Neural network acceleration simulation method for broadband target group radar scattering field

Abstract

The invention discloses a neural network acceleration simulation method for a broadband target group radar scattering field, and belongs to the technical field of electromagnetic simulation. According to the method, a KAN neural network model under a broadband center frequency is constructed, and KAN neural network models under other target frequencies are constructed in a transfer learning mode or a direct training mode; generating a multi-frequency-point numerical green function data set by using the trained KAN neural network model, and further constructing a broadband training sample to train a multi-task learning model to realize global prediction of the whole target broadband numerical green function; and fast simulation of a broadband target group scattering field is realized by using a numerical green function output by the multi-task learning model. According to the method, rapid and accurate simulation of the broadband target group radar scattering field can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic simulation, and particularly relates to a neural network accelerated simulation method for radar scattering fields of a wide-band target group. Background Art

[0002] Efficiently obtaining the ultra-wideband electromagnetic scattering characteristics of a target is of great significance for the research of high-resolution radar, imaging technology, and target recognition. Solving the electromagnetic scattering problem in a wide frequency band by numerical methods involves calculations at multiple frequency points and is a task that consumes a large amount of computing resources. To improve the solution efficiency, one can directly start from accelerating the algorithm itself, such as using the fast multipole method, adaptive integral method, fast Fourier transform acceleration algorithm, etc., which have a very significant effect on accelerating the calculation process of a single point. Another approach is to make full use of interpolation algorithms for sampling point information, such as asymptotic waveform estimation, physics-based modeling methods, time-frequency collaboration methods, model order reduction techniques, fast frequency sweep algorithms, and compressive sensing.

[0003] In recent years, artificial intelligence technology has brought new breakthroughs to computational electromagnetics. Its powerful learning ability is applicable to electromagnetic problems and provides an efficient solution for the forward modeling and reverse design of complex electromagnetic systems. In terms of forward modeling, a high-precision forward prediction model is constructed using a deep neural network, significantly improving the computational efficiency of problems such as metasurface characteristic analysis and wave field propagation simulation. In the field of reverse design, through the fusion of a deep generative model and electromagnetic physical constraints, the intelligent design optimization of manufacturable metasurfaces has been realized. However, the existing technologies are all for the direct prediction of the target electromagnetic field. When a new target is added to the solution scenario, or some targets in the scenario change, retraining is required, and the solution efficiency is low.

[0004] To improve the solution efficiency of the radar scattering field in a complex scenario, a numerical Green's function is introduced to perform electromagnetic modeling on fixed targets in the solution scenario. The numerical Green's function encapsulates the influence of complex media as a pre-computed function, which can greatly reduce the unknowns in electromagnetic calculation problems. After obtaining the numerical Green's function of the fixed target, the scattering field can be obtained by convolving the equivalent current of the external target (or changing target) with the numerical Green's function. When some targets change, the problem of having to re-solve the entire scenario is avoided, thus greatly improving the solution efficiency. Since solving the numerical Green's function must consider all source and field relationships in the computational domain and 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. Chinese patent application with publication number CN119783556A discloses a simulation method for radar scattering fields of a three-dimensional moving target group combined with a neural network. However, the above existing technology only involves the solution of the numerical Green's function at a single frequency point and cannot accelerate the solution of the wide-band numerical Green's function problem. Summary of the Invention

[0005] The object of the present invention is to overcome the defects of the above-mentioned prior art, and provide a neural network accelerated simulation method for the radar scattering field of a broadband target group, which is applicable to the simulation of the scattering field of a complex target group including fixed targets and variable targets.

[0006] The technical problem proposed by the present invention is solved as follows:

[0007] A neural network accelerated simulation method for the radar scattering field of a broadband target group includes the following steps:

[0008] Step 1: Determine the center frequency according to the broadband to be solved As a target frequency, select other target frequencies within the broadband to be solved ;

[0009] Step 2: Use the center frequency and its corresponding numerical Green's function data to construct training samples, where is a positive integer, train the KAN neural network model to obtain the KAN neural network model at the center frequency ; ;

[0010] Step 3: Use the numerical Green's function corresponding to the center frequency as the source domain, and the numerical Green's function corresponding to other target frequencies as the target domain, and judge whether the source domain and the target domain are relevant;

[0011] Step 4: If the source domain and the target domain are relevant, adopt the transfer learning method to fine-tune the KAN neural network model at the center frequency to obtain the KAN neural network model corresponding to other target frequencies ; ;

[0012] If the source domain and the target domain are not relevant, use the target frequency and its corresponding numerical Green's function data to construct the training samples corresponding to the target frequency and train the KAN neural network model to obtain the KAN neural network model at the target frequency ; ;

[0013] Set the position vector information of different source points and the position vector information of field points, and input them into the KAN neural network model corresponding to the target frequency to obtain the target frequency The corresponding numerical Green's function data, and then construct the target frequency corresponding training samples;

[0014] Step 5: Combine the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, and train the multi-task model ;

[0015] Step 6: The current broadband target group to be solved includes a fixed target group and an additional target; the position of the additional target is used as the source point position vector information, and all other positions within the solution domain are used as the field point position vector information, and input into the multi-task model to obtain the numerical Green's function corresponding to the additional target; the numerical Green's function corresponding to the additional target is convolved with the equivalent current to obtain the scattered field of the additional target; the scattered field of the fixed target group is superimposed with the scattered field of the additional target to obtain the scattered field of the current broadband target group to be solved.

[0016] Further, in Step 1, the broadband to be solved is 3~9 GHz, and a target frequency is taken every 0.2 GHz.

[0017] Further, based on the equality of the scattered fields obtained from the free-space Green's function and the numerical Green's function, the following integral equation is obtained:

[0018] ;

[0019] where and respectively represent the position vector information of the field point and the source point, represents the equivalent current of the additional scatterer under the action of the target group, represents that the target group is generated by the numerical Green's function, represents the three-dimensional Green's function in free space, ,[[]]END]] is the unit dyad, the wavenumber ,[[]]END]] is the frequency, is the speed of light, represents the bi-gradient operation, represents taking the modulus, represents the equivalent current of the target group, represents the numerical Green's function generated by the target group from the incident current source ,[[]]END]] represents located at the incident current source at the place, represents the solution domain.

[0020] Further, in Step 2, based on the integral equation, the target group is inversely solved at the center frequency The corresponding numerical Green's function , using the center frequency , the position information of the target group, the position vector information of the source point, the position vector information of the field point, and the corresponding numerical Green's function of the target group at the center frequency The corresponding numerical Green's function Construct training samples; change the position vector information of the source point and the position vector information of the field point, construct multiple training samples corresponding to the center frequency , form a training sample set; use the training sample set to train the KAN neural network model to obtain the KAN neural network model at the center frequency . .

[0021] Furthermore, in step 2, the training sample corresponding to the center frequency is expressed as , where the wave number corresponding to the center frequency ; is the position vector information between the source point and the field point, ; is the position vector information between the source point and the center of the target group, , is the central position vector information of the target group; is the position vector information between the field point and the center of the target group .

[0022] Furthermore, the specific process of step 3 is as follows:

[0023] Use the KAN neural network model at the center frequency to solve the corresponding numerical Green's function of other target frequencies , and count the relative error between the predicted value and the true value output by the KAN neural network model at the center frequency : ; :

[0024] ;

[0025] Among them, represents the predicted value of the corresponding numerical Green's function of other target frequencies output by the KAN neural network model at the center frequency , represents the true value of the corresponding numerical Green's function of other target frequencies , represents the center frequency and other target frequencies The distance weight, , denotes the calculation of the second norm;

[0026] If it satisfies ≤15%, it is determined that the source domain and the target domain are relevant; otherwise, it is determined that the source domain and the target domain are not relevant.

[0027] Furthermore, in step 4, for the target frequency where the source domain and the target domain are relevant , load the parameters of the KAN neural network model as the initialization parameters of the KAN neural network model ; construct P training samples corresponding to the target frequency , where P is a positive integer, and fine-tune and train the initial KAN neural network model to obtain the KAN neural network model corresponding to the target frequency .

[0028] Furthermore, in step 5, during the process of combining the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, for the training samples with the same position vector information of the source point and the field point, combine the position vector information of the source point, the position vector information of the field point, the center frequency and its corresponding numerical Green's function , each target frequency and its corresponding numerical Green's function to generate broadband training samples.

[0029] Furthermore, in step 5, the mean square error between the predicted value output by the multi-task model and the true result is:

[0030]

[0031] where, represents the weight of the th target frequency, ; and respectively represent the real part and the imaginary part of the predicted value of the numerical Green's function corresponding to the th broadband training sample and the th target frequency output by the multi-task learning model, and respectively represent the real part and the imaginary part of the true result of the numerical Green's function corresponding to the th broadband training sample and the th target frequency; 1 ≤ ≤ , is the total number of target frequencies; 1 ≤ ≤ , is the total number of broadband training samples;

[0032] If the mean square error of the numerical Green's function meets the set threshold requirement, it is determined that the training of the multi-task model is completed; otherwise, the number of broadband training samples is increased and the multi-task model is retrained.

[0033] Furthermore, in step 6, the scattered field of the fixed target group is superimposed with the scattered field of the applied target to obtain the scattered field of the current broadband target group to be solved , expressed as:

[0034] ;

[0035] where represents the numerical Green's function term generated by the incident current source for the fixed target group, represents the equivalent current of the applied target under the action of the fixed target group.

[0036] The beneficial effects of the present invention are:

[0037] The method of the present invention realizes the accelerated solution of the numerical Green's function of the broadband target group by constructing a KAN neural network model, greatly improving the calculation efficiency of the radar scattered field of the broadband target group; in the method of the present invention, a transfer learning method is adopted, and by virtue of the similarity of the target scattering characteristics between adjacent frequencies, the calculation cost of the multi-frequency point numerical Green's function data set is effectively reduced. By training a multi-task learning model, the real-time output of the numerical Green's function of the broadband target group is realized, reducing the unknowns in the subsequent scattered field calculation, saving a large amount of memory storage and calculation time, and providing an efficient and feasible new approach for the electromagnetic modeling of the broadband target group. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a schematic flow chart of the method of the present invention.

[0039] Figure 2 is a schematic structural diagram of the target group to be solved in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The present invention will be further described below with reference to the drawings and embodiments.

[0041] For the technical problem of efficiently simulating and solving the numerical Green's function of a wide-band target group, this embodiment provides a neural network accelerated simulation method for the radar scattering field of a wide-band target group, which is based on transfer learning and multi-task learning, and realizes the rapid calculation of the wide-band electromagnetic response through a multi-stage machine learning strategy. A Kolmogorov-Arnold (KAN) neural network model at the central frequency of the wide band is constructed to characterize the numerical Green's function of the target group at the central frequency; the central frequency The corresponding numerical Green's function is used as the source domain, and the other target frequencies The corresponding numerical Green's function is used as the target domain to determine whether the source domain and the target domain are relevant; if the source domain and the target domain are relevant, the transfer learning method is used to optimize and adjust the parameters of the KAN neural network model at the central frequency, and quickly generate KAN neural network models at multiple target frequencies; if the source domain and the target domain are not relevant, other target frequencies and their corresponding numerical Green's function data are used to construct training samples, and the KAN neural network model is trained to obtain the KAN neural network model at the target frequency; the KAN neural network models at other target frequencies are used to construct a multi-frequency point numerical Green's function data set; based on the multi-frequency point numerical Green's function data set, a wide-band training sample is constructed to train the multi-task learning model to realize the global prediction of the numerical Green's function of the entire target wide band; the numerical Green's function output by the multi-task learning model is used to realize the rapid simulation of the scattering field of the wide-band target group.

[0042] The flow schematic diagram of the method in this embodiment is as Figure 1 shown, and includes the following steps:

[0043] Step 1. Determine the central frequency as a target frequency, and select other target frequencies within the range of the wide band to be solved;

[0044] Step 2. Use the central frequency and its corresponding numerical Green's function data to construct training samples, and train the Kolmogorov-Arnold (KAN) neural network model to obtain the KAN neural network model at the central frequency ;

[0045] In step 2, based on the equality of the scattering fields obtained from the free space Green's function and the numerical Green's function, it can be obtained that:

[0046]

[0047] Among them, and respectively represent the position vector information of the field point and the source point, represents the equivalent current of the external scatterer under the action of the target group, represents that the target group is composed of the generated numerical Green's function, represents the three-dimensional Green's function in free space, , is the unit dyad, the wavenumber , is the frequency, is the speed of light, represents the bi-gradient operation, represents taking the modulus, represents the equivalent current of the target group, represents the numerical Green's function generated by the target group from the incident current source , represents being located at the incident current source at, represents the solution domain.

[0048] Based on the above integral equation, the numerical Green's function corresponding to the target group at the center frequency is inversely solved, using the center frequency , the position information of the target group, the position vector information of the source point, the position vector information of the field point, and the numerical Green's function corresponding to the target group at the center frequency to construct a training sample. By changing the position vector information of the source point and the position vector information of the field point, multiple training samples corresponding to the center frequency are constructed to form a training sample set. The Kolmogorov - Arnold (KAN) network model is trained using the training sample set, so that the KAN model is used to characterize the numerical Green's function of the target group at the center frequency, and the KAN neural network model at the center frequency is obtained. .

[0049] The training sample corresponding to the center frequency is expressed as , where the wavenumber corresponding to the center frequency , is the speed of light; the position vector information between the source point and the field point ; the position vector information between the source point and the center of the target group , is the center position vector information of the target group; the position vector information between the field point and the center of the target group .

[0050] Step 3, the center frequency The corresponding numerical Green's function is used as the source domain, and other target frequencies The corresponding numerical Green's function is used as the target domain to determine whether the source domain and the target domain are relevant;

[0051] Using the center frequency The KAN neural network model at Solve other target frequencies The corresponding numerical Green's function, and statistically analyze the KAN neural network model The relative error between the predicted value and the true value :

[0052]

[0053] Wherein, Indicates the center frequency The KAN neural network model at The predicted value of the corresponding numerical Green's function of other target frequencies output by ; Indicates other target frequencies The true value of the corresponding numerical Green's function, Indicates the center frequency And other target frequencies The distance weight between , Indicates the calculation of the two-norm.

[0054] If it satisfies ≤15%, it is determined that the source domain and the target domain are relevant and transfer learning can be performed; otherwise, it is determined that the source domain and the target domain are not relevant, and a large number of target frequency The corresponding numerical Green's function data Build training samples to train the KAN neural network model for other target frequencies The corresponding KAN neural network model.

[0055] Step 4. If the source domain and the target domain are relevant, use the transfer learning method to fine-tune the KAN neural network model To obtain the KAN neural network model corresponding to other target frequencies ; ;

[0056] If the source domain and the target domain are not relevant, use the target frequency And its corresponding numerical Green's function data Build training samples to train the KAN neural network model to obtain the KAN neural network model at the target frequency ; ;

[0057] Set the position vector information of different source points and the position vector information of field points, and input them to the target frequency The corresponding KAN neural network model to obtain the target frequency The corresponding multiple sets of numerical Green's function data, and then construct multiple sets of target frequencies The corresponding training samples

[0058] In step 4, for the target frequencies where the source domain and the target domain are correlated , load the parameters of the KAN neural network model as the initialization parameters of the KAN neural network model ; construct several training samples corresponding to the target frequency , and fine-tune and train the initial KAN neural network model to obtain the KAN neural network model corresponding to the target frequency .

[0059] Step 5: Combine the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, and train the multi-task model so that the multi-task model represents the numerical Green's function of the broadband

[0060] In the process of combining the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, for the training samples with the same position vector information of the source point and the position vector information of the field point, combine the position vector information of the source point, the position vector information of the field point, the center frequency and its corresponding numerical Green's function , each target frequency and its corresponding numerical Green's function to generate broadband training samples

[0061] In step 5, the mean square error between the predicted value of the multi-task model and the true result is measured as follows

[0062]

[0063] where represents the weight of the th target frequency ; and respectively represent the real part and the imaginary part of the predicted value of the numerical Green's function corresponding to the th broadband training sample and the th target frequency output by the multi-task learning model and respectively represent the real part and the imaginary part of the true result of the numerical Green's function corresponding to the -th broadband training sample at the -th target frequency; 1 ≤ ≤ , being the total number of target frequencies; 1 ≤ ≤ , being the total number of broadband training samples.

[0064] If the mean square error of the numerical Green's function meets the set threshold requirement, it is determined that the training of the multi-task model is completed; otherwise, the number of broadband training samples is increased and the multi-task model is retrained.

[0065] Step 6. Based on the multi-task model obtain the numerical Green's function corresponding to the applied target, and calculate the scattering field of the current target group to be solved;

[0066] Take the position of the applied target as the source point position vector information, and take all other positions in the solution domain as the field point position vector information, and input them into the multi-task model to obtain the numerical Green's function corresponding to the applied target; convolve the numerical Green's function corresponding to the applied target with the equivalent current to obtain the scattering field of the applied target; superimpose the scattering field of the fixed target group and the scattering field of the applied target to obtain the scattering field of the current broadband target group to be solved.

[0067] In Step 6, the scattering field of the fixed target group and the scattering field of the applied target are superimposed to obtain the scattering field of the current broadband target group to be solved , which is expressed as:

[0068]

[0069] where represents the numerical Green's function term generated by the incident current source for the fixed target group, represents the equivalent current of the applied target under the action of the fixed target group.

[0070] In the method described in this embodiment, the single-frequency-point KAN neural network model adopts a three-layer KAN network, where the three layers of the network respectively include 234, 10, and 1 neurons, and the Adam optimizer is used with a learning rate of 0.005. The multi-task model adopts a four-layer KAN neural network model, which respectively includes 234, 150, 50, and 31 neurons, and the Adam optimizer is used with a learning rate of 0.005.

[0071] The structural schematic diagram of the target group to be solved in this embodiment is asFigure 2 As shown in Figure 2 , the target group consists of 15 different dielectric unit targets distributed within a range of 0.12m × 0.12m × 0.3m. The relative permittivity of the unit target is 40, the shape of the unit target is a sphere, and the radius is 0.005m. The frequency is set from 3 to 9GHz, and a target frequency point is taken every 0.2GHz.

[0072] After the training based on the constructed training samples is completed, the mean square error of the real part and the mean square error of the imaginary part of the broadband numerical Green's function are 6.3×10 -3 and 4.3×10 -3 .

[0073] In the method described in this embodiment, the broadband scattering field is calculated by the broadband numerical Green's function output by the multi-task model and compared with the scattering field calculated by the free-space Green's function. The two are highly consistent, which verifies the effectiveness and high-precision characteristics of the method described in this embodiment. By outputting the numerical Green's function terms of the full frequency band through the trained multi-task model, the solution of the scattering effect of the modeled target can be omitted. Only the influence of the externally added target scatterer needs to be considered to solve the total scattering field, saving computational resources and time.

[0074] In summary, the method described in this embodiment improves the calculation efficiency of the numerical Green's function of the broadband target group while ensuring accuracy. Through the transfer learning technique, by leveraging the similarity of the target scattering characteristics between adjacent frequencies, the computational load of solving the numerical Green's function dataset at other frequencies is reduced. The network model outputs a large number of datasets at other target frequencies obtained through the transfer learning technique for training the multi-task model to achieve the real-time output of the numerical Green's function of the broadband target group, greatly reducing the unknowns in the subsequent scattering field calculation problem and saving a large amount of memory and time resources, providing a feasible new approach for the efficient calculation of the radar scattering field of the broadband target group.

Claims

1. A neural network acceleration simulation method for the radar scattering field of a wideband target group, characterized in that It includes the following steps: Step 1. Determine the center frequency according to the broadband to be solved As a target frequency, select other target frequencies within the broadband to be solved ; Step 2: Using the center frequency and its corresponding numerical Green's function data to construct training samples, where K is a positive integer, training the KAN neural network model to obtain the KAN neural network model at the center frequency ; Step 3, center frequency The corresponding numerical Green's function is used as the source domain, and the other target frequencies The corresponding numerical Green's function is used as the target domain to determine whether the source domain and the target domain are correlated; Step 4. If the source domain and the target domain are relevant, adopt transfer learning to fine-tune the KAN neural network model at the center frequency to obtain other target frequencies corresponding to the KAN neural network model ; If the source domain and the target domain are not relevant, use the target frequency and its corresponding numerical Green's function data to construct training samples corresponding to the target frequency and train the KAN neural network model to obtain the KAN neural network model at the target frequency ; Set the position vector information of different source points and the position vector information of field points, and input them to the target frequency The corresponding KAN neural network model to obtain the target frequency The corresponding numerical Green's function data, and then construct the target frequency corresponding training samples; Step 5. Combine the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, and train the multi-task model therewith; Step 6: The current broadband target group to be solved includes a fixed target group and additional targets; the positions of the additional targets are used as the source point position vector information, and all other positions within the solution domain are used as the field point position vector information, which is input into the multi-task model to obtain the numerical Green's function corresponding to the additional targets; the numerical Green's function corresponding to the additional targets is convolved with the equivalent current to obtain the scattered field of the additional targets; the scattered field of the fixed target group is superimposed with the scattered field of the additional targets to obtain the scattered field of the current broadband target group to be solved.

2. The neural network accelerated simulation method for the wideband target group radar scattering field according to claim 1, wherein In step 1, the broadband to be solved is 3 - 9 GHz, and a target frequency is taken every 0.2 GHz.

3. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 1, wherein Based on the equality of the scattered fields obtained from the free - space Green's function and the numerical Green's function, the following integral equation is obtained: ； Among them, and respectively represent the position vector information of the field point and the source point, represents the equivalent current of the external scatterer under the action of the target group, represents that the target group is composed of the numerical Green's function generated, represents the three-dimensional Green's function in free space, , is the unit dyad, the wavenumber , is the frequency, is the speed of light, represents the bi-gradient operation, represents taking the modulus, represents the equivalent current of the target group, represents the numerical Green's function generated by the target group from the incident current source , represents located at the incident current source at the place, represents the solution domain.

4. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 3, wherein In step 2, based on the integral equation, the numerical Green's function corresponding to the target group at the center frequency is inversely solved and a training sample is constructed by using the center frequency , the position information of the target group, the position vector information of the source point, the position vector information of the field point, and the numerical Green's function corresponding to the target group at the center frequency . By changing the position vector information of the source point and the position vector information of the field point, multiple training samples corresponding to the center frequency are constructed to form a training sample set . ​ Train the KAN neural network model using the training sample set to obtain the KAN neural network model at the center frequency under .

5. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 4, characterized in that, In step 2, the center frequency The corresponding training sample is expressed as , where the wavenumber corresponding to the center frequency ; is the position vector information between the source point and the field point, ; is the position vector information between the source point and the center of the target group, , is the central position vector information of the target group; is the position vector information between the field point and the center of the target group .

6. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 1, wherein, The specific process of step 3 is as follows: Using the center frequency of the KAN neural network model to solve other target frequencies and the corresponding numerical Green's function, and statistically analyzing the center frequency of the KAN neural network model of the relative error between the predicted value and the true value output : ； Among them, represents the center frequency of the KAN neural network model for other target frequencies output and the corresponding predicted values of the numerical Green's function represents other target frequencies and the corresponding true values of the numerical Green's function represents the center frequency and other target frequencies and the distance weight therebetween , represents the calculation of the two-norm; If ≤ 15%, it is determined that the source domain and the target domain are relevant; otherwise, it is determined that the source domain and the target domain are not relevant.

7. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 1, characterized in that In step 4, for the target frequency where the source domain and the target domain are relevant , load the parameters of the KAN neural network model as the initialization parameters of the KAN neural network model ; construct P training samples corresponding to the target frequency , where P is a positive integer, and fine-tune the initial KAN neural network model to obtain the KAN neural network model corresponding to the target frequency .

8. The neural network acceleration simulation method for the wideband target group radar scattering field according to claim 1, wherein In step 5, during the process of combining the training samples corresponding to the center frequency and other target frequencies to generate broadband training samples, for the training samples with the same position vector information of the source point and the field point, the position vector information of the source point, the position vector information of the field point, the center frequency and its corresponding numerical Green's function , each target frequency and its corresponding numerical Green's function are combined to generate broadband training samples.

9. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 1, wherein In step 5, the mean square error between the predicted value output by the multi-task model and the true result is as follows: ； Among them, represents the weight of the th target frequency, ; and respectively represent the real and imaginary parts of the predicted value of the numerical Green's function corresponding to the th target frequency of the th broadband training sample output by the multi-task learning model, and respectively represent the real and imaginary parts of the true result of the numerical Green's function corresponding to the th target frequency of the th broadband training sample; 1 ≤ ≤ , is the total number of target frequencies; 1 ≤ ≤ , is the total number of broadband training samples; If the mean square error of the numerical Green's function meets the set threshold requirement, it is determined that the training of the multi-task model is completed; otherwise, the number of broadband training samples is increased and the multi-task model is retrained.

10. The neural network accelerated simulation method for the radar scattering field of a broadband target group according to claim 3, wherein In step 6, the scattered field of the fixed target group is superimposed with the scattered field of the externally applied target to obtain the scattered field of the current broadband target group to be solved, which is expressed as: , which is expressed as: ； Among them, represents the numerical Green's function term generated by the incident current source for the fixed target group and represents the equivalent current of the applied target under the action of the fixed target group.

Images