the calculation time, seriously affecting the timeliness of imaging.

[0003] When optimizing multi - band and multi - polarization imaging parameters in a radar target imaging system and adjusting imaging algorithms for radar targets with complex structures and special materials, it is necessary to repeatedly seek the global optimal solution of imaging parameters in a multi - variable domain. This requires recalculating the electromagnetic response characteristics of the target every time the parameters change. When using traditional electromagnetic numerical calculation methods to participate in this optimization iteration process, due to the cumbersome calculation process and long calculation time, it has become a bottleneck restricting the overall design efficiency and imaging quality improvement of the radar target imaging system. In recent years, the rapid development of artificial intelligence (AI) technology has brought new solutions to the field of radar target imaging, promoting the rise of the emerging research direction of "intelligent electromagnetic calculation". A large number of past practices have shown that AI technology has powerful data - mining capabilities and can discover the internal relationships between massive data. In the field of radar target imaging, as long as enough electromagnetic simulation data is accumulated, the direct correlation mapping relationship between the geometric parameters, material properties of the radar target and its electromagnetic scattering characteristics and imaging results can be mined and established with the help of AI technology, realizing the rapid prediction of the imaging characteristics of specific radar targets. Intelligent electromagnetic calculation is expected to break through the limitation of traditional electromagnetic calculation relying on numerically solving Maxwell's equations, get rid of the constraints of calculation scale and resources on the algorithm, decouple the algorithm from specific imaging problems, and significantly improve the generality, autonomy and efficiency of the imaging algorithm, providing a new path for the innovation of radar target imaging technology. However, at present, the full application of AI technology in radar target imaging still faces many problems. On the one hand, how to obtain high - quality, large - scale electromagnetic simulation data covering various scenarios to train an excellent - performance AI model is the primary challenge. On the other hand, how to ensure the stability and accuracy of the AI model under different environments and complex target conditions and avoid model over - fitting still needs in - depth research. In addition, the compatibility problem between AI algorithms and existing radar target imaging systems also needs to be solved urgently. Summary of the Invention

[0004] Based on this, it is necessary to provide an intelligent calculation method for the electromagnetic characteristics of radar three - dimensional targets to achieve the rapid prediction of the imaging characteristics of specific radar targets in view of the above - mentioned technical problems.

[0005] An intelligent calculation method for the electromagnetic characteristics of radar three - dimensional targets, the method comprising:

[0006] Obtain a radar three - dimensional target; perform equivalent modeling of the radar three - dimensional target using a discrete dipole array;

[0007] Construct a moment - of - momentum information neural network model; predict the current coefficients of the discrete dipole array obtained by the modeling according to the moment - of - momentum information neural network model;

[0008] According to the principle of electromagnetic field superposition, the radiation field of the discrete dipole array is calculated based on the geometric characteristic parameters of the discrete dipole array and the current coefficient, so as to achieve the equivalent reconstruction of the electromagnetic characteristics of the target space.

[0009] The above-mentioned intelligent calculation method for electromagnetic characteristics of radar three-dimensional targets first performs discrete dipole array equivalent modeling on the radar three-dimensional target. By utilizing the equivalent characteristics of triangulated grids and dipoles, the three-dimensional target is equivalent to a discrete dipole array, and the geometric characteristics of the target are converted into two parameters: the position of the dipole array and the normalized dipole moment, to achieve the mapping of geometric characteristics to data space and provide input data for the subsequent moment information neural network model. Secondly, a moment information neural network model is constructed. This model innovatively adopts a moment feature extraction method based on quadratic calculation, incorporates high-order statistical features such as data autocoupling / mutual coupling, makes up for the shortcomings of traditional convolutional neural networks, and combines high-order features with traditional convolutional neural networks through additive neurons, so that the model can mine more data features, enhance the generalization ability of the model, and improve its adaptability to different complex targets. Finally, the moment information neural network model is used to predict the current coefficient of the discrete dipole array, and the radiation field of the discrete dipole array is calculated based on the principle of electromagnetic field superposition, so as to achieve equivalent reconstruction of the electromagnetic characteristics of the target space, quickly obtain the electromagnetic characteristics of specific radar targets, improve computing efficiency, and meet the timeliness requirements of radar target imaging. This application incorporates the grid division information and geometric characteristics of the target into the input data; introduces the analytical formula of the dipole radiation field in the post-processing of the output data, and uses the vector synthesis characteristics of the electromagnetic field to reduce the absolute accuracy requirements for the tiny amount of the dipole current coefficient. By creating a "physics-data" dual-driven intelligent electromagnetic computing method, the model's dependence on a large amount of training data is reduced, and the prediction accuracy and generalization ability are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of a flow chart of an intelligent calculation method for electromagnetic characteristics of radar three-dimensional targets in one embodiment;

[0011] Figure 2 A schematic diagram of the construction of a moment information neural network model in one embodiment. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0013] In one embodiment, Figure 1 As shown, an intelligent calculation method for electromagnetic characteristics of radar three-dimensional targets is provided, comprising the following steps:

[0014] Step 102, obtain the radar three-dimensional target; perform equivalent modeling of the radar three-dimensional target using a discrete dipole array.

[0015] Utilize the good simulation characteristics of the triangular meshing grid for three-dimensional targets and the equivalent characteristics between triangular meshes and dipoles to transform any three-dimensional radar target into a discrete dipole array for analysis. This equivalent modeling method can simplify the processing of complex three-dimensional targets and realize the transformation from geometric space to data space. Compared with directly performing traditional electromagnetic numerical calculations on three-dimensional targets, the computational complexity based on the discrete dipole array model is lower, reducing the demand for computing resources.

[0016] Step 104, construct a moment information neural network model; predict the current coefficients of the discrete dipole array obtained by modeling according to the moment information neural network model.

[0017] A moment information neural network based on quadratic form calculation and convolutional neural network is designed. Through the proposed moment feature extraction method, it can effectively incorporate high-order statistical features such as data self-coupling / mutual coupling. For the analysis of the electromagnetic characteristics of array structures, there are self-coupling / mutual coupling effects between different array elements, and traditional convolutional neural networks cannot mine these high-order features. However, this model incorporates high-order features into the traditional convolutional neural network through additive neurons, enabling the network to not only mine the first-order linear features of the data but also mine high-order cross features, thus providing support for constructing a more generalizable intelligent electromagnetic calculation architecture, improving the adaptability of the model to different scenarios and complex targets, and helping to solve the limitations of traditional convolutional neural networks in the face of complex multi-target coupling scenarios.

[0018] Step 106, according to the principle of electromagnetic field superposition, calculate the radiation field of the discrete dipole array by combining the current coefficients based on the geometric feature parameters of the discrete dipole array, and realize the equivalent reconstruction of the electromagnetic characteristics of the target space.

[0019] Predict the current coefficients of the discrete dipole array according to the moment information neural network model, and then calculate the radiation field based on the discrete dipole array according to the principle of electromagnetic field superposition, and finally realize the equivalent reconstruction of the electromagnetic characteristics of the target space. In this way, the electromagnetic characteristics of the radar three-dimensional target can be quickly obtained, avoiding the complex process of solving the Maxwell equations in traditional electromagnetic numerical calculations, greatly improving the calculation efficiency, meeting the requirements of radar target imaging for timeliness, and at the same time providing an effective way to solve the problem of long calculation time in the imaging calculation of large-size targets by traditional algorithms. At the same time, intelligent electromagnetic calculation is a new electromagnetic calculation method developed in the past two or three years. At present, the main application objects are still limited to two-dimensional ideal problems or three-dimensional symmetric problems, and there is less research on arbitrary three-dimensional targets. This application cleverly uses the good simulation characteristics of triangular meshes for arbitrary three-dimensional targets and the equivalent characteristics of triangular meshes with dipoles to innovatively establish an intelligent electromagnetic calculation model based on a three-dimensional randomly distributed dipole array, which is expected to expand the intelligent electromagnetic calculation technology from two-dimensional problems and three-dimensional symmetric problems to arbitrary three-dimensional problems. On the other hand, this application proposes an intelligent electromagnetic calculation method driven by both "physics-data". The input data incorporates the mesh dissection information and geometric features of the target, and at the same time incorporates the analytical formula of the dipole radiation field in subsequent calculations, implicitly constraining the physical characteristics of the electromagnetic far-field radiation. From the perspective of vector field synthesis, the prediction error of the small current term has a weak impact on the calculation result of the synthesized total field, reducing the stringent requirements of the model for the absolute accuracy of the small current coefficient terms. This calculation method effectively reduces the dependence of the model on a large amount of training data, improves the prediction accuracy and generalization ability of the model, solves the efficiency problem caused by complex calculation and long time in the optimization iteration process of traditional methods, and the stability and accuracy problems of the model under different environments and complex target conditions.

[0020] The above-mentioned intelligent calculation method for electromagnetic characteristics of radar three-dimensional targets first performs discrete dipole array equivalent modeling on radar three-dimensional targets. By utilizing the good simulation characteristics of triangulated grids for any three-dimensional targets and the equivalent characteristics of dipoles, the complex three-dimensional targets are simplified and equivalent to discrete dipole arrays, and the geometric characteristics of the targets are converted into two parameters: the position of the dipole array and the normalized dipole moment, so as to realize the mapping of geometric characteristics to data space and provide input data for the moment information neural network. Secondly, a moment information neural network model is constructed. This model innovatively adopts a moment feature extraction method based on quadratic calculation, incorporates high-order statistical features such as data autocoupling / mutual coupling, makes up for the shortcomings of traditional convolutional neural networks, and combines high-order features with traditional convolutional neural networks through additive neurons, so that the model can mine more data features, enhance the generalization ability of the model, and improve its adaptability to different complex targets. Finally, the moment information neural network model is used to predict the current coefficient of the discrete dipole array, and the radiation field of the discrete dipole array is calculated by combining the electromagnetic field superposition principle to achieve equivalent reconstruction of the electromagnetic characteristics of the target space, quickly obtain the target electromagnetic characteristics, improve computing efficiency, and meet the timeliness requirements of radar target imaging. This application incorporates the grid division information and geometric characteristics of the target into the input data; introduces the dipole radiation field analytical formula in the post-processing of the output data, and uses the vector synthesis characteristics of the electromagnetic field to reduce the absolute accuracy requirements for the tiny amount of the dipole current coefficient. By creating a "physics-data" dual-driven intelligent electromagnetic computing method, the model's dependence on a large amount of training data is reduced, and the prediction accuracy and generalization ability are improved.

[0021] In one embodiment, discrete dipole array equivalent modeling is performed on a radar three-dimensional target, including:

[0022] The radar 3D target is meshed with unstructured triangular meshes using a meshing algorithm. All the obtained triangular meshes are clustered. Every two triangular meshes with common edges define a dipole. The established dipoles are numbered and recorded as , N is the sum of the numbers of all discrete dipoles;

[0023] Define the position of each dipole as the center of the common edge of the pair of triangle meshes. n The position of the dipole is r n , define the normalized dipole moment of each dipole as the product of the three-dimensional vector pointing from the center of gravity of the negative film to the center of gravity of the positive film and the length of the common side, record n The normalized dipole moment of a dipole is ,in, l n Indicates n The length of the common side of the dipoles; Indicatesn The centroid coordinates of the positive / negative triangular patches of a dipole.

[0024] In a specific embodiment, for a specific target, an unstructured triangular mesh division is performed on it using a mature mesh division algorithm. This process can be implemented with the help of existing software, such as Feko, Ansys ICEM CFD, etc. Then, all discrete nodes are extracted and recorded, that is, the three-dimensional spatial coordinates of all triangular mesh vertices. Cluster all the obtained triangular meshes. Each pair of triangular meshes with a common edge can define a dipole model. Randomly specify one triangular patch as the positive patch and the other triangular patch as the negative patch. Number the established dipole models in sequence, denoted as ( N which is the total sum of the number of all discrete dipoles). Define the position of each dipole as the center of the common edge of this pair of triangular meshes. Denote the position of the n th dipole as r n . Define the normalized dipole moment of each dipole as the product of the three-dimensional vector pointing from the centroid of the negative patch to the centroid of the positive patch and the length of the common edge. Denote the normalized dipole moment of the n th dipole as . Among them, l n represents the length of the common edge of the n th dipole; represents the centroid coordinates of the positive / negative triangular patches of the n th dipole. Then, according to the node coordinates of the triangular mesh, the physical structure information such as the position and normalized dipole moment of each dipole is extracted respectively.

[0025] In one of the embodiments, constructing the moment information neural network model includes a moment feature extraction module, a convolutional feature extraction module, and a fully connected layer.

[0026] In one of the embodiments, the moment feature extraction module is used to extract the statistical features of the input sample data and output them after combination through additive neurons; the convolutional feature extraction module is used to refine the abstract correlation features from the statistical features of the sample data through a traditional convolutional neural network architecture; the fully connected layer is used to directly map and output the abstract correlation features and the current coefficients.

[0027] In one of the embodiments, the statistical features include first-order moment features and second-order moment features; the moment feature extraction module is used to extract the statistical features of the input sample data, including:

[0028] The extraction of the statistical features of the input sample data is respectively:

[0029] ;

[0030] ;

[0031] Among them, is the parameter to be estimated, represents the input sample data, and T represents the transpose operation.

[0032] In one of the embodiments, the calculation process of the parameter to be estimated includes:

[0033] Establish a gradient calculation formula, and solve the gradient calculation formula according to the gradient descent algorithm to obtain the value of the parameter to be estimated; the process of establishing the gradient calculation formula includes:

[0034] Denote the loss function as L , then L The gradient with respect to is:

[0035] ;

[0036] From it can be obtained that , and , so:

[0037] ;

[0038] In the above formula, is obtained by the backpropagation algorithm of the traditional CNN;

[0039] Similarly, it can be obtained that L The gradient with respect to is:

[0040] ;

[0041] In the above formula, represents the vectorization operation rule of the sample data.

[0042] In one of the embodiments, the current coefficients of the discrete dipole array predicted and modeled according to the moment information neural network model include:

[0043] Encode the parameter information of the modeled discrete dipole array and the external irradiation wave information in a data format as input data and input it into the moment information neural network model for current coefficient prediction; the data format is:

[0044] ;

[0045] Among them, represents the external irradiation wave information.

[0046] In a specific embodiment, the model needs to use a neural network to intelligently predict the induced current coefficient of each dipole array under the irradiation of external electromagnetic waves according to the physical characteristic parameters of the discrete dipole array. First, the input / output data sets of the model are set. The input set contains two parts of information. One is the external irradiation wave information, and the other is the discrete dipole array parameter information. The data format of the input set is:

[0047] .

[0048] The above input data set contains both information such as the polarization, frequency, and propagation direction of the incident wave, and information such as the spatial distribution, length, and polarization of the dipole array.

[0049] For the output data set, its form is relatively simple, which is the current coefficient of each discrete dipole. Therefore, its form is an N-dimensional vector:

[0050] .

[0051] The moment information neural network model depends on the classic network model in the field of computer vision: convolutional neural network (CNN), and is obtained through improved design. Traditional convolution operations can only extract the first-order moment information of the input sample data. In the analysis of electromagnetic problems, there are a large number of electromagnetic self-coupling and mutual-coupling effects between different dipole models. The essence reflected by this electromagnetic coupling effect is the quadratic term and cross-term operations between sample data. In this application, by introducing quadratic form calculation, the second-order moment information of the sample data is incorporated into the traditional CNN network model, which is more conducive to the learning of the high-order statistical features of the input data. More importantly, the introduction of quadratic terms and cross-terms fully analogizes the calculation rules of self-impedance elements and mutual-impedance elements in the moment method, and indirectly incorporates the constraint information of electromagnetic self-coupling / mutual-coupling. Therefore, to a certain extent, the generalization ability of the network model and the prediction accuracy of the current coefficient of the discrete dipole array are improved.

[0052] The architecture of the moment information neural network model is as Figure 2 shown, and it includes three modules:

[0053] Moment feature extraction module: responsible for extracting the statistical features of the input samples, and realizing the extraction of statistical features such as first-order moment, second-order moment, and cross-terms through quadratic form calculation and convolution calculation, and combining them through additive neurons.

[0054] Assume that the input sample data is a vector , then the operation rule of its first-order moment is: ; the operation rule of the second-order moment is . Among them, is the parameter to be estimated.

[0055] The output of the moment feature extraction module is . Denote the loss function as L , then L With respect to The gradient is:

[0056] ;

[0057] From we get , and , so:

[0058] (Equation 1)

[0059] In the above formula, can be obtained by the backpropagation algorithm of traditional CNN;N.

[0060] Similarly, we can get L With respect to The gradient is:

[0061] (Equation 2)

[0062] In the above formula, for the convenience of writing, use to represent the vectorization operation rule of sample data.

[0063] After obtaining the gradient calculation formulas such as (Equation 1) and (Equation 2), the gradient descent algorithm can be used for the training and learning of the weight parameters of the moment feature extraction module.

[0064] Convolutional feature extraction module: Extract abstract correlation features from the statistical features of sample data through a traditional convolutional neural network architecture. Among them, the convolutional layer is responsible for data feature extraction and contains multiple layers; the non-linear activation layer uses the ReLU function to achieve and is responsible for the sparsification of data features; the pooling layer is responsible for data dimensionality reduction and improving feature invariance.

[0065] Fully connected layer: Responsible for the mapping of output nodes.

[0066] This application innovatively designs a moment information neural network based on quadratic calculation and convolutional neural network, effectively integrating high-order statistical features such as data autocoupling / mutual coupling, and providing support for building a more generalized intelligent electromagnetic computing architecture. For the electromagnetic characteristics analysis of array structures, there are a large number of autocoupling / mutual coupling effects between different array elements, which are reflected in mathematical form as operations such as second-order moments and cross terms of sample data, while traditional convolutional neural networks generally do not involve the extraction of these high-order statistical features, and are unable to mine the mutual coupling features between data. The present invention innovatively proposes a moment feature extraction method based on quadratic calculation, and integrates it into the traditional convolutional neural network through additive neurons, so that the network model can not only mine the first-order linear features of the data, but also mine the higher-order cross features, providing support for building a more generalized intelligent electromagnetic computing architecture.

[0067] In one embodiment, the radiation field calculation is performed based on the discrete dipole array according to the electromagnetic field superposition principle, including:

[0068] According to electromagnetic theory, the time-harmonic dipole at the origin of the coordinate system is calculated in space. r The radiation field at the point and the single time-harmonic dipole located at the origin of the coordinate system in space r The radiation field at the point is replaced by coordinate variables to obtain the radiation field for a dipole whose spatial position is not at the origin;

[0069] Based on the superposition principle of electromagnetic fields, the radiation field of the discrete dipole array is obtained by vector superposition of the radiation fields of all dipoles.

[0070] In one embodiment, the time-harmonic dipole at the origin of a coordinate system is calculated based on electromagnetic theory. r The radiation field at a point, including:

[0071] According to electromagnetic theory, the time-harmonic dipole at the origin of the coordinate system is calculated in space. r The radiation field at the point is:

[0072] ;

[0073] in, m represents the dipole moment, represents the wave impedance of the medium, Represents the spatial wave number.

[0074] In one embodiment, based on the superposition principle of electromagnetic fields, vector superposition is performed on the radiation fields of all dipoles to obtain the radiation field of the discrete dipole array, including:

[0075] Based on the superposition principle of electromagnetic fields, the radiation field of the discrete dipole array is obtained by vector superposition of the radiation fields of all dipoles:

[0076] ;

[0077] In the above formula, N represents the total number of dipoles, r n represents the n th spatial position of the dipole, E n , H n represents the n th dipole's radiation field in space.

[0078] In a specific embodiment, for the equivalent modeling problem of electromagnetic scattering / radiation characteristics of three-dimensional targets, when a discrete dipole reference array model is established based on a triangular mesh and the corresponding physical structure characteristic parameters and current coefficients are obtained, the equivalent reconstruction of the target's spatial electromagnetic characteristics can be achieved.

[0079] To solve the above problem, first, the problem of accurately calculating the radiation characteristics of an arbitrarily distributed dipole array needs to be solved. According to electromagnetic theory, for a single time-harmonic dipole located at the origin of coordinates, its radiation field at a point r in space is:

[0080] ;

[0081] In the above formula, m represents the dipole moment, represents the medium wave impedance, represents the spatial wave number.

[0082] For a dipole whose spatial position is not at the origin, its radiation field can be obtained by replacing the coordinate variables in the above formula. In addition, according to the superposition principle of electromagnetic fields, for an array composed of multiple dipoles with different polarization characteristics, its radiation field can be obtained by vectorially superposing the radiation fields of individual dipoles. Now, assuming that the parameters such as the number N , position , and the dipole moment m of each dipole in a discrete dipole array are all known, then its radiation field in space can be calculated as follows:

[0083] ;

[0084] In the above formula, N represents the total number of dipoles; r n represents the n th spatial position of the dipole; E n , H n represents the nThe radiation field of a single dipole in space.

[0085] It is not difficult to find that the above calculation process only involves simple operations of vectors and matrices (summation, multiplication, etc.), without involving any high-time-consuming calculation processes such as iteration and inversion. Therefore, its calculation efficiency is extremely high, laying a foundation for the rapid prediction of the electromagnetic characteristics of any three-dimensional target.

[0086] This application is mainly used for the rapid prediction of the electromagnetic characteristics of radar three-dimensional targets. In the input data, the mesh division information of the target is incorporated, effectively integrating the structural physical characteristics of the calculation object. On the other hand, the intelligent calculation model is only used to predict the current coefficients of the dipole array. In the subsequent vector field synthesis process, the analytical formula of the dipole radiation field is incorporated, implicitly containing the physical characteristic constraints of electromagnetic far-field radiation. And, from the perspective of vector field synthesis, the prediction error of the tiny current term has a weak impact on the calculation result of the synthesized total field, effectively reducing the strict requirements of the model for the absolute accuracy of the current coefficients predicted by the network. Through the above incorporation of physical information, it is expected to reduce the dependence of the model on data, establish an intelligent electromagnetic calculation model driven by both "physics - data", and improve the accuracy and generalization ability of the model.

[0087] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the indication of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0088] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0089] The above-described embodiments only represent several implementation manners of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application should be subject to the appended claims.

Claims

1. An intelligent calculation method for the electromagnetic characteristics of radar three-dimensional targets, characterized in that, The method includes: Obtaining a radar three-dimensional target; performing equivalent modeling of the radar three-dimensional target with a discrete dipole array; the process of the modeling includes performing unstructured triangular mesh division on the target using a mesh division algorithm, defining each pair of triangular meshes with a common edge as a dipole, and extracting the position and normalized dipole moment of the dipole as geometric feature parameters; Constructing a moment information neural network model; the moment information neural network model includes a moment feature extraction module, where the moment feature extraction module extracts statistical features from the input sample data; the statistical features include first-order moment features and second-order moment features; the moment feature extraction module is used to extract statistical features from the input sample data, including: The extraction of statistical features from the input sample data respectively is: Among them, and are parameters to be estimated, represents the input sample data, and T represents the transpose operation; Predicting the current coefficient of the discrete dipole array obtained by modeling according to the moment information neural network model, including: Encoding the parameter information of the discrete dipole array obtained by modeling and the external irradiation wave information in a data format as input data and inputting them into the moment information neural network model for current coefficient prediction; the data format is: Among them, represents the external irradiation wave information; according to the electromagnetic field superposition principle, the radiation field of the discrete dipole array is calculated by combining the current coefficient based on the geometric characteristic parameters of the discrete dipole array, and the equivalent reconstruction of the electromagnetic characteristics of the target space is realized. r n represents the n position of the N th dipole, and l n represents the common edge length of the n th dipole, n represents the centroid coordinates of the positive triangular patch of the th dipole, n n represents the centroid coordinates of the negative triangular patch of the n th dipole.

2. The method according to claim 1, wherein Performing equivalent modeling of the radar three-dimensional target with a discrete dipole array, including: Use the grid division algorithm to perform unstructured triangular grid division on the 3D radar target, cluster all the obtained triangular grids, define a dipole for each pair of triangular grids with a common edge, number the established dipoles, denoted as , N which is the total number of all discrete dipoles; Define the position of each dipole as the center of the common edge of this pair of triangular meshes. Denote the position of the n -th dipole as r n . Define the normalized dipole moment of each dipole as the product of the three-dimensional vector from the centroid of the negative patch to the centroid of the positive patch and the length of the common edge. Denote the normalized dipole moment of the n -th dipole as , where l n represents the length of the common edge of the n -th dipole; represents the centroid coordinates of the positive triangular patch of the n -th dipole, and represents the centroid coordinates of the negative triangular patch of the n -th dipole.

3. The method according to claim 1, characterized in that The construction of the moment information neural network model includes a convolutional feature extraction module and a fully connected layer.

4. The method according to claim 3, characterized in that, The moment feature extraction module is used to extract statistical features from the input sample data and output them after combination through additive neurons; the convolutional feature extraction module is used to refine abstract correlation features from the statistical features of the sample data through a traditional convolutional neural network architecture; the fully connected layer is used to directly map and output the abstract correlation features and the current coefficient.

5. The method according to claim 1, wherein The calculation process of the parameter to be estimated includes: Establishing a gradient calculation formula and solving the gradient calculation formula according to the gradient descent algorithm to obtain the value of the parameter to be estimated; the process of establishing the gradient calculation formula includes: Let the loss function be L , then L with respect to the gradient is: From it follows that and since therefore: In the above formula, is obtained by the backpropagation algorithm of the traditional CNN; Similarly, we can obtain L with respect to the gradient is: In the above formula, Vec(·) represents the vectorization operation rule of the sample data.

6. The method according to claim 1, wherein Performing radiation field calculation based on the superposition principle of electromagnetic fields on the basis of the discrete dipole array, including: Calculate the radiation field of a single time - harmonic dipole located at the origin of coordinates in space according to electromagnetic theory r and obtain the radiation field of a dipole whose spatial position is not at the origin by performing a coordinate variable substitution on the radiation field of a single time - harmonic dipole located at the origin of coordinates at a point in space r in space; Based on the superposition principle of electromagnetic fields, performing vector superposition on the radiation fields of all dipoles to obtain the radiation field of the discrete dipole array.

7. The method according to claim 6, wherein Calculate the radiation field of a single time - harmonic dipole located at the origin of coordinates in space according to electromagnetic theory, including: r at the point, including: According to electromagnetic theory, the radiation field of a single time-harmonic dipole located at the origin of coordinates in space r at a point is as follows: Among them, m represents the dipole moment, represents the medium wave impedance, represents the spatial wave number.

8. The method according to claim 6, characterized in that Based on the superposition principle of electromagnetic fields, performing vector superposition on the radiation fields of all dipoles to obtain the radiation field of the discrete dipole array, including: Based on the superposition principle of electromagnetic fields, performing vector superposition on the radiation fields of all dipoles to obtain the radiation field of the discrete dipole array as: In the above formula, N represents the total sum of the number of all discrete dipoles, r n represents the n th spatial position of the dipole, E n , H n represents the n th radiation field of the dipole in space.

Images