Radiometer dense feed source array directional diagram optimization method based on generative adversarial network

Abstract

The invention relates to a radiation meter dense feed source array directional diagram optimization method based on a generative adversarial network. The method comprises the following steps: constructing a training data set, a condition vector containing a target directional diagram performance index and a corresponding optimal amplitude-phase weight vector; designing a conditional generative adversarial network model; constructing a physical constraint loss function by embedding a differentiable electromagnetic calculation module; loading the constructed training data set, and training a complete generative adversarial network model based on a loss function embedded with a differentiable electromagnetic calculation module; and when the system antenna pattern obtained according to the amplitude-phase adjustment parameters output by the generative adversarial network model meets the antenna pattern evaluation index, completing dense feed source array antenna pattern optimization. According to the method, the target weight is dynamically adjusted and optimized on orbit, so that a single model can adapt to beam forming requirements in different observation scenes, and flexible switching of multiple task modes such as high-resolution detection and wide-range general survey is realized.

Description

A Radiation Pattern Optimization Method for Dense Feed Arrays of Radiometers Based on Generative Adversarial Networks Technical Field

[0001] This invention belongs to the interdisciplinary fields of microwave remote sensing technology, antenna engineering and artificial intelligence, and relates to a radiation meter dense feed array pattern optimization method based on generative adversarial networks. Background Technology

[0002] High-precision microwave radiometers are key equipment for modern meteorological, oceanographic, and environmental monitoring. To achieve high spatial resolution and radiation accuracy, advanced radiometer systems commonly employ dense feed arrays and digital beamforming technology. By optimizing the amplitude and phase excitation weights of each array element, low-sidelobe, high-efficiency, and flexibly directional beams can be formed, and antenna profile manufacturing errors can be compensated for.

[0003] However, finding the optimal amplitude and phase weights for large-scale dense arrays is a complex, high-dimensional, non-convex optimization problem. Traditional optimization algorithms, such as genetic algorithms and particle swarm optimization, are computationally intensive and time-consuming; gradient-based algorithms are prone to getting trapped in local optima. These methods are all insufficient to meet the requirements of real-time on-orbit processing.

[0004] In recent years, machine learning methods have been introduced into this field, but existing methods are mostly based on pure data and lack constraints from physical laws, which may lead to results that violate the fundamental principles of electromagnetism. Therefore, there is an urgent need for a new optimization method that combines high computational efficiency, superior global performance, and compliance with physical laws. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks, so as to realize intelligent optimization of the radiation pattern of dense feed arrays of radiometers.

[0006] The solution to the technical problem of this invention is: a radiation meter dense feed array pattern optimization method based on generative adversarial networks (GANs), comprising the following steps: constructing a training dataset containing conditional vectors of target pattern performance indicators and corresponding optimal amplitude and phase weight vectors; designing a conditional GAN ​​model, including generator network design, discriminator network design, and network parameter initialization; constructing a loss function of physical constraints by embedding a differentiable electromagnetic computation module; loading the constructed training dataset and training a complete GAN model based on the loss function of the embedded differentiable electromagnetic computation module; and completing the dense feed array antenna pattern optimization when the system antenna pattern obtained according to the amplitude and phase adjustment parameters output by the GAN model satisfies the antenna pattern evaluation indicators.

[0007] Furthermore, the construction of the training dataset specifically involves: determining the target pattern performance parameter range: the main lobe pointing angle range, including the coverage range of azimuth and elevation angles, the 3dB beamwidth range, and the maximum sidelobe level setting; generating optimal amplitude and phase weight vectors: using optimization algorithms to generate condition vectors, including genetic algorithms, particle swarm optimization, and convex optimization algorithms, each algorithm running independently, and selecting the optimal result as a training sample; for each condition vector, generating at least three different optimal amplitude and phase weight vectors by introducing amplitude and phase errors; normalizing the amplitude weights and mapping them to the [0,1] interval; normalizing the phase weights and mapping them to the [-π,π] interval; standardizing each parameter in the condition vectors to eliminate the influence of dimensions; and dividing the dataset as follows: 70% for the training set, 20% for the validation set, and 10% for the test set.

[0008] Furthermore, the generator network design is as follows: Input layer: receives conditional vectors and random noise vectors; Hidden layer structure: 5 fully connected layers with 512, 256, 128, 64, and 32 neurons respectively; Activation function: the hidden layer uses the Leaky ReLU function with a parameter set to 0.2; Output layer design: the first N nodes use the Sigmoid function to output amplitude weights, and the last N nodes use a linear function to output phase weights.

[0009] Furthermore, the discriminator network design specifically includes: an input layer that simultaneously receives amplitude and phase weight vectors and conditional vectors; feature fusion that fuses weighted features and conditional features through concatenation; a hidden layer structure consisting of four fully connected layers with 256, 128, 64, and 32 neurons respectively; an output layer with a single-node output using the Sigmoid activation function; and a regularization measure employing Dropout technology with a dropout rate of 0.3.

[0010] Furthermore, the network parameter initialization specifically includes: weight initialization: using the Xavier uniform distribution initialization method; bias initialization: all biases are initialized to zero; learning rate setting: the initial learning rate is set to 0.0002.

[0011] Furthermore, the loss function for constructing physical constraints by embedding a differentiable electromagnetic calculation module is specifically implemented as follows: the differentiable electromagnetic calculation module is based on the principle of pattern product and superposition, mathematically expressed as: F(θ,φ) = Σ[w_i · e (jφ_i) · EL_i(θ,φ) · e (j·(2π / λ)·r_i·r_hat)] where w_i and φ_i are the amplitude and phase weights of the i-th unit, EL_i(θ,φ) is the unit radiation pattern, r_i is the unit position vector, r_hat is the observation direction unit vector, and λ is the wavelength; this calculation process is completely differentiable and supports gradient backpropagation; Physics-driven loss function design: Adversarial loss: using Wasserstein distance metric to improve training stability; Sidelobe loss function: using an asymmetric penalty strategy to strengthen sidelobe suppression; Main lobe loss function: including beamwidth loss and beam pointing accuracy loss; Efficiency loss function: focusing on optimizing the main beam efficiency index; Joint loss function configuration: the total loss function expression is: L_physics = α·L_SLL + β·L_beamwidth + γ·L_pointing + δ • L_efficiency, where L_SLL is for sidelobe suppression, L_beamwidth is for main lobe beamwidth loss, L_pointing is for main lobe beam pointing accuracy, L_efficiency is for main beam efficiency, and α, β, γ, δ are configurable weight coefficients that are dynamically adjusted according to the training process for adaptive adjustment.

[0012] Furthermore, the training of the complete generative adversarial network model specifically involves: initializing the network parameters of the generator and discriminator and loading them into the constructed training dataset; training employs an alternating optimization strategy, with each training iteration comprising two stages: In the first stage, the generator parameters are fixed, and real amplitude and phase weight data are sampled from the training dataset. Simultaneously, the generator generates simulated weight data based on random noise and target performance conditions. The discriminator calculates the discrimination loss by comparing the real weight data with the generated weight data and updates the discriminator parameters to improve discrimination capability; In the second stage, the discriminator parameters are fixed, and the generator generates new weight data based on the target conditions. On the one hand, the adversarial loss is calculated to evaluate the effect of the newly generated weight data in deceiving the discriminator. On the other hand, the newly generated weights are input into the differentiable electromagnetic calculation module to obtain the radiation pattern. The physical driving loss, which is composed of the difference between the performance indicators of the radiation pattern and the target radiation pattern, is further calculated. The two are weighted and combined, and the generator parameters are updated through backpropagation; During the training process, the changes in the total loss function and the performance of the validation set are continuously monitored. When the key indicators of the radiation pattern output by the generator meet the design requirements and the model converges, the final network parameters are saved, and training is completed.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the radiation meter dense feed array pattern optimization method based on generative adversarial networks.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the radiation meter dense feed array pattern optimization method based on generative adversarial networks.

[0015] A computer program product includes a computer program that, when executed by a processor, implements the steps of the radiation meter dense feed array pattern optimization method based on generative adversarial networks.

[0016] The beneficial effects of this invention compared with the prior art are: (1) The method of this invention shortens the traditional optimization calculation of several hours to the forward inference of milliseconds, realizes real-time beam control and reconstruction, and provides possibilities for dynamic remote sensing scenarios.

[0017] (2) The conditional adversarial grid in the method of the present invention can learn and explore complex weight distributions. Combined with the direct guidance of physical loss, it can often find solutions with better performance (such as lower sidelobes and more accurate beam shapes) that exceed the search range of traditional algorithms.

[0018] (3) The method of the present invention, through training including errors, enables the model to learn to synthesize robust beams in the presence of array distortion, thereby reducing the system’s requirements for processing and assembly accuracy and improving its engineering practicality.

[0019] (4) A well-trained model obtained by the method of the present invention is a universal beam designer that can cover all possible beamforming requirements in its training space without having to rerun the optimization program for each new task.

[0020] (5) Compared with reinforcement learning, the supervised adversarial training framework adopted in this invention is more stable, has better convergence, and generates high-quality and highly consistent beams. Attached Figure Description

[0021] Figure 1 is a schematic diagram of digital beamforming provided in an embodiment of the present invention; Figure 2 is a flowchart of a radiation meter dense feed array pattern optimization method based on generative adversarial networks provided in an embodiment of the present invention; Figure 3 is a schematic diagram of a deep neural network structure provided in an embodiment of the present invention; Figure 4 is an optimized antenna pattern provided in an embodiment of the present invention. Detailed Implementation

[0022] This invention proposes a method for optimizing the radiation pattern of a dense feed array of radiometers based on generative adversarial networks (GANs), comprising the following steps: S1: Constructing a training dataset, including conditional vectors of target radiation pattern performance indicators and corresponding optimal amplitude and phase weight vectors; S2: Designing a conditional GAN ​​model, including generator network design, discriminator network design, and network parameter initialization; S3: Constructing a loss function for physical constraints by embedding a differentiable electromagnetic computation module; S4: Loading the constructed training dataset and training a complete GAN model based on the loss function embedded with the differentiable electromagnetic computation module; S5: When the system antenna pattern obtained according to the amplitude and phase adjustment parameters output by the GAN model satisfies the antenna pattern evaluation indicators, the optimization of the dense feed array antenna pattern is completed.

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Example 1: Figure 1 is a schematic diagram of digital beamforming provided in an embodiment of the present invention, detailed as follows: The core components of the digital beamforming system include a dense feed array, a receiving channel array, and a data processor. Its workflow is as follows: the dense feed array provides an initial single-feed pattern; the data processor optimizes and adds amplitude and phase weight parameters accordingly to dynamically form the final optimal antenna pattern.

[0025] As shown in Figure 2, the radiation pattern optimization method for dense feed array of radiometers based on generative adversarial networks provided by the present invention specifically includes the following steps: S1: Construct a training dataset, which includes the condition vector of the target radiation pattern performance index and the corresponding optimal amplitude and phase weight vector. Specifically: (1) First, determine the range of target radiation pattern performance parameters: the main lobe pointing angle range, including the coverage range of azimuth and elevation angles; the 3dB beamwidth range; the maximum sidelobe level setting, etc.

[0026] (2) Next, generate the optimal amplitude and phase weight vector: use traditional optimization algorithms to generate condition vector C. The traditional optimization algorithms include genetic algorithm, particle swarm algorithm and convex optimization algorithm. Each algorithm runs independently and selects the optimal result as the training sample. For each condition vector, at least 3 different optimal amplitude and phase weight vectors are generated by introducing amplitude error and phase error.

[0027] (3) Finally, the data is preprocessed: the amplitude weights are normalized and mapped to the [0,1] interval; the phase weights are normalized and mapped to the [-π,π] interval; the parameters in the condition vector are standardized to eliminate the influence of the units; the dataset is divided as follows: the training set accounts for 70%, the validation set accounts for 20%, and the test set accounts for 10%.

[0028] S2: Design conditional generative adversarial network model, including generator network design, discriminator network design, and network parameter initialization.

[0029] Figure 3 shows a schematic diagram of the generative adversarial network model structure provided in this embodiment of the invention, as follows: (1) Generator network design: Input layer: receiving conditional vector C and random noise vector Z; Hidden layer structure: 5 fully connected layers with 512, 256, 128, 64 and 32 neurons respectively; Activation function: Leaky activation function is used in the hidden layer. ReLU function, parameter set to 0.2; Output layer design: the first N nodes use the Sigmoid function to output amplitude weights, and the last N nodes use the linear function to output phase weights; (2) Discriminator network design: Input layer: simultaneously receives amplitude and phase weight vector W and condition vector C; Feature fusion: fused weight features and condition features by splicing; Hidden layer structure: 4 fully connected layers, with the number of neurons being 256, 128, 64, and 32 respectively; Output layer: single node output, using the Sigmoid activation function; Regularization measures: using Dropout technology, with the dropout rate set to 0.3; (3) Network parameter initialization: Weight initialization: using the Xavier uniform distribution initialization method; Bias initialization: all initialized to zero; Learning rate setting: the initial learning rate is set to 0.0002.

[0030] S3: Construct a loss function for physical constraints by embedding a differentiable electromagnetic calculation module.

[0031] This step is used to embed physical constraints. The specific implementation process includes: (1) Implementation of the differentiable electromagnetic calculation module: Based on the array antenna theory, a radiation pattern calculation model is constructed to realize the radiation pattern product principle, support calculation of arbitrary observation directions, include unit radiation pattern model, accurately reflect unit characteristics, ensure that all calculation processes are differentiable, and support gradient propagation.

[0032] The differentiable electromagnetic calculation module is implemented based on the principle of pattern product and the principle of superposition, and its mathematical expression is: F(θ,φ) = Σ[w_i · e (jφ_i) · EL_i(θ,φ) · e (j·(2π / λ) ·r_i·r_hat)] where w_i and φ_i are the amplitude and phase weights of the i-th cell, EL_i(θ,φ) is the cell radiation pattern, r_i is the cell position vector, r_hat is the observation direction unit vector, and λ is the wavelength; this calculation process is completely differentiable and supports gradient backpropagation.

[0033] (2) Physically driven loss function design: Adversarial loss: Wasserstein distance metric is used to improve training stability; Side lobe loss function: Asymmetric penalty strategy is used to strengthen side lobe suppression; Main lobe loss function: Includes beamwidth loss and beam pointing accuracy loss; Efficiency loss function: Focus on optimizing the main beam efficiency index.

[0034] (3) Joint Loss Function Configuration: The total loss function is expressed as: L_physics = α·L_SLL + β·L_beamwidth + γ·L_pointing + δ·L_efficiency, where L_SLL is for sidelobe suppression, L_beamwidth is for main lobe beamwidth loss, L_pointing is for main lobe beam pointing accuracy, and L_efficiency is for main beam efficiency. α, β, γ, and δ are configurable weight coefficients. The weight coefficients are dynamically adjusted according to the training process for adaptive adjustment.

[0035] S4: Load the constructed training dataset and train the complete generative adversarial network model based on the loss function embedded with the differentiable electromagnetic computation module. Specifically: First, initialize the network parameters of the generator and discriminator, and load the training dataset constructed in step S1. Training employs an alternating optimization strategy, with each training iteration containing two stages: In the first stage, the generator parameters are fixed, and real amplitude and phase weight data are sampled from the training dataset. Simultaneously, the generator generates simulated weight data based on random noise and target performance conditions. The discriminator calculates the discrimination loss by comparing the real weight data with the generated weight data and updates the discriminator parameters to improve discrimination capability. In the second stage, the discriminator parameters are fixed, and the generator generates new weight data based on the target conditions. On one hand, it calculates the adversarial loss to evaluate the effect of the newly generated weight data in deceiving the discriminator; on the other hand, it inputs the newly generated weights into the differentiable electromagnetic computation module to obtain the radiation pattern. Further calculation of the physical driving loss constituted by the difference in performance indicators between the radiation pattern and the target radiation pattern is performed. The two weighted combinations are then used to update the generator parameters through backpropagation. During training, the changes in the total loss function and the performance on the validation set are continuously monitored. Once the key metrics of the generator's output orientation graph meet the design requirements and the model converges, the final network parameters are saved, and training is complete. This process, through the close integration of adversarial training and physical constraints, ensures the accuracy and physical rationality of the generated weights.

[0036] S5: When the system antenna pattern obtained from the amplitude and phase adjustment parameters output by the generative adversarial network model meets the antenna pattern evaluation index, the dense feed array antenna pattern optimization is completed, and the final optimized antenna pattern is obtained.

[0037] To further illustrate the radiation pattern optimization method for a dense feed array antenna based on a generative adversarial network provided in this embodiment of the invention, a detailed description is provided below with reference to a specific example: In this embodiment, the radiation pattern of a 25-element dense feed array is optimized.

[0038] (1) In a microwave anechoic chamber environment, antenna pattern tests were performed on each independent feed in the 25-element dense feed array to obtain the original pattern data of each feed; (2) The main beam efficiency was used as the core evaluation index, and the optimization target was set to be no less than 99%; (3) The measured pattern data of each feed and the initial amplitude and phase weights were input into the constructed generative adversarial network model; (4) The system synthetic pattern performance corresponding to the amplitude and phase weights generated by forward inference was used as feedback, and the network parameters were iteratively optimized in combination with the physical driving loss function; (5) When the main beam efficiency of the system antenna pattern output by the model exceeds 98.5%, the network training was deemed complete.

[0039] Experimental results show that the main beam efficiency of the system optimized using this method is improved to 98.7%, as shown in Figure 4. Compared with the 80.9% main beam efficiency of the single feed before optimization, the performance improvement is significant. The method proposed in this invention can effectively improve the main beam efficiency of the radiation pattern of dense feed array antennas, meeting the application requirements of high-precision radiometer systems.

[0040] The above embodiments verify the technical effects of the present invention: the present invention can effectively optimize the amplitude and phase distribution of dense feed arrays and significantly improve antenna pattern performance. Specifically, through a physically constrained generative adversarial network, it can simultaneously achieve effective suppression of sidelobe levels, significant improvement of main lobe efficiency, and accurate control of beam pointing accuracy, fundamentally solving the problem that traditional optimization methods struggle to obtain globally optimal solutions under complex constraints. Furthermore, this method, through an end-to-end intelligent optimization architecture, can replace multiple independent optimization modules and error compensation units in traditional systems, significantly improving computational efficiency and reducing system complexity, providing an innovative solution for the lightweight and intelligent design of radiometer systems.

[0041] This application provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the method described in FIG2.

[0042] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0043] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0046] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0047] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks, characterized in that, Includes the following steps: Construct a training dataset containing the conditional vector of the target pattern performance index and the corresponding optimal amplitude and phase weight vector; The design of the conditional generative adversarial network model includes generator network design, discriminator network design, and network parameter initialization; a loss function for physical constraints is constructed by embedding a differentiable electromagnetic computation module; the constructed training dataset is loaded, and the complete generative adversarial network model is trained based on the loss function of the embedded differentiable electromagnetic computation module; when the system antenna pattern obtained from the amplitude and phase adjustment parameters output by the generative adversarial network model meets the antenna pattern evaluation index, the dense feed array antenna pattern optimization is completed.

2. The radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks according to claim 1, characterized in that, The construction of the training dataset specifically involves: determining the target pattern performance parameter range: the main lobe pointing angle range, including the coverage range of azimuth and elevation angles, the 3dB beamwidth range, and the maximum sidelobe level setting; generating the optimal amplitude and phase weight vector: using an optimization algorithm to generate a condition vector, the optimization algorithm including genetic algorithm, particle swarm optimization algorithm and convex optimization algorithm, each algorithm runs independently, and the optimal result is selected as the training sample; For each conditional vector, at least three different optimal amplitude and phase weight vectors are generated by introducing amplitude error and phase error; the amplitude weights are normalized and mapped to the [0,1] interval; the phase weights are normalized and mapped to the [-π,π] interval; the parameters in the conditional vectors are standardized to eliminate the influence of dimensions; the dataset is divided as follows: training set accounts for 70%, validation set accounts for 20%, and test set accounts for 10%.

3. The radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks according to claim 1, characterized in that, The generator network design is as follows: Input layer: receives conditional vector and random noise vector; Hidden layer structure: 5 fully connected layers with 512, 256, 128, 64 and 32 neurons respectively; Activation function: the hidden layer uses the Leaky ReLU function with a parameter set to 0.2; Output layer design: the first N nodes use the Sigmoid function to output amplitude weights, and the last N nodes use a linear function to output phase weights.

4. The radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks according to claim 3, characterized in that, The discriminator network design is as follows: Input layer: simultaneously receives amplitude and phase weight vectors and conditional vectors; Feature fusion: weight features and conditional features are fused by concatenation; Hidden layer structure: 4 fully connected layers with 256, 128, 64 and 32 neurons respectively; Output layer: single-node output, using the Sigmoid activation function; Regularization measure: Dropout technology is adopted, with a dropout rate set to 0.

3.

5. The radiation pattern optimization method for a dense feed array of radiometers based on generative adversarial networks according to claim 4, characterized in that, The network parameter initialization is specifically as follows: weight initialization: using the Xavier uniform distribution initialization method; bias initialization: all biases are initialized to zero; learning rate setting: the initial learning rate is set to 0.0002.

6. The radiation pattern optimization method for dense feed arrays of radiometers based on generative adversarial networks according to claim 1, characterized in that, The loss function for constructing physical constraints by embedding a differentiable electromagnetic calculation module is specifically implemented based on the principle of pattern product and superposition, mathematically expressed as: F(θ,φ) = Σ[w_i · e (jφ_i) · EL_i(θ,φ) · e (j·(2π / λ) ·r_i·r_hat)] where w_i and φ_i are the amplitude and phase weights of the i-th unit, EL_i(θ,φ) is the unit radiation pattern, r_i is the unit position vector, r_hat is the observation direction unit vector, and λ is the wavelength; this calculation process is completely differentiable and supports gradient backpropagation; physics-driven loss function design: adversarial loss: using Wasserstein distance metric to improve training stability; sidelobe loss function: using an asymmetric penalty strategy to strengthen sidelobe suppression; main lobe loss function: including beamwidth loss and beam pointing accuracy loss; Efficiency loss function: Focuses on optimizing the main beam efficiency index; Joint loss function configuration: The total loss function expression is: L_physics = α·L_SLL +β ·L_beamwidth + γ·L_pointing + δ ·L_efficiency, where L_SLL is for sidelobe suppression, L_beamwidth is for main lobe beamwidth loss, L_pointing is for main lobe beam pointing accuracy, and L_efficiency is for main beam efficiency. α, β, γ, and δ are configurable weight coefficients that are dynamically adjusted according to the training process for adaptive adjustment.

7. The radiation pattern optimization method for a dense feed array of radiometers based on generative adversarial networks according to claim 6, characterized in that, The training of the complete generative adversarial network model specifically involves: initializing the network parameters of the generator and discriminator, and loading the constructed training dataset; training adopts an alternating optimization strategy, and each training iteration includes two stages: in the first stage, the generator parameters are fixed, real amplitude and phase weight data are sampled from the training dataset, and at the same time, the generator generates simulated weight data based on random noise and target performance conditions; the discriminator calculates the discrimination loss by comparing the real weight data with the generated weight data, and updates the discriminator parameters to improve the discrimination ability; In the second stage, the discriminator parameters are fixed, and the generator generates new weight data based on the target conditions. On the one hand, the adversarial loss is calculated to evaluate the effect of the newly generated weight data in deceiving the discriminator. On the other hand, the newly generated weights are input into the differentiable electromagnetic calculation module to obtain the radiation pattern. The physical driving loss formed by the difference between the performance index of the radiation pattern and the target radiation pattern is further calculated. The two are weighted and combined and then the generator parameters are updated through backpropagation. During training, the changes in the total loss function and the performance on the validation set are continuously monitored. When the key indicators of the generator's output orientation graph meet the design requirements and the model converges, the final network parameters are saved, and training is completed.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images