Multi-target pointing intelligent anti-interference beam generation method

Abstract

The invention discloses a multi-target pointing intelligent anti-interference beam generation method, and belongs to the crossing field of digital communication and artificial intelligence. The method comprises the following main steps: constructing a data set, generating a moving target pointing prediction intelligent model, generating an antenna pattern which aims at multi-target pointing and has an anti-interference capability, and realizing multi-target tracking and interference suppression. According to the invention, for the problems of errors and time delay in moving target positioning and the requirements of multi-direction pointing and multi-direction suppression, a method of predicting a target position by deep learning and generating a multi-antenna phase control antenna beam in parallel by an optimization algorithm is adopted, so that enhancement and suppression in multiple directions can be realized at the same time; the modular design is adopted, so that the module reusability is high, the framework processing logic is visual, and the process is clear and concise.

Description

Technical Field

[0001] This invention discloses a multi-target pointing intelligent anti-interference beam generation method, which belongs to the interdisciplinary field of digital communication and artificial intelligence. Background Technology

[0002] In complex electromagnetic scenarios such as aerospace telemetry and control and unmanned collaborative operations, it is necessary to generate beam pointing in a timely manner in response to changes in motion position and suppress interference in the direction of interference, thereby improving communication performance and stability. This is an important research topic in the field of electromagnetic spectrum.

[0003] In the paper "A 3D Beamforming Image Compression Method Based on a Generative Model" published in *Chinese Journal of Mechanical Engineering*, Zhao Yunjie et al. proposed a 3D beamforming image compression method based on an improved vectorized variational autoencoder model, which can effectively calculate beamforming results in 3D space. However, the proposed 3D image compression method has a relatively complex structure and fails to consider issues such as position errors in complex motion scenarios and the suppression of crosstalk generated by a specified direction of arrival. In the paper "Effect Analysis of Particle Swarm Optimization Algorithm on Beamforming" published in the *Journal of Chengdu University of Information Technology*, Teng Yupeng et al. discussed the effect of the particle swarm optimization algorithm on beamforming and analyzed the impact of particle swarm optimization on beamforming performance under different variable spaces, providing many improvement suggestions. However, this method focuses on the application problems of the algorithm itself and gives little consideration to the application background of complex electromagnetic scenarios. The patent application "Optimization Method for Integrated Trajectory and Beamforming of Multiple UAVs Based on MADT3" filed by Nanjing University of Posts and Telecommunications (Publication No. CN120186624B) employs the MATD3 algorithm, which combines centralized training with distributed execution, to train UAVs and achieve optimization of integrated trajectory and beamforming. However, the deep reinforcement learning-based integrated paradigm suffers from poor interpretability and issues with the inability to modularly reuse and improve functions such as tracking allocation and beam generation. Summary of the Invention

[0004] Traditional beam generation methods suffer from shortcomings in beam tracking and interference suppression for multiple moving targets. Firstly, delays and errors during localization introduce deviations in target pointing, reducing accuracy and interference suppression. Secondly, using deep learning and other intelligent methods to predict the trajectories of multiple moving targets simultaneously faces limitations in input quantity and representation difficulties. Therefore, this invention proposes an intelligent anti-interference beam generation method for multi-target pointing. Firstly, based on deep learning and feature engineering, long-term continuous localization results are correlated, and temporal differential information and localization system errors are superimposed to construct features for predicting the pointing of moving targets. Further, by reusing the prediction model, the positions of each target are predicted, guiding the generation of the objective function for multi-target pointing. Finally, differential evolution calculation methods are used to optimize the antenna phase, generating an antenna pattern that enhances pointing accuracy and reduces interference, thus achieving intelligent anti-interference beam generation with multi-target pointing capability.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for generating intelligent anti-jamming beams with multi-target pointing, comprising the following steps:

[0007] Step 1: Obtain continuous positioning information of multiple moving targets, including position and relative pointing angle, and perform differential operation on the relative pointing angle to obtain the difference of relative pointing angle between adjacent time moments as a supplement to the feature. Then, superimpose the positioning error range to construct feature data, and use the pointing angle at the last time moment as sample data to form a dataset. The dataset includes feature data and sample data. Finally, the dataset is split into training dataset and test dataset.

[0008] Step 2: Construct a dedicated intelligent model for predicting the pointing of moving targets based on a deep neural network. Train the dedicated intelligent model for predicting the pointing of moving targets using a training dataset, and select and save the optimal model based on its performance on the test dataset.

[0009] Step 3: Based on the existing multi-target localization information, use the intelligent dedicated model for moving target pointing prediction to predict the multi-target localization results. Construct an objective function based on the multi-target localization results, and optimize the antenna phase based on the objective function to obtain the optimal phase. Then, set each phase of the antenna to the optimal phase to form an antenna pattern that is oriented towards multiple targets and has anti-interference capabilities, thereby realizing multi-target tracking and interference suppression.

[0010] Furthermore, the specific process of step 1 is as follows:

[0011] Step 1-1: Obtain continuous localization information of multiple moving targets at consecutive time points T+1, including their positions at each time point. and and relative pointing angle And perform difference calculations based on the relative pointing angle. Then add the positioning error range Constituent features and the pointing angle at time T+1. As a sample l, where:

[0012]

[0013] ;

[0014] Step 1-2: Repeat step 1-1 a set number of times to form a dataset, including a feature set. and tag set ;

[0015] Steps 1-3: According to proportion The dataset is divided into a training dataset and a test dataset. The training dataset includes a training feature set and a training label set, while the test dataset includes a test feature set and a test label set.

[0016] Furthermore, the specific process of step 2 is as follows:

[0017] Step 2-1: Construct a dedicated intelligent model for predicting moving targets. The structure includes a convolutional block (ConvBlock), an upsampling convolutional block (UpConvBlock), a pooling layer (Pool), and an output block (OutLine).

[0018] ConvBlock consists of 1D convolutional layers and batch normalization layers, used to extract local features of the input signal. The input parameters are the number of input channels in_channels, the number of output channels out_channels, the kernel size kernel_size, and padding.

[0019] Upsampling convolutional blocks (UpConvBlock) employ transposed convolutional layers to upsample and restore spatial dimensions. The input parameters are the number of input channels (in_channels), the number of output channels (out_channels), the kernel size (kernel_size), and the stride.

[0020] The pooling layer uses max pooling to reduce the spatial dimension of the input while preserving key features, thus aligning the dimensions between modules. The input parameters are the kernel size and stride.

[0021] The output block OutLine consists of 1D convolutional layers and fully connected layers, and is used to output the fitting results. The input parameters are the number of input channels in_channels, the number of output channels out_channels, the kernel size kernel_size, and the padding.

[0022] The processing procedure is as follows:

[0023] Based on the parameters of the ConvBlock convolutional block, it is instantiated as the enc1 module, and the moving target is pointed to the input of the prediction intelligent dedicated model to obtain the feature xe1.

[0024] Based on the pooling layer, parameters are set to instantiate a pool module, which processes feature xe1 to obtain features. ;

[0025] Based on the ConvBlock convolutional block parameter setting instantiation into the enc2 module, the features The processing yields feature xe2;

[0026] Based on the pooling layer, feature xe2 is processed to obtain features. ;

[0027] Based on the ConvBlock convolutional block parameter setting instantiation as an enc3 module, the features The processing yields feature xe3;

[0028] Based on the pooling layer, feature xe3 is processed to obtain features. ;

[0029] Based on the upsampling convolutional block UpConvBlock, the parameters are instantiated as up3 modules, and the features are... The feature xu3 is obtained through processing;

[0030] xu3 and The data is concatenated according to the second dimension and merged into xw3;

[0031] Based on the ConvBlock convolutional block parameter setting instantiation into a dec3 module, the feature xw3 is processed to obtain the feature ;

[0032] Based on the upsampling convolutional block UpConvBlock, the parameters are instantiated as up2 module, and the features are... The feature xu2 is obtained through processing;

[0033] xu2 and The data is concatenated according to the second dimension and merged into xw2;

[0034] Based on the ConvBlock convolutional block, the parameters are instantiated as a dec2 module, and the feature xw2 is processed to obtain... ;

[0035] Based on the parameter settings of the 1D convolutional layer in the output block OutLine, the features are instantiated into the out_conv module. The processing yields feature c1;

[0036] Dimensionality reduction is performed on the second dimension of feature c1;

[0037] Based on the settings parameters of the fully connected layer in the output block OutLine, the out_line module is instantiated, and the feature c1 is processed to obtain the feature output.

[0038] Step 2-2: Initialize the variable epoch = 0 to record the number of training iterations, and initialize the minimum loss value Min_Loss to infinity to record the minimum loss value during training and testing; set the maximum number of training iterations num_epochs;

[0039] Steps 2-3: During the training phase, switch the model to training mode, randomly select B data points from the training feature set and training label set to construct a batch of data, then clear the gradient cache, perform forward propagation, calculate the model output, and use the mean squared error loss function to calculate the loss between the model output and the labels in the batch of data, perform back propagation, calculate the gradient, and update the model parameters; where B is a set value.

[0040] Steps 2-4: Enter the testing phase. Switch the model to evaluation mode, disable gradient calculation, perform forward propagation, output predicted data based on the test feature set, and use the mean squared error loss function to calculate the loss value Loss of the model output predicted data and the test label set. If the loss Loss is lower than the minimum loss value Min_Loss, save the current model and assign the test loss value to the minimum loss value Min_Loss. Otherwise, increment the training count epoch by 1 and return to step 2-3 until all data in the training feature set and training label set are taken or the maximum training count num_epochs is reached.

[0041] Furthermore, the specific process of step 3 is as follows:

[0042] Step 3-1: Construct feature data based on multi-target localization results to generate predicted multi-target localization results:

[0043] The optimal moving target pointing prediction intelligent dedicated model is loaded. Continuous positioning information of multiple moving targets requiring antenna beam pointing, including position, relative pointing angle, differential calculation results based on the relative pointing angle, and positioning error range, is sequentially input into the model to obtain the relative pointing angles of the moving targets requiring pointing. ,in P represents the number of targets that need to be pointed towards by the antenna beam. The continuous positioning information of the multiple moving targets requiring antenna beam suppression, including their position, relative pointing angle, the result of differential calculation based on the relative pointing angle, and the positioning error range, is sequentially input into the model to obtain the relative angles of the moving targets to be suppressed. ,in Q represents the number of targets for which antenna beam suppression is required;

[0044] Step 3-2: Construct a multi-target localization result generation objective function, with the input being the relative angles of all moving targets whose orientation needs to be determined. The relative angle of the moving target that needs to be suppressed and antenna array element vector The output is the antenna optimization effect; among which Phase of antenna elements composition, N is the number of antenna array elements;

[0045] The objective function is:

[0046]

[0047] Where d is the antenna spacing;

[0048] Step 3-3: Perform population initialization. Uniformly initialize a two-dimensional array as the initial population within the given phase range phase_range, representing the possible phase values ​​of the antenna elements. The dimension is num_particles×N, where num_particles is the number of individuals in the population and N is the number of antennas.

[0049] Set the first individual as the current best solution, and calculate and initialize it to the best fitness value, best_score, through the objective function; perform a specified number of iterations, with the maximum iteration value set to num_iterations; and execute steps 3-4 after completing the iterations.

[0050] In each iteration, every individual in the population is traversed. An individual 'a' is randomly selected from the initial population as the current individual, and two different individuals 'b' and 'c' are randomly selected. A mutation operation is performed to form a new individual 'mutant', whose value is equal to a + F × (b - c), where F is the difference weight coefficient, and the value is restricted to the phase range 'phase_range'. Then, according to the crossover probability 'CR', two gene positions of the individual 'mutant' are randomly selected, and the two selected genes are exchanged to generate a new candidate solution 'trial'. If the fitness value of the newly generated candidate solution 'trial' is smaller than that of individual 'a', then the original individual 'a' is replaced by the candidate solution 'trial' in the population. If the fitness of the newly generated candidate solution is less than the best fitness value 'best_score', then the best fitness value is updated, the fitness of the candidate solution is assigned to the best fitness value 'best_score', and the candidate solution 'trial' is assigned to the best solution 'best_solution'.

[0051] Steps 3-4: Based on the best_solution, set the various phases of the antenna to form an antenna pattern with anti-interference capability for multi-target pointing, so as to realize multi-target tracking and interference suppression.

[0052] The present invention has the following advantages over the prior art:

[0053] 1) High-precision target pointing prediction. By fusing deep learning and feature engineering techniques, and combining temporal difference information and positioning system errors to construct multi-dimensional features, this method effectively correlates long-term continuous positioning results. This significantly improves the accuracy of pointing prediction for multiple moving targets and alleviates the pointing deviation problem caused by positioning delay and errors in traditional methods.

[0054] 2) Multi-target trajectory representation and reuse. A deep learning model is used to predict the trajectories of multiple moving targets. By reusing the prediction model, the collaborative prediction of the positions of each target is achieved, which solves the problems of limited input quantity and high representation complexity in traditional methods, thereby generating objective functions pointing to multiple targets more efficiently.

[0055] 3) Intelligent Anti-interference Beam Generation. Based on differential evolution calculation methods, the antenna phase is optimized, and the antenna pattern is dynamically adjusted to achieve intelligent beam generation that enhances signal strength at the target pointing location and suppresses interference signals in interference areas. This method effectively improves the beam's ability to track multiple moving targets through a multi-target pointing strategy, while simultaneously enhancing anti-interference performance. Attached Figure Description

[0056] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0057] Figure 2This illustrates the relationship between the number of iterations and the error value in the embodiments of the present invention.

[0058] Figure 3 This is the optimized antenna pattern according to an embodiment of the present invention. Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0060] A multi-target pointing intelligent anti-jamming beam generation method, the overall process of which is as follows: Figure 1 As shown, the specific steps include the following:

[0061] Step 1: Construct the dataset. Obtain continuous positioning information for multiple moving targets, including their position and relative pointing angle. Perform a difference operation on the relative pointing angles to obtain the difference in relative pointing angles between adjacent time points as supplementary features. Then, superimpose the positioning error range to construct feature data. Use the pointing angle at the last time point as sample data to form the dataset. The dataset includes feature data and sample data. Finally, split the dataset into training dataset and test dataset.

[0062] Step 1 includes the following specific processes:

[0063] Step 1-1: Obtain continuous localization information of multiple moving targets at consecutive time points T+1, including their positions at each time point. and and relative pointing angle T=7, and differential calculation is performed based on the relative pointing angle. Then add the positioning error range Constituent features and the pointing angle at time T+1. As a sample l, where:

[0064]

[0065] ;

[0066] Step 1-2: Repeat Step 1-1M times, M=550441, to form a dataset, including the feature set. and tag set ;

[0067] Steps 1-3: According to proportion =0.9 splits the dataset into training and testing datasets. Training feature set Training label set Test feature set and test tag set .

[0068] Step 2: Construct a dedicated intelligent model for predicting the pointing of moving targets based on a deep neural network. Train the dedicated intelligent model for predicting the pointing of moving targets using a training dataset, and select and save the optimal model based on its performance on the test dataset.

[0069] Step 2 includes the following specific steps:

[0070] Step 2-1: Construct a dedicated intelligent model for predicting the pointing of moving targets. The input is "input", and the output is "output". The computation process can be described as follows: , where f represents the computation process of the dedicated model.

[0071] The dedicated model structure consists of four parts: ConvBlock, UpConvBlock, Pooling layer, and OutLine.

[0072] A ConvBlock consists of 1D convolutional layers and batch normalization layers, used to extract local features of the input signal. The input parameters are the number of input channels (in_channels), the number of output channels (out_channels), the kernel size (kernel_size), and the padding.

[0073] Upsampling convolutional blocks (UpConvBlock) employ transposed convolutional layers to upsample and restore spatial dimensions. The input parameters are the number of input channels (in_channels), the number of output channels (out_channels), the kernel size (kernel_size), and the stride.

[0074] The pooling layer uses max pooling to reduce the spatial dimension of the input while preserving key features, thus aligning the dimensions between modules. The input parameters are the kernel size and stride.

[0075] The output block OutLine consists of 1D convolutional layers and fully connected layers, and is used to output the fitting results. The input parameters are the number of input channels in_channels, the number of output channels out_channels, the kernel size kernel_size, and the padding.

[0076] The processing procedure is as follows:

[0077] Based on the ConvBlock convolutional block, the parameters are set to instantiate the enc1 module, with in_channels=1, out_channels=64, kernel_size=3 and padding=1. The moving target is pointed to the input of the prediction intelligent dedicated model to obtain the feature xe1.

[0078] Based on the pooling layer, a pool module is instantiated with parameters set to kernel_size=2 and stride=2. This module processes feature xe1 to obtain the feature... ;

[0079] Based on the ConvBlock convolutional block, instantiate it as an enc2 module with the parameters set to in_channels=64, out_channels=128, kernel_size=3, and padding=1, and then use the features... The processing yields feature xe2;

[0080] Based on the pool module, feature xe2 is processed to obtain features. ;

[0081] Based on the ConvBlock convolutional block, instantiate it as an enc3 module with the parameters set to in_channels=128, out_channels=256, kernel_size=3, and padding=1, and then use the features... The processing yields feature xe3;

[0082] Based on the pool module, feature xe3 is processed to obtain features. ;

[0083] Based on the upsampling convolutional block UpConvBlock, instantiate it as an up3 module with the following parameters: in_channels=256, out_channels=128, kernel_size=2, and stride=2. The feature xu3 is obtained through processing;

[0084] xu3 and The data is concatenated according to the second dimension and merged into xw3;

[0085] Based on the ConvBlock convolutional block, instantiate it as a dec3 module with parameters set to in_channels=256, out_channels=128, kernel_size=3, and padding=1. Process the feature xw3 to obtain the feature. ;

[0086] Based on the upsampling convolutional block UpConvBlock, instantiate the up2 module with the parameters set to in_channels=128, out_channels=64, kernel_size=2, and stride=2, and then use the features... The feature xu2 is obtained through processing;

[0087] xu2 and The data is concatenated according to the second dimension and merged into xw2;

[0088] Based on the ConvBlock convolutional block, instantiate it as a dec2 module with parameters set to in_channels=128, out_channels=64, kernel_size=3, and padding=1. Process the feature xw2 to obtain... ;

[0089] Based on the parameter settings of the 1D convolutional layer in the output block OutLine, an out_conv module is instantiated with parameters set to in_channels=64, out_channels=1, kernel_size=2, and padding=1, which will then be used to store the features. The processing yields feature c1;

[0090] Dimensionality reduction is performed on the second dimension of feature c1;

[0091] Based on the settings of the fully connected layer in the output block OutLine, the module is instantiated as out_line with parameters set to in_features=17 and out_features=1. Feature c1 is then processed to obtain feature output.

[0092] Step 2-2: Initialize the variable epoch = 0 to record the number of training iterations, and initialize the minimum loss value Min_Loss to infinity to record the minimum loss value during training and testing; set the maximum number of training iterations num_epochs;

[0093] Steps 2-3: During the training phase, switch the model to training mode, randomly select B data points from the training feature set and training label set to construct a batch of data, then clear the gradient cache, perform forward propagation, calculate the model output, and use the mean squared error loss function to calculate the loss between the model output and the labels in the batch of data, perform back propagation, calculate the gradient, and update the model parameters; where B is a set value.

[0094] Steps 2-4: Enter the testing phase. Switch the model to evaluation mode, disable gradient calculation, perform forward propagation, output predicted data based on the test feature set, and calculate the loss value Loss of the model's output predicted data and the test label set using the mean squared error loss function. If the loss Loss is lower than the minimum loss value Min_Loss, save the current model and assign the test loss value to the minimum loss value Min_Loss; otherwise, increment the training count by 1 (epochs) and return to step 2-3. Repeat this process until all data in the training feature set and training label set has been processed or the maximum number of training counts (num_epochs) has been reached. The relationship between the number of counts and the error value is as follows: Figure 2 As shown.

[0095] Step 3: Generate an antenna pattern with anti-interference capabilities for multi-target pointing, enabling multi-target tracking and interference suppression. Based on existing multi-target localization information, a dedicated intelligent model for moving target pointing prediction is used to predict multi-target localization results. An objective function is constructed based on the multi-target localization results, and the antenna phase is optimized based on the objective function to obtain the optimal phase. Then, each phase of the antenna is set to the optimal phase, forming an antenna pattern with anti-interference capabilities for multi-target pointing, enabling multi-target tracking and interference suppression.

[0096] Step 3 includes the following specific steps:

[0097] Step 3-1: Construct feature data based on multi-target localization results to generate predicted multi-target localization results:

[0098] The optimal moving target pointing prediction intelligent dedicated model is loaded. Continuous positioning information of multiple moving targets requiring antenna beam pointing, including position, relative pointing angle, differential calculation results based on the relative pointing angle, and positioning error range, is sequentially input into the model to obtain the relative pointing angles of the moving targets requiring pointing. ,in P represents the number of targets to be targeted by the antenna beam, and its value is 2. The continuous positioning information of the multiple moving targets requiring antenna beam suppression, including their position, relative pointing angle, the result of differential calculation based on the relative pointing angle, and the positioning error range, is sequentially input into the model to obtain the relative angles of the moving targets to be suppressed. ,in Q represents the number of targets requiring antenna beam suppression, and its value is 2. The features, predicted values, and the difference between predicted values ​​and labels are shown in the table below:

[0099] direction feature Predicted value (degrees) Difference (degrees) Point to 1 -1.2997e+03, 1.0084e+04, -7.3443e+00,1.0000e+02, -1.6309e+03, 1.0032e+04, -9.2334e+00, 1.0000e+02, -1.9983e+03,1.0043e+04, -1.1253e+01, 1.0000e+02, -2.2845e+03, 9.9685e+03, -1.2908e+01,1.0000e+02, -2.6160e+03, 1.0011e+04, -1.4646e+01, 1.0000e+02, -2.8224e+03,1.0094e+04, -1.5622e+01, 1.0000e+02, -3.1817e+03, 1.0078e+04, -1.7521e+01,1.0000e+02, -1.8891e+00, -2.0195e+00, -1.6549e+00, -1.7378e+00, -9.7658e-01 -20.2392 -0.5836 Point to 2 2.0419e+03, 1.0094e+04, 1.1436e+01,1.0000e+02, 1.5921e+03, 9.9957e+03,9.0500e+00, 1.0000e+02, 1.2374e+03,9.9955e+03, 7.0571e+00, 1.0000e+02,1.0159e+03, 1.0061e+04, 5.7661e+00,1.0000e+02, 6.4937e+02, 9.9343e+03,3.7399e+00, 1.0000e+02, 2.6844e+02,9.9470e+03, 1.5458e+00, 1.0000e+02,2.0167e+00, 1.0017e+04, 1.1536e-02,1.0000e+02, -2.3856e+00, -1.9929e+00, -1.2910e+00, -2.0261e+00, -2.1941e+00 -2.3110 0.0666 Inhibit 1 5.0780e+03, 1.0093e+04, 2.6709e+01,1.0000e+02, 4.8241e+03, 9.9237e+03,2.5925e+01, 1.0000e+02, 4.4614e+03,9.9545e+03, 2.4141e+01, 1.0000e+02,4.0560e+03, 1.0037e+04, 2.2004e+01,1.0000e+02, 3.7096e+03, 1.0058e+04,2.0246e+01, 1.0000e+02, 3.3480e+03,9.9672e+03, 1.8567e+01, 1.0000e+02,3.0750e+03, 9.9179e+03, 1.7226e+01,1.0000e+02, -7.8317e-01, -1.7844e+00, -2.1369e+00, -1.7585e+00, -1.6785e+00 15.9584 0.4238 Inhibit 2 8.0641e+03, 1.0078e+04, 3.8667e+01,1.0000e+02, 7.6978e+03, 1.0005e+04,3.7574e+01, 1.0000e+02, 7.3159e+03,9.9180e+03, 3.6414e+01, 1.0000e+02,7.0731e+03, 9.9498e+03, 3.5408e+01,1.0000e+02, 6.5676e+03, 1.0006e+04,3.3279e+01, 1.0000e+02, 6.3171e+03,1.0087e+04, 3.2057e+01, 1.0000e+02,5.9991e+03, 9.9423e+03, 3.1106e+01,1.0000e+02, -1.0924e+00, -1.1607e+00, -1.0053e+00, -2.1297e+00, -1.2216e+00 30.5710 1.4679

[0100] Step 3-2: Construct a multi-target localization result generation objective function, with the input being the relative angles of all moving targets whose orientation needs to be determined. The relative angle of the moving target that needs to be suppressed and antenna array element vector The output is the antenna optimization effect; among which Phase of antenna elements composition, N is the number of antenna array elements;

[0101] The objective function is:

[0102]

[0103] Where d is the antenna spacing, which is 0.5 meters.

[0104] Step 3-3: Optimize the antenna phase based on the objective function. Perform population initialization by uniformly initializing a two-dimensional array as the initial population within the given phase range phase_range=[-π, π], representing the possible antenna phase values. The dimension is num_particles×N, i.e., 50×32, where num_particles is the number of individuals in the population and N is the number of antenna elements. Set the first individual as the current best solution best_solution and calculate and initialize it to the best fitness value best_score using the objective function. Perform a specified number of iterations, with the maximum iteration value set to num_iterations, i.e., 10000. After completing the iterations, execute step 3-4.

[0105] In each iteration, every individual in the population is traversed. An individual 'a' is randomly selected from the initial population as the current individual, and two different individuals 'b' and 'c' are randomly selected. A mutation operation is performed to form a new individual 'mutant', whose value is equal to a + F × (b - c), where F is the differential weight coefficient, and the value is restricted to the phase range 'phase_range'. Then, based on the crossover probability 'CR', two gene positions of the individual 'mutant' are randomly selected, and the two selected genes are exchanged to generate a new candidate solution 'trial'. If the fitness value of the newly generated candidate solution 'trial' is smaller than that of individual 'a', then individual 'a' is replaced by the candidate solution 'trial' in the population. If the fitness of the newly generated candidate solution is less than the best fitness value 'best_score', then the best fitness value is updated, the fitness of the candidate solution is assigned to the best fitness value 'best_score', and the candidate solution 'trial' is assigned to the best solution 'best_solution', with a value of [1.5400658 2.04739019 2.45940538 -2.91063359 -0.2561683]. 1.3712924 2.10726917 2.84834991 -3.13269703 -2.09081933 -0.60556471 1.43064659 2.40148527 -3.06793231 -2.67869118 -1.78223138 -1.11285596 1.49037194 -3.14159265 -2.90485418 -2.2222814 -1.73683538 -0.31108245 1.01680481 -3.00290515 -1.90966869 -1.6191792 -0.95787977 -1.01975984 0.42659694 -2.93486724 -1.50113334).

[0106] Steps 3-4: Based on the best_solution, set the various phases of the antenna to form an antenna pattern with anti-interference capability for multi-target pointing, such as... Figure 3 As shown, the antenna pattern gain is clearly visible at the required directions of -20.2392° and -2.3110°, while the gain is clearly suppressed at the required suppression directions of 15.9584° and 30.5710°, effectively achieving multi-target tracking and interference suppression.

Claims

1. A method for generating intelligent anti-interference beams targeting multiple targets, characterized in that, Includes the following steps: Step 1: Obtain continuous positioning information of multiple moving targets, including position and relative pointing angle, and perform differential operation on the relative pointing angle to obtain the difference of relative pointing angle between adjacent time moments as a supplement to the feature. Then, superimpose the positioning error range to construct feature data, and use the pointing angle at the last time moment as sample data to form a dataset. The dataset includes feature data and sample data. Finally, the dataset is split into training dataset and test dataset. Step 2: Construct a dedicated intelligent model for predicting the pointing of moving targets based on a deep neural network. Train the dedicated intelligent model for predicting the pointing of moving targets using a training dataset, and select and save the optimal model based on its performance on the test dataset. Step 3: Based on the existing multi-target localization information, use the intelligent dedicated model for moving target pointing prediction to predict the multi-target localization results. Construct an objective function based on the multi-target localization results, and optimize the antenna phase based on the objective function to obtain the optimal phase. Then, set each phase of the antenna to the optimal phase to form an antenna pattern that is oriented towards multiple targets and has anti-interference capabilities, thereby realizing multi-target tracking and interference suppression.

2. The intelligent anti-interference beam generation method for multi-target pointing according to claim 1, characterized in that, Step 1 is as follows: Step 1-1: Obtain continuous localization information of multiple moving targets at consecutive time points T+1, including their positions at each time point. and and relative pointing angle And perform difference calculations based on the relative pointing angle. Then add the positioning error range Constituent features and the pointing angle at time T+1. As a sample l, where: ； Step 1-2: Repeat step 1-1 a set number of times to form a dataset, including a feature set. and tag set ; Steps 1-3: According to proportion The dataset is divided into a training dataset and a test dataset. The training dataset includes a training feature set and a training label set, while the test dataset includes a test feature set and a test label set.

3. The intelligent anti-interference beam generation method for multi-target pointing according to claim 2, characterized in that, Step 2 is as follows: Step 2-1: Construct a dedicated intelligent model for predicting moving targets. The structure includes a convolutional block (ConvBlock), an upsampling convolutional block (UpConvBlock), a pooling layer (Pool), and an output block (OutLine). ConvBlock consists of 1D convolutional layers and batch normalization layers, used to extract local features of the input signal. The input parameters are the number of input channels in_channels, the number of output channels out_channels, the kernel size kernel_size, and padding. Upsampling convolutional blocks (UpConvBlock) employ transposed convolutional layers to upsample and restore spatial dimensions. The input parameters are the number of input channels (in_channels), the number of output channels (out_channels), the kernel size (kernel_size), and the stride. The pooling layer uses max pooling to reduce the spatial dimension of the input while preserving key features, thus aligning the dimensions between modules. The input parameters are the kernel size and stride. The output block OutLine consists of 1D convolutional layers and fully connected layers, and is used to output the fitting results. The input parameters are the number of input channels in_channels, the number of output channels out_channels, the kernel size kernel_size, and the padding. The processing procedure is as follows: Based on the parameters of the ConvBlock convolutional block, it is instantiated as the enc1 module, and the moving target is pointed to the input of the prediction intelligent dedicated model to obtain the feature xe1. Based on the pooling layer, parameters are set to instantiate a pool module, which processes feature xe1 to obtain features. ; Based on the ConvBlock convolutional block parameter setting instantiation into the enc2 module, the features The processing yields feature xe2; Based on the pooling layer, feature xe2 is processed to obtain features. ; Based on the ConvBlock convolutional block parameter setting instantiation as an enc3 module, the features The processing yields feature xe3; Based on the pooling layer, feature xe3 is processed to obtain features. ; Based on the upsampling convolutional block UpConvBlock, the parameters are instantiated as up3 module, and the features are... The feature xu3 is obtained through processing; xu3 and The data is concatenated according to the second dimension and merged into xw3; Based on the ConvBlock convolutional block, the parameters are instantiated as a dec3 module, and the feature xw3 is processed to obtain the feature. ; Based on the upsampling convolutional block UpConvBlock, the parameters are instantiated as up2 module, and the features are... The feature xu2 is obtained through processing; xu2 and The data is concatenated according to the second dimension and merged into xw2; Based on the ConvBlock convolutional block, the parameters are instantiated as a dec2 module, and the feature xw2 is processed to obtain... ; Based on the parameter settings of the 1D convolutional layer in the output block OutLine, the features are instantiated into the out_conv module. The processing yields feature c1; Dimensionality reduction is performed on the second dimension of feature c1; Based on the settings parameters of the fully connected layer in the output block OutLine, the out_line module is instantiated, and the feature c1 is processed to obtain the feature output. Step 2-2: Initialize the variable epoch = 0 to record the number of training iterations, and initialize the minimum loss value Min_Loss to infinity to record the minimum loss value during training and testing; set the maximum number of training iterations num_epochs; Steps 2-3: During the training phase, switch the model to training mode, randomly select B data points from the training feature set and training label set to construct a batch of data, then clear the gradient cache, perform forward propagation, calculate the model output, and use the mean squared error loss function to calculate the loss between the model output and the labels in the batch of data, perform back propagation, calculate the gradient, and update the model parameters; where B is a set value. Steps 2-4: Enter the testing phase. Switch the model to evaluation mode, disable gradient calculation, perform forward propagation, output predicted data based on the test feature set, and use the mean squared error loss function to calculate the loss value Loss of the model output predicted data and the test label set. If the loss Loss is lower than the minimum loss value Min_Loss, save the current model and assign the test loss value to the minimum loss value Min_Loss. Otherwise, increment the training count epoch by 1 and return to step 2-3 until all data in the training feature set and training label set are taken or the maximum training count num_epochs is reached.

4. The intelligent anti-interference beam generation method for multi-target pointing according to claim 1, characterized in that, Step 3 is as follows: Step 3-1: Construct feature data based on multi-target localization results to generate predicted multi-target localization results: The optimal moving target pointing prediction intelligent dedicated model is loaded. Continuous positioning information of multiple moving targets requiring antenna beam pointing, including position, relative pointing angle, differential calculation results based on the relative pointing angle, and positioning error range, is sequentially input into the model to obtain the relative pointing angles of the moving targets requiring pointing. ,in P represents the number of targets that need to be pointed towards by the antenna beam. The continuous positioning information of the multiple moving targets requiring antenna beam suppression, including their position, relative pointing angle, the result of differential calculation based on the relative pointing angle, and the positioning error range, is sequentially input into the model to obtain the relative angles of the moving targets to be suppressed. ,in Q represents the number of targets for which antenna beam suppression is required; Step 3-2: Construct a multi-target localization result generation objective function, with the input being the relative angles of all moving targets whose orientation needs to be determined. The relative angle of the moving target that needs to be suppressed and antenna array element vector The output is the antenna optimization effect; among which Phase of antenna elements composition, N is the number of antenna array elements; The objective function is: Where d is the antenna spacing; Step 3-3: Perform population initialization. Uniformly initialize a two-dimensional array as the initial population within the given phase range phase_range, representing the possible phase values ​​of the antenna elements. The dimension is num_particles×N, where num_particles is the number of individuals in the population and N is the number of antenna elements. Set the first individual as the current best solution, and calculate and initialize it to the best fitness value, best_score, through the objective function; perform a specified number of iterations, with the maximum iteration value set to num_iterations; and execute steps 3-4 after completing the iterations. In each iteration, every individual in the population is traversed. An individual 'a' is randomly selected from the initial population as the current individual, and two different individuals 'b' and 'c' are randomly selected. A mutation operation is performed to form a new individual 'mutant', whose value is equal to a + F × (b - c), where F is the difference weight coefficient, and the value is restricted to the phase range 'phase_range'. Then, according to the crossover probability 'CR', two gene positions of the individual 'mutant' are randomly selected, and the two selected genes are exchanged to generate a new candidate solution 'trial'. If the fitness value of the newly generated candidate solution 'trial' is smaller than that of individual 'a', then the original individual 'a' is replaced by the candidate solution 'trial' in the population. If the fitness of the newly generated candidate solution is less than the best fitness value 'best_score', then the best fitness value is updated, the fitness of the candidate solution is assigned to the best fitness value 'best_score', and the candidate solution 'trial' is assigned to the best solution 'best_solution'. Steps 3-4: Based on the best_solution, set the various phases of the antenna to form an antenna pattern with anti-interference capability for multi-target pointing, so as to realize multi-target tracking and interference suppression.

Images