Intelligent real-time beam forming method for radar antenna

Abstract

The invention relates to an intelligent real-time beam forming method for a radar antenna. The method comprises the following steps: S1, constructing a training sample library; s2, constructing a neural network beam former which takes the directional diagram as input and outputs a predicted value of a first weight matrix, and training the neural network beam former through the directional diagram and the corresponding weight matrix; s3, constructing a neural network null former which takes the first weight matrix and the null angle as input and outputs a predicted value of the second weight matrix, and training the neural network null former through the first weight matrix, the null angle and the second weight matrix; and S4, according to the angle and gain of the input directional diagram, completing on-orbit beam forming through the trained neural network beam former and neural network null generator. The method has the capability of all-day and all-weather sea observation, the dependence on instruction control of a ground measurement and control station is reduced, the detection efficiency is high, the information timeliness is high, and the acquisition time of sea surface target information is shortened.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to an intelligent real-time beamforming method for a radar antenna. Background Art

[0002] For a long time, large-scale sea surface ship target detection has been a hot topic in scientific research. Due to the influence of foggy weather at sea, optical sensors can hardly play their advantages. The satellite-borne synthetic aperture radar (SAR) system can penetrate fog and has the ability to observe the sea all day and all weather, which is very suitable for the detection and tracking of sea surface ship targets. However, the existing satellite-borne SAR system is heavily dependent on the injection commands from the ground measurement and control station, and cannot track the detection targets of interest in a timely manner, which greatly reduces the timeliness of satellite-borne SAR information, especially for wide-band search and long-term tracking. It is difficult to achieve multiple working modes, so beamforming technology is needed to adjust the phase and amplitude of each antenna unit in the radar antenna array to control the directionality and shape of the radar beam to track the detection targets of interest.

[0003] The traditional beamforming method is simple to implement, has low computational complexity, and has relatively stable performance, but it cannot intelligently generate null sinks according to the interference direction to reduce the impact of interference, thus affecting the observation capability. Although the scheme based on multi-objective optimization (genetic algorithm, bee colony algorithm, etc.) can generate null sinks to suppress interference signals, the computational complexity is too large to track the detection target of interest in a timely manner. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a radar antenna intelligent real-time beamforming method, a training neural network beamformer and a neural network null generator, so as to realize real-time beamforming of any angle and no less than 4 nulls, which has the advantages of high robustness and fast calculation speed.

[0005] To achieve the above-mentioned object of the invention, the present invention provides a radar antenna intelligent real-time beamforming method, comprising the following steps:

[0006] Step S1, constructing a training sample library, wherein each sample in the training sample library includes a directional pattern and a first weight matrix corresponding to the directional pattern, a null angle and a second weight matrix corresponding to the null angle;

[0007] Step S2, constructing a neural network beamformer, training the neural network beamformer through the directional patterns in the training sample library and their corresponding weight matrices, wherein the neural network beamformer takes the directional patterns as input and outputs the predicted values ​​of the first weight matrix;

[0008] Step S3, constructing a neural network null sink former, training the neural network null sink former through the directional map in the training sample library, the first weight matrix, the null sink angle and the second weight matrix, wherein the neural network null sink former takes the first weight matrix and the null sink angle as input and outputs a predicted value of the second weight matrix;

[0009] Step S4: According to the input directional diagram, the trained neural network beamformer and the neural network null generator, on-orbit beamforming is completed.

[0010] According to a technical solution of the present invention, in the step S1, it specifically includes:

[0011] Step S11, generating a first simulation database by using a traditional MVDR beamformer, wherein samples of the first simulation database include a first weight matrix and its corresponding directional pattern;

[0012] Step S12: Generate a second simulation database by using a genetic algorithm beamformer, wherein samples of the second simulation database include a first weight matrix, a null angle, and a second weight matrix corresponding to the null angle.

[0013] According to a technical solution of the present invention, in step S11, it specifically includes:

[0014] Assume that the array 40 is a two-dimensional uniform array of M×N, and the antenna units are of the same type, size, and feeding amplitude. The antenna unit spacing 401 along the y direction and the antenna unit spacing 402 along the x direction are d x and d y , the number of units is M and N respectively, the wavelength is λ, and the step is 1°. Use the traditional MVDR beamformer to generate beams, calculate and save the first weight matrix w and the directional pattern It means that the antenna has different receiving capabilities for electromagnetic waves coming from different directions. The value range of azimuth angle θ is 0-180°, and the elevation angle The value range of is 0-180°, then the first weight matrix W and the array response vector The calculation formula is

[0015]

[0016] Among them, R -1 Represents the inverse matrix of the R matrix, the array response vector The array response vector of a uniform linear array arranged along the x and y directions is and The calculation formulas are

[0017]

[0018] According to a technical solution of the present invention, in the step S12, it specifically includes:

[0019] Step S121, chromosome encoding and population initialization: The weighted vector of the array antenna is used as the chromosome of the genetic algorithm, the number of chromosomes is P, and the dimension of each chromosome is a decimal encoding vector of L = M*N. The optimized search initial point is established, and the initial parameter value of the chromosome is W = w1, w2, ..., w L , where w L is the gene on the Lth array element, that is, the weighted current amplitude of the Lth array element, L is the number of array elements of the array antenna, where L = M*N;

[0020] Assuming that all randomly initialized populations conform to Gaussian distribution, the population is expressed as

[0021] C=δ·W P×L

[0022] Among them, δ is the disturbance amplitude, P is the group size, L is the number of array elements of the array antenna, and W P×L is a randomly initialized matrix, C=[C1,C2,…,C P ] T Initialize the matrix to be P×L;

[0023] Step S122, evaluate the quality of chromosomes: calculate the chromosome fitness of the group through the fitness function f, expressed as

[0024]

[0025] Among them, α and β are weight coefficients, is the maximum relative sidelobe level, SLVL is the sidelobe gain, max is the maximum value function, S is the sidelobe area of ​​the pattern, NLVL is the null depth, It means that at a given θ, Generate a zero depression not less than the depth NLVL in the direction;

[0026] Step S123, judging whether the calculated chromosome fitness meets the requirement according to the fitness threshold, if yes, executing step S127, otherwise executing step S124;

[0027] Step S124, selection operation: use the roulette method to extract the excellent chromosomes in the population. First, calculate the W of each chromosome in the population. l The cumulative probability q l , cumulative probability q l Expressed as

[0028]

[0029] Wherein, l represents the chromosome number, l = 1 to P;

[0030] In the interval (0,q P ) generates a uniformly distributed random number r l , if q l-1 <r l ≤q l , then select the lth chromosome w l , repeat the above steps for P times to obtain P replicated chromosomes;

[0031] Step S125, crossover operation: generate P random numbers t in the interval [0,1] s1 , where s1 represents the s1th chromosome, 1≤s1≤P; if the random number t s1 Less than the crossover probability threshold t c1 Then select chromosome W from the population s1 As the parent generation, the selected parent chromosomes are randomly paired and crossed to produce new offspring chromosomes W l ';

[0032] Step S126, mutation operation: use the random number generation function to generate P random numbers t in the interval [0,1] s2 , where s2 represents the s2th chromosome, 1≤s2≤P; if t s2 Less than the crossover probability threshold t c2 Then select the s2th chromosome W from the population s2 And perform random number mutation in the mutation direction r to generate the mutant offspring chromosome W s2 ', execute step S122;

[0033] Step S126: According to the chromosome W of the group l The fitness of meets the fitness threshold, forming the azimuth, pitch angle, and weight data pair Save it to the second simulation database.

[0034] According to a technical solution of the present invention, the step S1 further includes:

[0035] Step S13, generating beamforming simulation data based on deep learning and genetic algorithm;

[0036] In the step S13, it specifically includes:

[0037] Randomly generate k interference angles at angles [0°, 180°] As a zero-pit generation list, k≤4;

[0038] Input the directional pattern to the neural network beamformer and generate a predicted value W' of the first weight matrix at [0°, 180°];

[0039] The predicted value W' of the first weight matrix and the null angle that needs to be oriented to generate the null The output result is the predicted value of the second weight matrix of the neural network null sink generator according to the angle prediction of the directional null sink generator. And use the predicted values ​​of the second weight matrix Add L random noises to establish the optimized search starting point, with the initial parameter value of the chromosome as Generate simulation data based on genetic algorithm to generate no less than 200 sample pairs;

[0040] When the generated sample pairs are used to train the neural network null former, the training is stopped when the loss no longer decreases within 5 rounds, and the beamforming simulation data generation based on deep learning and genetic algorithm is continued.

[0041] According to a technical solution of the present invention, in the step S2, it specifically includes:

[0042] The neural network beamformer is constructed and trained by taking the directional patterns in the training sample library as input and the predicted values ​​of the first weight matrix as output;

[0043] The input of the neural network beamformer is the directional pattern in the training sample library The output is the predicted value W' of the first weight matrix, based on the directional map in the training sample library The corresponding first weight matrix W is the true value, and the loss is calculated with the predicted value W' of the first weight matrix, using the mean square error loss function J(W,W') = (W-W') 2 Calculate the losses of W and W', use SGD as the optimizer, set the learning rate to 0.01, train for 100 rounds and back-propagate to update the neural network weights, and save the optimal training model;

[0044] According to a technical solution of the present invention, in step S3, it specifically includes:

[0045] The neural network null sink former is constructed and trained by taking the first weight matrix and the null sink angle in the training sample library as input, taking the predicted value of the second weight matrix as output, and taking the second weight matrix corresponding to the null sink angle in the training sample library as the true value W';

[0046] The output of the neural network null former is the predicted value of the second weight matrix Using the mean squared error loss function Calculate W' and The loss was 0.01, SGD was used as the optimizer, the learning rate was 0.01, and the neural network weights were updated by backpropagation. The training was repeated for 100 rounds, and the optimal training model was saved.

[0047] According to a technical solution of the present invention, the neural network beamformer includes:

[0048] The first fully connected layer is used for feature refinement and forward calculation;

[0049] L second fully connected layers, used for feature refinement and adjustment of neuron dimensions, where L = M*N is the number of array elements;

[0050] The first convolutional layer has a kernel size of (1,1), a stride of 1, and a padding of 0;

[0051] The first batch of normalization layers is calculated as x = γ * x norm +β, where γ and β are two learnable parameters, with initial values ​​set to 0 and 1; ε is a constant, set to 0.0001, to prevent the denominator from being 0; is the mean, b is the batch size of a single forward pass, σ 2 =1 / bΣ(x-μ) 2 , σ 2 is the variance;

[0052] The first activation function layer, whose activation function is relu=max(0,x), max is the maximum value function;

[0053] The second convolutional layer has a convolution kernel size of (1,1), a stride of 1, and a padding of 0;

[0054] The second batch of normalization layers and the second activation function layer.

[0055] According to a technical solution of the present invention, the input of the neural network null former includes two paths, the first input of the neural network null former includes a third convolutional layer, a third batch normalization layer, a third activation function layer, a fourth convolutional layer, a fourth batch normalization layer and a fourth activation function layer connected in sequence, the input of the third convolutional layer is a weight matrix, the size of its convolution kernel is (1,1), the step size is 1, and it is filled with 0; the activation function of the third activation function layer is relu=max(0,x); the size of the convolution kernel of the fourth convolutional layer is (1,1), the step size is 1, and it is filled with 0;

[0056] The second input of the neural network null former includes a third fully connected layer, a fourth fully connected layer, a fifth convolutional layer, a fifth batch normalization layer, and a fifth activation function layer; the input of the third fully connected layer is a null angle; the number of neurons in the third fully connected layer is 128, which is used for feature refinement and forward calculation; the number of neurons in the fourth fully connected layer is L, which is used for feature refinement and adjustment of the dimension of neurons, where L=M*N is the number of array elements;

[0057] The output results of the third activation function layer and the fifth activation function layer are flattened and added together, and then output after the sixth convolutional layer, the sixth batch normalization layer and the sixth activation function layer. The output result is the weight matrix predicted by the neural network null sink generator according to the null sink angle.

[0058] According to a technical solution of the present invention, in the step S4, it specifically includes:

[0059] In the on-orbit beamforming stage, the predicted value of the first weight matrix is ​​obtained by inputting the directional pattern into the neural network beamformer, and then the predicted value of the first weight matrix and the list of null angles with n null angles are input into the neural network null angle former, and the predicted value of the second weight matrix is ​​formed in a loop. When the directional pattern satisfies the requirement of having n null angles or the iteration reaches the threshold, the loop is stopped, the beam is formed, and the target is tracked.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The radar antenna intelligent real-time beamforming method proposed in the present invention has the advantages of high signal-to-interference ratio and low peak-to-sidelobe ratio. The present invention realizes beamforming of any angle and no less than 4 nulls through a deep learning algorithm, and has the advantages of high detection efficiency, high information timeliness, high robustness and fast calculation speed. It significantly improves the intelligent autonomy of the satellite-borne SAR system and reduces the dependence on command control from ground measurement and control stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 A technical block diagram schematically showing a radar antenna intelligent real-time beamforming method according to an embodiment of the present invention;

[0064] Figure 2A schematic diagram of an antenna array model used in an embodiment of the present invention is schematically shown;

[0065] Figure 3 A flowchart schematically shows an anti-interference beamforming method based on a genetic algorithm according to an embodiment of the present invention;

[0066] Figure 4 A schematic diagram of a deep learning-based beamforming and nulling process according to an embodiment of the present invention is schematically shown;

[0067] Figure 5 A flowchart of on-orbit beamforming according to an embodiment of the present invention is schematically shown;

[0068] Figure 6 A network structure diagram schematically showing a neural network beamformer according to an embodiment of the present invention;

[0069] Figure 7 A network structure diagram schematically showing a neural network null former according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The description of the embodiments of this specification should be combined with the corresponding drawings, which should be considered as part of the complete specification. In the drawings, the shape or thickness of the embodiments may be enlarged and indicated for simplification or convenience. Furthermore, the parts of each structure in the drawings will be described separately. It is worth noting that the elements not shown in the drawings or not described in words are in a form known to ordinary technicians in the relevant technical field.

[0071] The description of the embodiments herein and any reference to directions and orientations are only for the convenience of description and are not to be construed as any limitation on the scope of protection of the present invention. The following description of the preferred embodiments may involve combinations of features, which may exist independently or in combination, and the present invention is not particularly limited to the preferred embodiments. The scope of the present invention is defined by the claims.

[0072] This embodiment is based on the method of the present invention, such as Figure 1 As shown, it is divided into three stages: simulation database building 10, model building and training 20, and on-orbit beamforming 30, which specifically include the following steps:

[0073] Step S1, constructing a training sample library, each sample in the training sample library includes a directional pattern and a first weight matrix corresponding to the directional pattern, a null angle and a second weight matrix corresponding to the null angle;

[0074] Step S11, generating a first simulation database by using a traditional MVDR beamformer, wherein samples of the first simulation database include a first weight matrix and its corresponding directional pattern;

[0075] In step S11, it specifically includes:

[0076] The antenna array model used in this embodiment is as follows: Figure 2 As shown, it is assumed that the array is a two-dimensional uniform array of M×N, and the antenna units are of the same type, size, and feeding amplitude. The spacing between antenna units in the x and y directions is d x and d y , the number of units is M and N respectively, the wavelength is λ, and the step is 1°. Use the traditional MVDR beamformer to generate beams, calculate and save the first weight matrix W and the directional pattern The azimuth angle θ ranges from 0 to 180°, and the elevation angle The value range of is 0-180°, then the first weight matrix W and the array response vector The calculation formula is

[0077]

[0078] Among them, R- 1 Represents the inverse matrix of the R matrix, the array response vector The array response vector of a uniform linear array arranged along the x and y directions is and The calculation formulas are:

[0079]

[0080] Step S12, genetic algorithm beam simulation part, generates a second simulation database through a genetic algorithm beamformer, and the samples of the second simulation database include a first weight matrix, a nulling angle, and a second weight matrix corresponding to the nulling angle.

[0081] A genetic algorithm based beamformer is used to simulate a weight matrix that is resistant to interference.

[0082] like Figure 3 As shown, in step S12, it specifically includes:

[0083] Step S121, chromosome encoding and population initialization: The weighted vector of the array antenna is used as the chromosome of the genetic algorithm, the number of chromosomes is P, and the dimension of each chromosome is a decimal encoding vector of L = M*N. The optimized search initial point is established, and the initial parameter value of the chromosome is W = w1, w2, ..., w L , where w L is the gene on the Lth array element, that is, the weighted current amplitude of the Lth array element, L is the number of array elements of the array antenna, where L = M*N;

[0084] Assuming that all randomly initialized populations conform to Gaussian distribution, the population is expressed as

[0085] C=δ·W P×L

[0086] Among them, δ is the disturbance amplitude, P is the group size, L is the number of array elements of the array antenna, and W P×L is a randomly initialized matrix, C=[C1,C2,…,C P ] T Initialize the matrix to be P×L;

[0087] Step S122, evaluate the quality of chromosomes: calculate the chromosome fitness of the group through the fitness function f, expressed as

[0088]

[0089] Among them, α and β are weight coefficients, is the maximum relative sidelobe level, SLVL is the sidelobe gain, max is the maximum value function, S is the sidelobe area of ​​the pattern, NLVL is the null depth, It means that at a given θ, Generate a zero depression not less than the depth NLVL in the direction;

[0090] Step S123, judging whether the calculated chromosome fitness meets the requirement according to the fitness threshold, if yes, executing step S127, otherwise executing step S124;

[0091] Step S124, selection operation: use the roulette method to extract the excellent chromosomes in the population. First, calculate the W of each chromosome in the population. l The cumulative probability q l , expressed as

[0092]

[0093] Wherein, l represents the chromosome number, l = 1 to P;

[0094] In the interval (0,q P ) generates a uniformly distributed random number r l , if q l-1 <r l ≤q l , then select the lth chromosome w l , repeat the above steps for P times to obtain P replicated chromosomes;

[0095] Step S125, crossover operation: generate P random numbers t in the interval [0,1] s1 , where s1 represents the s1th chromosome, 1≤s1≤P; if the random number ts1 Less than the crossover probability threshold t c1 Then select chromosome W from the population s1 As the parent generation, the selected parent chromosomes are randomly paired and crossed to produce new offspring chromosomes W l ';

[0096] Step S126, mutation operation: use the random number generation function to generate P random numbers t in the interval [0,1] s2 , where s2 represents the s2th chromosome, 1≤s2≤P; if t s2 Less than the crossover probability threshold t c2 Then select the s2th chromosome W from the population s2 And perform random number mutation in the mutation direction r to generate the mutant offspring chromosome W s2 ', execute step S122;

[0097] Step S126: According to the chromosome W of the group l The fitness of meets the fitness threshold, forming the azimuth, pitch angle, and weight data pair Save it to the second simulation database.

[0098] Step S1 also includes:

[0099] Step S13, generating beamforming simulation data based on deep learning and genetic algorithm;

[0100] like Figure 4 As shown, in step S13, it specifically includes:

[0101] Randomly generate k interference angles at angles [0°, 180°] As a zero-pit generation list, k≤4;

[0102] The directional pattern is input to the neural network beamformer, and the predicted value W of the first weight matrix is ​​generated at [0°, 180°];

[0103] The predicted value W' of the first weight matrix and the null angle that needs to be oriented to generate the null Input to the neural network null sink generator, the output result is the predicted value of the second weight matrix predicted by the neural network null sink generator according to the angle of the directional null sink And use the predicted values ​​of the second weight matrix Add L random noises to establish the optimized search starting point, with the initial parameter value of the chromosome as Generate simulation data based on genetic algorithm to generate no less than 200 sample pairs;

[0104] When the generated sample pairs are used to train the neural network null former, the training is stopped when the loss does not decrease within 5 rounds, and the beamforming simulation data generation based on deep learning and genetic algorithm is continued.

[0105] Step S14: Generate a third simulation database using the measured data.

[0106] Therefore, the simulation data may include simulation data generated by a traditional MVDR beamformer, simulation data generated by a genetic algorithm beamformer, simulation data generated by a genetic algorithm + deep learning hybrid beamformer, and simulation data generated based on measured data.

[0107] Step S2, constructing a neural network beamformer, through the directional pattern and its corresponding weight matrix neural wave network beamformer, the neural wave network beamformer takes the directional pattern as input and outputs the predicted value of the first weight matrix;

[0108] In step S2, it specifically includes:

[0109] A neural network beamformer is constructed and trained with the directional pattern as input and the first weight matrix as output.

[0110] The input of the neural network beamformer is the directional pattern in the training sample library The output is the predicted value W' of the first weight matrix, based on the directional map in the training sample library. The corresponding first weight matrix W is the true value, and the loss is calculated with the predicted value W' of the first weight matrix, using the mean square error loss function J(W,W') = (W-W') 2 Calculate the losses of W and W', use SGD as the optimizer, set the learning rate to 0.01, train for 100 rounds and back-propagate to update the neural network weights, and save the optimal training model.

[0111] like Figure 5 As shown, the neural network beamformer includes a first fully connected layer, L second fully connected layers, a first convolutional layer, a first batch of normalization layers, a first activation function layer, a second convolutional layer, a second batch of normalization layers and a second activation function layer, which are connected in sequence.

[0112] The first fully connected layer is used for feature refinement and forward calculation. The L second fully connected layers are used for feature refinement and adjusting the neuron dimension, where L = M*N is the number of array elements. The convolution kernel size of the first convolution layer is (1,1), the step size is 1, and the padding is 0. The calculation method of the first batch of normalization layers is x' = γ*x' norm +β, where γ and β are two learnable parameters, with initial values ​​set to 0 and 1; ε is a constant, set to 0.0001, to prevent the denominator from being 0; is the mean, b is the batch size of a single forward pass, σ 2 =1 / bΣ(x'-μ) 2 , σ 2 is the variance. The activation function of the first activation function layer is relu=max(0,x'), where max is the maximum value function. The size of the convolution kernel of the second convolution layer is (1,1), the step size is 1, and the padding is 0.

[0113] The input of the neural network beamformer is the azimuth angle θ and the elevation angle Direction diagram And connect the first fully connected layer, then use the calculation result of the first fully connected layer as the input of the second fully connected layer, and use the running result as the input of the first convolutional layer, and input the running result to the first batch of normalization layers, and use the running result as the input of the second convolutional layer, and input the running result to the second batch of normalization layers and the second activation function layer. The output result is the weight matrix predicted by the neural network.

[0114] Step S3, constructing a neural network null sink former, the network structure of the neural network null sink former is trained by the first weight matrix, the null sink angle and the second weight matrix, and the neural network null sink former is trained by the first weight matrix W' and the null sink angle As input, output the predicted value of the second weight matrix;

[0115] like Figure 6 As shown, in step S3, it specifically includes:

[0116] The first weight matrix and the null angle in the training sample library are used as input, the predicted value of the second weight matrix is ​​used as output, and the second weight matrix W" corresponding to the null angle in the training sample library is used as the true value to construct and train a neural network null former.

[0117] The output of the neural network null former is the predicted value of the second weight matrix Using the mean squared error loss function Calculate W" and The loss was 0.01, SGD was used as the optimizer, the learning rate was 0.01, and the neural network weights were updated by backpropagation. The training was repeated for 100 rounds, and the optimal training model was saved.

[0118] The input of the neural network null former includes two paths. The first input of the neural network null former includes the third convolution layer, the third batch normalization layer, the third activation function layer, the fourth convolution layer, the fourth batch normalization layer and the fourth activation function layer connected in sequence. The input of the third convolution layer is a weight matrix, and the size of its convolution kernel is (1,1), the step size is 1, and the padding is 0; the activation function of the third activation function layer is relu=max(0,x'); the size of the convolution kernel of the fourth convolution layer is (1,1), the step size is 1, and the padding is 0;

[0119] The second input of the neural network null former includes the third fully connected layer, the fourth fully connected layer, the fifth convolutional layer, the fifth batch normalization layer, and the fifth activation function layer; the input of the third fully connected layer is the null angle; the number of neurons in the third fully connected layer is 128, which is used for feature refinement and forward calculation; the number of neurons in the fourth fully connected layer is L, which is used for feature refinement and adjustment of the dimension of neurons, where L = M*N is the number of array elements;

[0120] The output results of the third activation function layer and the fifth activation function layer are flattened and added together, and then output after the sixth convolutional layer, the sixth batch normalization layer and the sixth activation function layer. The output result is the weight matrix predicted by the neural network null sink generator according to the null sink angle.

[0121] Therefore, the neural network null sink former can be trained based on the simulation data of the traditional beamformer and the genetic algorithm with the null sink formed at the specified angle. The input of the neural network null sink former is the weight vector of the traditional beamforming algorithm and the list of angles that need to generate directional null sinks. The true value is the weight vector of the genetic algorithm beamformer that can generate null sinks at a specific angle. The neural network null sink former is constructed and trained.

[0122] Step S4: According to the angle and gain of the input directional pattern, the on-orbit beamforming is completed according to the trained neural network beamformer and neural network null generator.

[0123] like Figure 7 As shown, in step S4, it specifically includes:

[0124] During the on-orbit beamforming phase, the input pattern To the neural network beamformer, get the predicted value of the first weight matrix, then input the predicted value of the first weight matrix and the null angle list with n null angles into the neural network null angle former, loop to form the predicted value of the second weight matrix, when the direction map When n null angles are met or the iteration reaches the threshold, the loop stops, a beam is formed, and the target is tracked.

[0125] The radar antenna intelligent real-time beamforming method of this embodiment establishes a large number of high-quality sample databases with high signal-to-interference ratio and high signal-to-clutter ratio by utilizing traditional beamforming methods, multi-objective optimized beamforming methods and measured data, trains a neural network beamformer and a neural network null generator, and uses the directional pattern as the weight matrix of the neural network inversion array. It can realize real-time beam intelligent forming of any angle and no less than 4 nulls, with an inference speed of no less than 40ms, a sidelobe level lower than -30dB, and a zero point level lower than -60dB. It has the advantages of high detection efficiency, high information timeliness and good robustness, significantly improves the intelligent autonomy of the satellite-borne SAR system, and reduces the dependence on command control from ground measurement and control stations.

[0126] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0127] Finally, it should be noted that the above is a preferred embodiment of the present invention. It should be pointed out that although the preferred embodiment of the present invention has been described, for those skilled in the art, once the basic creative concept of the present invention is known, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention. Therefore, the attached claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. A radar antenna intelligent real-time beamforming method, characterized in that: The following steps are involved: Step S1, constructing a training sample library, wherein each sample in the training sample library includes a directional pattern and a first weight matrix corresponding to the directional pattern, a null angle and a second weight matrix corresponding to the null angle; Step S2, constructing a neural network beamformer, training the neural network beamformer through the directional patterns in the training sample library and their corresponding weight matrices, wherein the neural network beamformer takes the directional patterns as input and outputs the predicted values ​​of the first weight matrix; Step S3, constructing a neural network null sink former, training the neural network null sink former through the directional map in the training sample library, the first weight matrix, the null sink angle and the second weight matrix, wherein the neural network null sink former takes the first weight matrix and the null sink angle as input and outputs a predicted value of the second weight matrix; Step S4: According to the input directional diagram, the trained neural network beamformer and the neural network null generator, on-orbit beamforming is completed.

2. The radar antenna intelligent real-time beamforming method according to claim 1, characterized in that: In the step S1, it specifically includes: Step S11, generating a first simulation database by using a traditional MVDR beamformer, wherein samples of the first simulation database include a first weight matrix and its corresponding directional pattern; Step S12, generate a second simulation database through a genetic algorithm beamformer, wherein the samples of the second simulation database include a first weight matrix, a null angle, and a second weight matrix corresponding to the null angle. (Use a genetic algorithm beamformer, a genetic algorithm and a deep learning hybrid beamformer to generate a weight matrix at an azimuth angle of 0-180° and an elevation angle of 0-180°, with the signal-to-interference ratio, signal-to-clutter ratio, and null depth as the target, to form a null at a specified azimuth, and to form a directional pattern, establish a mapping relationship between the directional pattern and the weight matrix, and provide a large number of data samples for intelligent real-time beamforming generated by directional null).

3. The radar antenna intelligent real-time beamforming method according to claim 2, characterized in that: In step S11, it specifically includes: Assume that the array is a two-dimensional uniform array of M×N, and the antenna elements are of the same type, size, and feeding amplitude. The spacing between antenna elements in the x and y directions is d x and d y , the number of units is M and N respectively, the wavelength is λ, and the step is 1°. Use the traditional MVDR beamformer to generate beams, calculate and save the first weight matrix W and the directional pattern The azimuth angle θ ranges from 0 to 180°, and the elevation angle The value range of is 0-180°, then the first weight matrix W and the array response vector The calculation formula is Among them, R -1 Represents the inverse matrix of the R matrix, the array response vector The array response vector of a uniform linear array arranged along the x and y directions is and The calculation formulas are:

4. The radar antenna intelligent real-time beamforming method according to claim 2, characterized in that: In the step S12, it specifically includes: Step S121, chromosome encoding and population initialization: The weighted vector of the array antenna is used as the chromosome of the genetic algorithm, the number of chromosomes is P, and the dimension of each chromosome is a decimal encoding vector of L = M*N. The optimized search initial point is established, and the initial parameter value of the chromosome is W = w1, w2, ..., w L , where w L is the gene on the Lth array element, that is, the weighted current amplitude of the Lth array element, L is the number of array elements of the array antenna, where L = M*N; Assuming that all randomly initialized populations conform to Gaussian distribution, the population is expressed as C=δ·W P×L Among them, δ is the disturbance amplitude, P is the group size, L is the number of array elements of the array antenna, and W P×L is a randomly initialized matrix, C=[C1,C2,…,C P ] T Initialize the matrix to be P×L; Step S122, evaluate the quality of chromosomes: calculate the chromosome fitness of the group through the fitness function f, expressed as Among them, α and β are weight coefficients, is the maximum relative sidelobe level, SLVL is the sidelobe gain, max is the maximum value function, S is the sidelobe area of ​​the pattern, NLVL is the null depth, It means that at a given θ, Generate a zero depression not less than the depth NLVL in the direction; Step S123, judging whether the calculated chromosome fitness meets the requirement according to the fitness threshold, if yes, executing step S127, otherwise executing step S124; Step S124, selection operation: use the roulette method to extract the excellent chromosomes in the population. First, calculate the W of each chromosome in the population. l The cumulative probability q l , expressed as Wherein, l represents the chromosome number, l = 1 to P; In the interval (0,q P ) generates a uniformly distributed random number r l , if q l-1 <r l ≤q l , then select the lth chromosome W l , repeat the above steps for P times to obtain P replicated chromosomes; Step S125, crossover operation: generate P random numbers t in the interval [0,1] s1 , where s1 represents the s1th chromosome, 1≤s1≤P; if the random number t s1 Less than the crossover probability threshold t c1 Then select chromosome W from the population s1 As the parent generation, the selected parent chromosomes are randomly paired and crossed to produce new offspring chromosomes W l '; Step S126, mutation operation: use the random number generation function to generate P random numbers t in the interval [0,1] s2 , where s2 represents the s2th chromosome, 1≤s2≤P; if t s2 Less than the crossover probability threshold t c2 Then select the s2th chromosome W from the population s2 And perform random number mutation in the mutation direction r to generate the mutant offspring chromosome W s2 ', execute step S122; Step S126: According to the chromosome W of the group l The fitness of meets the fitness threshold, forming the azimuth, pitch angle, and weight data pair Save it to the second simulation database.

5. The radar antenna intelligent real-time beamforming method according to claim 2, characterized in that: The step S1 also includes: Step S13, generating beamforming simulation data based on deep learning and genetic algorithm; In the step S13, it specifically includes: Randomly generate k interference angles at angles [0°, 180°] As a zero-pit generation list, k≤4; Input the directional pattern to the neural network beamformer and generate a predicted value W' of the first weight matrix at [0°, 180°]; The predicted value W' of the first weight matrix and the null angle that needs to be oriented to generate the null The output result is the predicted value of the second weight matrix of the neural network null sink generator according to the angle prediction of the directional null sink generator. And use the predicted values ​​of the second weight matrix Add L random noises to establish the optimized search starting point, with the initial parameter value of the chromosome as Generate simulation data based on genetic algorithm to generate no less than 200 sample pairs; When the generated sample pairs are used to train the neural network null former, the training is stopped when the loss no longer decreases within 5 rounds, and the beamforming simulation data generation based on deep learning and genetic algorithm is continued.

6. The radar antenna intelligent real-time beamforming method according to claim 1, characterized in that: In the step S2, it specifically includes: The neural network beamformer is constructed and trained by taking the directional patterns in the training sample library as input and the predicted values ​​of the first weight matrix as output; The input of the neural network beamformer is the directional pattern in the training sample library The output is the predicted value W' of the first weight matrix, based on the directional map in the training sample library The corresponding first weight matrix W is the true value, and the loss is calculated with the predicted value W' of the first weight matrix, using the mean square error loss function J(W,W') = (W-W') 2 Calculate the losses of W and W', use SGD as the optimizer, set the learning rate to 0.01, train for 100 rounds and back-propagate to update the neural network weights, and save the optimal training model.

7. The radar antenna intelligent real-time beamforming method according to claim 1, characterized in that: In step S3, it specifically includes: Taking the first weight matrix and the null angle in the training sample library as input, taking the predicted value of the second weight matrix as output, and taking the second weight matrix W" corresponding to the null angle in the training sample library as the true value, constructing and training the neural network null former; The output of the neural network null former is the predicted value of the second weight matrix Using the mean squared error loss function Calculate W" and The loss was 0.01, SGD was used as the optimizer, the learning rate was 0.01, and the neural network weights were updated by backpropagation. The training was repeated for 100 rounds, and the optimal training model was saved.

8. The radar antenna intelligent real-time beamforming method according to claim 1, characterized in that: The neural network beamformer comprises: The first fully connected layer is used for feature refinement and forward calculation; L second fully connected layers, used for feature refinement and adjustment of neuron dimensions, where L = M*N is the number of array elements; The first convolutional layer has a kernel size of (1,1), a stride of 1, and a padding of 0; The first batch of normalization layers is calculated as x'=γ*x' norm +β, where γ and β are two learnable parameters, with initial values ​​set to 0 and 1; ε is a constant, set to 0.0001, to prevent the denominator from being 0; is the mean, b is the batch size of a single forward pass, σ 2 =1 / bΣ(x'-μ) 2 , σ 2 is the variance; The first activation function layer, whose activation function is relu=max(0,x'), max is the maximum value function; The second convolutional layer has a convolution kernel size of (1,1), a stride of 1, and a padding of 0; The second batch of normalization layers and the second activation function layer.

9. The radar antenna intelligent real-time beamforming method according to claim 1, characterized in that: The input of the neural network null former includes two paths, the first input of the neural network null former includes a third convolution layer, a third batch normalization layer, a third activation function layer, a fourth convolution layer, a fourth batch normalization layer and a fourth activation function layer connected in sequence, the input of the third convolution layer is a weight matrix, the size of its convolution kernel is (1,1), the step length is 1, and it is padded with 0; the activation function of the third activation function layer is relu=max(0,x'); the size of the convolution kernel of the fourth convolution layer is (1,1), the step length is 1, and it is padded with 0; The second input of the neural network null former includes a third fully connected layer, a fourth fully connected layer, a fifth convolutional layer, a fifth batch normalization layer, and a fifth activation function layer; the input of the third fully connected layer is a null angle; the number of neurons in the third fully connected layer is 128, which is used for feature refinement and forward calculation; the number of neurons in the fourth fully connected layer is L, which is used for feature refinement and adjustment of the dimension of neurons, where L=M*N is the number of array elements; The output results of the third activation function layer and the fifth activation function layer are flattened and added together, and then output after the sixth convolutional layer, the sixth batch normalization layer and the sixth activation function layer. The output result is the weight matrix predicted by the neural network null sink generator according to the null sink angle.

10. The radar antenna intelligent real-time beamforming method according to claim 6, characterized in that: In the step S4, it specifically includes: In the on-orbit beamforming stage, the predicted value of the first weight matrix is ​​obtained by inputting the directional pattern into the neural network beamformer, and then the predicted value of the first weight matrix and the list of null angles with n null angles are input into the neural network null angle former, and the predicted value of the second weight matrix is ​​formed in a loop. When the directional pattern satisfies the requirement of having n null angles or the number of iterations reaches the prediction number threshold, the loop is stopped, the beam is formed, and the target is tracked.

Images