Simulation method of radar scattering field of three-dimensional moving target group combined with neural network

Abstract

The present invention discloses a radar scattering field simulation method for a three-dimensional moving target group combined with a neural network, and belongs to the field of electromagnetic simulation technology. The method of the present invention converts the electromagnetic response of a three-dimensional moving target group into the response of the target when it is distributed at different positions, and extracts the three-dimensional numerical Green's function of the randomly distributed target group in the solution domain by changing the spatial distribution of the target group; the three-dimensional numerical Green's function is split into scattering terms and coupling terms, and the artificial neural network is trained separately; when training the artificial neural network, according to the characteristics of the moving target group, its position distribution and scattering terms / coupling terms are input into the artificial neural network as the main features; the scattering terms and coupling terms output by the artificial neural network model are combined into numerical Green's function terms, which can be applied to integral equations to achieve fast and accurate solution of the scattering field of the three-dimensional moving target group.

Description

Technical Field

[0001] The invention belongs to the technical field of electromagnetic simulation, and in particular relates to a three-dimensional moving target group radar scattering field simulation method combined with a neural network. Background Art

[0002] Electromagnetic modeling and scattering analysis are widely used in many fields of science and engineering, including biomedical imaging, radar systems and target detection, atmospheric scattering analysis, antenna design, wireless communication channel modeling, metamaterial research, etc. In theory, the target can be treated as a whole and analyzed using computational methods such as the method of moments, finite element method, finite difference time domain method, and multi-layer fast multipole algorithm. However, when the target is in slow motion or randomly distributed, these prior art methods become inefficient due to the need to recalculate to account for position changes.

[0003] In recent years, deep learning has become an important research direction in the field of artificial intelligence and has been widely used in electromagnetic scattering and electromagnetic inverse scattering problems. Its application methods are divided into pure data-driven methods and electromagnetic physics-driven methods. The initial research was based on the deep learning black box method, which was purely data-driven and directly input the data to be solved into the neural network to perform reverse gradient descent to optimize the network parameters. In order to be suitable for applications in the electromagnetic field, a physics-driven neural network model was proposed, which combines electromagnetic field knowledge and its mathematical expression into the input or internal architecture of the neural network.

[0004] The numerical Green's function encapsulates the influence of complex media into a pre-calculated function, which can greatly reduce the unknowns in electromagnetic calculation problems. Since solving the numerical Green's function must consider all the sources and field relationships in the calculation domain, it is a task that consumes a huge amount of computing resources. Artificial intelligence technology can be used to accelerate the solution of the numerical Green's function. The prior art "Solving two-dimensional scattering from multiple dielectric cylinders by artificial neural network accelerated numerical Green's function" discloses an accelerated solution method for the numerical Green's function of a two-dimensional dielectric cylinder based on an artificial neural network. By using the powerful function fitting ability of the neural network, the implicit numerical Green's function of the associated field source coordinates and the point source response is approximated and fitted in the form of a composite function, thereby achieving accurate and efficient solution of the numerical Green's function. On this basis, for multi-body scattering scenarios, the prior art "Recursive solution of numerical Green's function for multibody scattering scenarios using artificial neural networks acceleration" discloses a recursive method for accelerating the acquisition of numerical Green's function using artificial neural networks. The recursive method is used to decompose the multi-body problem, and the scattering characteristics of the multi-body problem are decomposed into the scattering characteristics of a single target, which is conducive to the accelerated calculation of the neural network. However, the above-mentioned existing technology only solves the problem of solving the numerical Green's function in two dimensions. The two-dimensional numerical Green's function is a simple scalar form, that is, it verifies the feasibility of using neural networks to accelerate the fitting of scalar Green's function from a theoretical level, but cannot solve the problem of electromagnetic scattering calculation of targets in a three-dimensional solution domain.

[0005] The prior art "Calculation of numerical Green's function of three-dimensional medium based on neural network" discloses a method of fitting three-dimensional numerical Green's function using deep neural network. The numerical Green's function in the prior art refers to the electric field of the field point under the excitation of unit current element. Its essence is the direct fitting of the scattered field, rather than solving the numerical Green's function term that can be used for integral equations. The prior art also discloses an accelerated solution method for three-dimensional numerical Green's function that can be applied to integral equations. The solution scene is three-dimensional, but the actual solution data is only selected on the two-dimensional observation surface. At the same time, the prior art only involves the solution of the numerical Green's function of a single fixed target, and is not suitable for the solution of the numerical Green's function of a three-dimensional moving target group and the radar scattering field. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and provide a method for simulating the radar scattering field of a three-dimensional moving target group in combination with a neural network.

[0007] The technical problem proposed by the present invention is solved in this way:

[0008] A three-dimensional moving target group radar scattering field simulation method combined with a neural network includes the following steps:

[0009] Step 1: Using the quasi-stationary approximation method, randomly change the position of each target in the target group, and extract the three-dimensional numerical Green's function of the target group in the solution domain, including scattering terms and coupling terms;

[0010] Step 2: construct training samples using the scattering term and the coupling term respectively, and use the training samples to train the first artificial neural network model for fitting the scattering term and the second artificial neural network model for fitting the coupling term respectively;

[0011] Step 3: Input the position distribution of the current three-dimensional motion target group to be solved into the first artificial neural network model and the second artificial neural network model respectively, obtain the scattering terms and coupling terms of each target of the current three-dimensional motion target group to be solved, and then obtain the three-dimensional numerical Green's function of the current three-dimensional motion target group to be solved in the solution domain, and calculate the scattering field of the current three-dimensional motion target group to be solved.

[0012] Furthermore, in step 1, for any target group with a certain position distribution, the three-dimensional numerical Green's function of the target group in the solution domain is extracted: The specific process is:

[0013]

[0014] in, represents the scattering term of the i-th target in the target group, 1≤i≤n, n is the number of targets in the target group; It represents the coupling term of the interaction among the targets in the target group.

[0015] Furthermore, in step 1, the scattering term of the i-th target in the target group is for:

[0016]

[0017] in, is the impedance matrix of the ith target, is the equivalent current of the ith target, is the scattered field generated by the i-th target under the irradiation of external current, is the equivalent current of the external scatterer under the action of the i-th target.

[0018] Furthermore, in step 1, the coupling term of the interaction between the targets in the target group for:

[0019]

[0020] in, is the impedance matrix of the target group, is the equivalent current of the target group, is the weight parameter, The target group is composed of an incident current source The resulting numerical Green's function term, For the i-th target, the incident current source The resulting numerical Green's function term, is the equivalent current of the external scatterer under the action of the target group.

[0021] Furthermore, in step 1, the weight parameter .

[0022] Furthermore, in step 2, the position distribution and field source relationship of the i-th target in the target group are used as training data, and the scattering term of the i-th target in the target group is used as a label to construct the first type of training samples;

[0023] The location distribution and field source relationship of the target group are used as training data, and the coupling items of the interactions among the targets in the target group are used as labels to construct the second type of training samples.

[0024] Furthermore, in step 2, for any polarization component in the three-dimensional numerical Green's function, the mean square error MSE between the predicted value of the artificial neural network and the true value is expressed as:

[0025]

[0026] Where P and Q represent the number of first-category training samples and second-category training samples, respectively, 1≤p≤P, 1≤q≤Q; and denote the real and imaginary parts of the predicted values ​​of the first artificial neural network model used to fit the scattering term, respectively; and denote the real and imaginary parts of the scattering term of the i-th target respectively; and represent the real and imaginary parts of the predicted values ​​of the second artificial neural network model used to fit the coupling term, respectively; and denote the real and imaginary parts of the coupling term of the target group respectively;

[0027] For all polarization components in the three-dimensional numerical Green's function, if the mean square error (MSE) meets the set threshold requirement, the training of the artificial neural network model is considered to be completed. Otherwise, the number of training samples is increased and the artificial neural network model is retrained.

[0028] Furthermore, in step 3, the scattering field of the three-dimensional moving target group to be solved is It is expressed as:

[0029]

[0030] in, and denote the position vectors of the field point and the source point respectively, represents the three-dimensional Green's function in free space, Indicates that it is located The incident current source at , V represents the solution domain, The target group is represented by the incident current source The resulting numerical Green's function term, represents the equivalent current of the external scatterer under the action of the current three-dimensional moving target group to be solved, Represents the three-dimensional numerical Green's function of the three-dimensional moving target group to be solved in the solution domain.

[0031] The beneficial effects of the present invention are:

[0032] The method described in the present invention realizes the real-time output of the numerical Green's function of a three-dimensional moving target group by constructing an artificial neural network model, thereby greatly improving the electromagnetic modeling efficiency of the moving target scene; in the method described in the present invention, it is only necessary to input the position of the three-dimensional moving target group to be solved into the trained artificial neural network model to obtain the corresponding numerical Green's function information, thereby simplifying the electromagnetic modeling pre-processing process, and at the same time reducing the unknown quantities in the subsequent calculation of the radar scattering field problem based on the numerical Green's function, thereby saving a large amount of memory and time resources, and providing a feasible new approach for the electromagnetic modeling of a three-dimensional moving target group. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0034] Figure 2 It is a schematic diagram of a solution process of a three-dimensional numerical Green's function of a three-dimensional moving target group to be solved in the method of the present invention;

[0035] Figure 3 It is a schematic diagram of the structure of the three-dimensional moving target group to be solved in the method described in the embodiment. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0037] This embodiment aims at the problem of efficient solution of numerical Green's function of three-dimensional moving target group, and provides a three-dimensional moving target group radar scattering field simulation method combined with neural network. The quasi-stationary approximation method is adopted to convert the electromagnetic response of the moving target group into the response of the target when it is distributed at different positions. By changing the spatial distribution of the target group, the three-dimensional numerical Green's function of the randomly distributed target group in the solution domain is extracted; the three-dimensional numerical Green's function is split into scattering terms and coupling terms, and artificial neural networks are trained separately; when training the artificial neural network, according to the characteristics of the moving target group, its position distribution and scattering terms / coupling terms are input into the artificial neural network as main features; the scattering terms and coupling terms output by the artificial neural network model are combined into numerical Green's function terms, which can be applied to integral equations to achieve fast and accurate solution of the radar scattering field of the three-dimensional moving target group.

[0038] The schematic flow chart of the method described in this embodiment is as follows Figure 1 As shown, the following steps are included:

[0039] Step 1: Using the quasi-stationary approximation method, the electromagnetic response problem of the moving target group is converted into the response problem of the target at different positions; the position of each target in the target group is randomly changed, and the three-dimensional numerical Green's function of the target group in the solution domain is extracted, including the scattering term and the coupling term;

[0040] In step 1, for any target group with a certain position distribution, the three-dimensional numerical Green's function of the target group in the solution domain is extracted. The specific process is:

[0041]

[0042] in, represents the scattering term of the i-th target in the target group, 1≤i≤n, n is the number of targets in the target group; The coupling term represents the interaction between the targets in the target group;

[0043] The scatter term of the i-th target in the target group for:

[0044]

[0045] in, is the impedance matrix of the ith target, is the equivalent current of the ith target, is the scattered field generated by the i-th target under the irradiation of external current, is the equivalent current of the external scatterer under the action of the i-th target.

[0046] The coupling term of the interaction between the targets in the target group for:

[0047]

[0048] in, is the impedance matrix of the target group, is the equivalent current of the target group, is the weight parameter, The target group is composed of an incident current source The resulting numerical Green's function term, For the i-th target, the incident current source The resulting numerical Green's function term, is the equivalent current of the external scatterer under the action of the target group.

[0049] In this embodiment, the weight parameter .

[0050] Step 2, constructing training samples using the scattering term and the coupling term respectively, and using the training samples to train the first artificial neural network model ANNs for fitting the scattering term and the second artificial neural network model ANNc for fitting the coupling term respectively;

[0051] The process of constructing training samples using scatter terms is:

[0052] The position distribution and source relationship of the i-th target in the target group are used as training data, and the scattering term of the i-th target in the target group is used as a label to construct the first type of training samples;

[0053] The first type of training samples is represented as , is the location information of the i-th target, is the location information of the source point, It is the location information of the field point.

[0054] The process of constructing training samples using coupling terms is:

[0055] The location distribution and source relationship of the target group are used as training data, and the coupling items of the interaction between the targets in the target group are used as labels to construct the second type of training samples;

[0056] The second type of training samples is represented as .

[0057] For any polarization component in the three-dimensional numerical Green's function, the mean square error (MSE) between the predicted value and the true value of the artificial neural network is expressed as:

[0058]

[0059] Where P and Q represent the number of first-category training samples and second-category training samples, respectively, 1≤p≤P, 1≤q≤Q; and They represent the real and imaginary parts of the predicted values ​​of the first artificial neural network model ANNs used to fit the scattering term, respectively; and denote the real and imaginary parts of the scattering term of the i-th target respectively; and They respectively represent the real and imaginary parts of the predicted values ​​of the second artificial neural network model ANNc used to fit the coupling term; and represent the real and imaginary parts of the coupling term of the target group respectively.

[0060] For all polarization components in the three-dimensional numerical Green's function, if the mean square error (MSE) meets the set threshold requirement, the training of the artificial neural network model is considered to be completed. Otherwise, the number of training samples is increased and the artificial neural network model is retrained.

[0061] Since the field points and source points in the training data are randomly selected, this means that there is no intrinsic relationship between adjacent elements in the input training data. Therefore, the convolutional neural network that uses convolution and pooling to capture adjacent elements of adjacent training data is not suitable for this problem, while the artificial neural network can effectively and quickly establish the functional relationship between the training data and the label. The artificial neural network model used in this embodiment is an eight-layer artificial neural network, in which the hidden units of the eight-layer neural network include 250, 300, 300, 250, 200, 150, 100 and 50 neurons respectively, and the activation function used by the neuron is the ReLU function, and the Adam optimizer is used, and the learning rate is 0.003.

[0062] Step 3: Figure 2 As shown, the position distribution of the current three-dimensional moving target group to be solved is input into the first artificial neural network model ANNs and the second artificial neural network model ANNc respectively, and the scattering terms and coupling terms of each target of the current three-dimensional moving target group to be solved are obtained, and then the three-dimensional numerical Green's function of the current three-dimensional moving target group to be solved in the solution domain is obtained, and the scattering field of the current three-dimensional moving target group to be solved is calculated.

[0063] The scattering field of the three-dimensional moving target group to be solved It is expressed as:

[0064]

[0065] in, and denote the position vectors of the field point and the source point respectively, represents the three-dimensional Green's function in free space, Indicates that it is located The incident current source at , V represents the solution domain, The target group is represented by the incident current source The resulting numerical Green's function term, represents the equivalent current of the external scatterer under the action of the current three-dimensional moving target group to be solved, Represents the three-dimensional numerical Green's function of the three-dimensional moving target group to be solved in the solution domain.

[0066] The schematic diagram of the structure of the three-dimensional moving target group to be solved in this embodiment is as follows Figure 3 As shown in the figure, the target group consists of 8 different inhomogeneous dielectric unit targets distributed in the range of 1.2 m×1.2 m×1.2 m. The relative dielectric constant of the unit target is randomly selected in the range of [3.0,4.0]. The shape of the unit target is a cube with a side length of 0.12 m. The frequency is set to 500 MHz.

[0067] After the training based on the constructed training samples is completed, the real mean square error and imaginary mean square error of the scattering term, coupling term, and three-dimensional numerical Green's function are shown in Table 1.

[0068] Table 1 Mean square error table

[0069] Real mean square error Imaginary mean square error Scattering term 0.0002 0.0005 Coupling term 0.0016 0.0070 Three-dimensional numerical Green's function 0.0018 0.0075

[0070] In the method described in this embodiment, the radar scattering field calculated by the numerical Green's function output by the artificial neural network model is compared with the radar scattering field calculated by the free space Green's function, and the two maintain a high degree of consistency, which confirms the effectiveness and high precision characteristics of the method described in this embodiment. By using the numerical Green's function term output by the trained artificial neural network model, the scattering effect of the modeled target can be omitted, and only the influence of the external target scatterer needs to be considered to solve the total scattering field, saving computing resources and time.

[0071] In summary, the method described in this embodiment improves the calculation efficiency of the numerical Green's function of the three-dimensional moving target group while ensuring accuracy. Through the quasi-stationary approximation method, the changes of the three-dimensional moving target group in the time domain are converted into its position changes in the space domain. With the help of the learning ability of the neural network, a suitable artificial neural network model is constructed to characterize the impact of the spatial position change of the target group, and the real-time output of the numerical Green's function of the three-dimensional moving target group in the domain is achieved, which greatly reduces the unknown quantity of the subsequent radar scattering field calculation problem, saves a lot of memory and time resources, and provides a feasible new way for electromagnetic modeling of three-dimensional moving target groups.

Claims

1. A method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network, characterized in that: The following steps are involved: Step 1: Using the quasi-stationary approximation method, randomly change the position of each target in the target group, and extract the three-dimensional numerical Green's function of the target group in the solution domain, including scattering terms and coupling terms; In step 1, for any target group with position distribution, extract the three-dimensional numerical Green's function of the target group in the solution domain The specific process is: ;in, represents the scattering term of the i-th target in the target group, 1≤i≤n, n is the number of targets in the target group; The coupling term that represents the interaction between the targets in the target group; The scatter term of the i-th target in the target group for: ;in, is the impedance matrix of the ith target, is the equivalent current of the ith target, is the scattered field generated by the i-th target under the irradiation of external current, is the equivalent current of the external scatterer under the action of the i-th target; Step 2: construct training samples using the scattering term and the coupling term respectively, and use the training samples to train the first artificial neural network model for fitting the scattering term and the second artificial neural network model for fitting the coupling term respectively; Step 3: Input the position distribution of the current three-dimensional motion target group to be solved into the first artificial neural network model and the second artificial neural network model respectively, obtain the scattering terms and coupling terms of each target of the current three-dimensional motion target group to be solved, and then obtain the three-dimensional numerical Green's function of the current three-dimensional motion target group to be solved in the solution domain, and calculate the scattering field of the current three-dimensional motion target group to be solved.

2. The method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network according to claim 1 is characterized in that: In step 1, the coupling term of the interaction between the targets in the target group is for: ;in, is the impedance matrix of the target group, is the equivalent current of the target group, is the weight parameter, The target group is composed of an incident current source The resulting numerical Green's function term, For the i-th target, the incident current source The resulting numerical Green's function term, is the equivalent current of the external scatterer under the action of the target group.

3. The method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network according to claim 2 is characterized in that: In step 1, the weight parameter .

4. The method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network according to claim 1, characterized in that: In step 2, the position distribution and field source relationship of the i-th target in the target group are used as training data, and the scattering term of the i-th target in the target group is used as a label to construct the first type of training samples; The location distribution and field source relationship of the target group are used as training data, and the coupling items of the interactions among the targets in the target group are used as labels to construct the second type of training samples.

5. The method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network according to claim 1, characterized in that: In step 2, for any polarization component in the three-dimensional numerical Green's function, the mean square error MSE between the predicted value of the artificial neural network and the true value is expressed as: ; Wherein, P and Q represent the number of first-category training samples and second-category training samples, respectively, 1≤p≤P, 1≤q≤Q; and denote the real and imaginary parts of the predicted values ​​of the first artificial neural network model used to fit the scattering term, respectively; and denote the real and imaginary parts of the scattering term of the i-th target respectively; and represent the real and imaginary parts of the predicted values ​​of the second artificial neural network model used to fit the coupling term, respectively; and denote the real and imaginary parts of the coupling term of the target group respectively; For all polarization components in the three-dimensional numerical Green's function, if the mean square error (MSE) meets the set threshold requirement, the training of the artificial neural network model is considered to be completed. Otherwise, the number of training samples is increased and the artificial neural network model is retrained.

6. The method for simulating the radar scattering field of a three-dimensional moving target group combined with a neural network according to claim 1, characterized in that: In step 3, the scattering field of the three-dimensional moving target group to be solved is It is expressed as: ;in, and denote the position vectors of the field point and the source point respectively, represents the three-dimensional Green's function in free space, Indicates that it is located The incident current source at , V represents the solution domain, The target group is represented by the incident current source The resulting numerical Green's function term, represents the equivalent current of the external scatterer under the action of the current three-dimensional moving target group to be solved, Represents the three-dimensional numerical Green's function of the three-dimensional moving target group to be solved in the solution domain.

Images