Staged beamforming method and system based on fully connected network

Abstract

This invention discloses a staged beamforming method and system based on a fully connected network, mainly addressing the problems of high computational complexity, long computation time, and performance loss in heuristic beamforming. The implementation scheme includes: pre-learning the fully connected network to form an amplitude weight predictor; the amplitude weight predictor outputs an amplitude weight vector based on the angle difference between the interference direction and the beam pointing angle; calculating the corresponding phase weight vector based on the beam pointing angle; taking the Hadamard product of the amplitude weight vector and the phase weight vector to obtain the array element complex weight vector; multiplying and superimposing the received or transmitted signals of each antenna with the corresponding array element complex weight vector to complete beamforming. This invention reduces online computational complexity by fully utilizing the fitting ability of neural networks and the staged beamforming strategy, simplifies the beamforming steps and network dimensions, and improves beamforming efficiency and performance. It can be used in signal enhancement and anti-interference equipment for communication, radar, and satellite applications.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and specifically relates to a staged beamforming method and system, which can be used in fields such as communication, radar, and satellite. Background Technology

[0002] When two parties are transmitting and receiving information, or when radar receives target echoes, the receiving end receives a mixed signal containing the target signal and other spatial interference signals. While ensuring the quality of the received target signal, it is necessary to suppress other interference signals to achieve a higher standard of target signal reception. The presence of interference signals affects the receiver's reception and detection of target signals. Adaptive beamforming technology is a technique for suppressing interference from different directions and forms. The signal received by the array antenna can be enhanced through adaptive anti-interference technology to achieve higher gain in the target direction and create nulls in the interference and noise directions, suppressing interference and noise. This results in increased gain in the main lobe direction and decreased gain in the interference and noise direction of the array antenna.

[0003] Existing optimization methods for beamforming technology mainly employ specific iterative and convex optimization algorithms. These iterative algorithms suffer from high computational complexity and long computation time, thus beamforming technology cannot meet the high real-time requirements of communication processes in certain scenarios. To address this challenge, some researchers have proposed simple heuristic beamforming schemes. These heuristic beamforming schemes are calculated directly based on channel state information, requiring no iteration and thus exhibiting lower computational latency. However, due to factors such as errors in angle of arrival estimation, amplitude and phase errors in element channels, and element position errors during beamforming, the MVDR algorithm cannot obtain accurate steering vectors and covariance matrices. Subsequently, a series of robust beamforming algorithms were proposed, mainly including diagonal loading algorithms, worst-case performance optimization algorithms based on steering vector uncertainty set constraints, projected subspace algorithms, and interference plus noise covariance matrix reconstruction algorithms. However, these algorithms reduce latency at the expense of performance; this trade-off between latency and performance may limit the potential of beamforming technology and its practical application. Due to recent advancements in neural network technology, real-time prediction of optimal beamforming weights has become particularly important, taking into account both performance and computational latency.

[0004] Neural network technology allows for offline training of neural networks using large amounts of data, retaining high-performing network parameters for online optimization. This method shifts the complex computations of previous online optimization to iterative offline training, utilizing the trained neural network parameters to obtain the optimal beamforming solution, significantly reducing computational complexity and latency.

[0005] Patent application CN202310899181.5 discloses a robust adaptive beamforming method based on a deep unfolded network. This method uses a neural network to learn and predict interference signals and errors, then reconstructs the corresponding covariance matrix, and finally calculates the required weight vector using other methods. However, this method does not fully utilize the learning ability and forward propagation of the neural network; it only uses it to estimate errors, thereby increasing the robustness of beamforming. Furthermore, some methods in other literature use neural networks to predict the imaginary and real parts of the weight vector separately, but this increases the output dimension of the neural network, slowing down both the learning and prediction speeds. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the existing technology by proposing a staged beamforming method and system based on a fully connected network. This method fully utilizes offline resources, improves beamforming efficiency, and leverages the fitting ability of neural networks to simplify network output, reduce computational complexity and real-time running time, thereby enhancing beamforming efficiency.

[0007] The technical solutions to achieve the above objectives include:

[0008] Technical Solution 1: A staged beamforming method based on a fully connected network, characterized by comprising the following:

[0009] The neural network is pre-learned to form an amplitude weight predictor Net;

[0010] The amplitude weight predictor outputs an amplitude weight vector w based on the difference in the interference angle. norm ;

[0011] Obtain the beam pointing angle and calculate the corresponding phase weight vector w. phase ;

[0012] The desired weight vector w is obtained by multiplying the amplitude weight vector and the phase weight vector by the Hadamard product.

[0013] The beamforming process is completed by multiplying the signals received or transmitted by each antenna with the corresponding weight vector w and then superimposing them.

[0014] Furthermore, the pre-learning of the neural network to form the magnitude weight predictor Net includes:

[0015] 2a) Construct a linear array containing N array elements with a spacing of d between adjacent array elements, where d is the element spacing, d = λ / 2, λ is the signal wavelength, and N is an integer greater than or equal to 1;

[0016] 2b) Based on the number of array elements N, a real magnitude weight matrix A in the range [0,1] is randomly generated as training data;

[0017] 2c) Use IFFT to calculate the antenna pattern f of each column of the real amplitude weighting matrix A. m This forms the pattern matrix F:

[0018] F = [f1, f2, ..., f m ,…,f M ], where M is the dimension of the training dataset;

[0019] 2d) The set V is formed by searching the local minima of each column of the direction graph matrix F. m , where set V m The minimum value is the null position u. m (u m ≠±1), according to u m Calculate the null angle vector Θ corresponding to the pattern matrix F. null :

[0020] Θ null =[arcsin(u1),arcsin(u2),…,arcsin(u m ...,arcsin(u M )];

[0021] 2e) Using the ReLU function as the activation function of the intermediate layer, the Sigmoid function as the output layer function, and the mean squared error (MSE) function as the loss function, a magnitude weighted predictor Net with a single-dimensional input and an N-dimensional output is constructed.

[0022] 2f) The null angle Θ null Using the columns of the real magnitude weight matrix A as input values, the network magnitude weight predictor Net is fitted.

[0023] Furthermore, the amplitude weight predictor Net outputs an amplitude weight vector w based on the interference angle difference. norm ,include:

[0024] 3a) Based on the received mixed signal X(l), the interference angle θ is calculated using the MUSIC algorithm. J ;

[0025] 3b) Calculate the interference angle θ J The angle θ between the beam pointing angle θ0 and the beam direction angle θ0 is used as the network input for forward propagation, and the output amplitude weight w is used as the output. norm :

[0026] w norm =Net(θ)

[0027] Net is the magnitude weight predictor.

[0028] Furthermore, the phase weight vector w is calculated based on the beam pointing angle. phase The formula is as follows:

[0029]

[0030] Where λ is the signal wavelength, θ0 is the beam pointing angle, [·] T This indicates transpose, d is the element spacing, and j is the imaginary unit.

[0031] The desired weight vector w is obtained by multiplying the amplitude weight vector and the phase weight vector using the Hadamard product, as shown in the following formula:

[0032] w = w phase ⊙w norm , where ⊙ represents the Hadama product.

[0033] Technical Solution 2: A staged beamforming system based on a fully connected network, characterized in that it includes:

[0034] The training data generation module is used to construct the data required for the magnitude weight predictor fitting process.

[0035] The interference angle estimation module is used to estimate the direction of arrival of the interference signal based on the currently received signal and output the included angle.

[0036] The amplitude weight prediction module first fits the data generated by the training data generation module, and then predicts the amplitude weight vector of the array element based on the fitted network parameters and the angle output by the interference angle estimation module.

[0037] The phase weight generation module is used to construct a phase weight vector based on the beam direction;

[0038] The array element weight synthesis module is used to synthesize the amplitude weight vector of the array element predicted by the amplitude weight prediction module and the phase weight vector constructed by the phase weight generation module into the final array element weight vector.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] Firstly, by employing pre-learning of the neural network, this invention can transfer the complex calculations of online optimization to the continuous iteration of offline training, and obtain the optimal beamforming solution using the trained neural network parameters, thereby reducing the complexity of online computation.

[0041] Secondly, this invention directly constructs the angle θ between interference and beam pointing and the amplitude weight vector w. norm The mapping relationship fully leverages the learning capabilities of fully connected networks and simplifies the beamforming process.

[0042] Third, by adopting a strategy of forming amplitude weight vector and phase weight vector separately, this invention simplifies the network output dimension of the amplitude weight predictor and shortens the training time and prediction time, thereby improving beamforming efficiency. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the implementation of the phased beamforming method based on a fully connected network according to the present invention.

[0044] Figure 2 yes Figure 1 A schematic diagram of the forward propagation of the predictor and the synthesis of the final weights.

[0045] Figure 3 This is a block diagram of the phased beamforming system based on a fully connected network according to the present invention. Detailed Implementation

[0046] The invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0047] Example 1: A staged beamforming method based on a fully connected network.

[0048] Reference Figure 1 The implementation scheme of this example includes:

[0049] (a) The neural network is pre-learned to form the magnitude weight predictor Net.

[0050] The neural networks can be divided into three main types: feedforward neural networks, feedback neural networks, and graph neural networks, among which:

[0051] A feedforward neural network is a simple type of neural network, also known as a multilayer perceptron (MLP). Signals are transmitted unidirectionally from the input layer to the output layer without any feedback. Commonly used model structures include convolutional neural networks, backpropagation (BP) neural networks, and backpropagation (RBF) neural networks.

[0052] Feedback neural networks are directed or undirected cyclic graphs with strong associative memory and optimization capabilities. Their output depends not only on the current input and network weights but also on previous inputs. Commonly used model structures include RNNs, Hopfield networks, Boltzmann machines, and LSTMs.

[0053] Graph Neural Networks (GNNs) are a type of neural network based on graphs, which are sets of functions organized in a graph structure within a topological space for relational reasoning. Several GNN models include: graph convolutional networks, graph autoencoders, graph generative networks, graph recurrent networks, and graph attention networks.

[0054] Feedforward neural networks possess strong function fitting capabilities, while feedback neural networks exhibit strong associative memory and can effectively associate contextual information. Graph neural networks, on the other hand, are better suited for modeling and analyzing complex and structured data. The angle θ between interference and beam pointing, and the amplitude weight vector w... norm The mapping relationship is similar to a functional relationship, therefore this invention uses a feedforward neural network. Furthermore, since the network input is only one-dimensional data and the network output is multi-dimensional data, convolutional operations are not required for feature extraction from the input data. Therefore, this invention preferably uses a BP neural network, i.e., a fully connected network, but other types of network models can also be used.

[0055] Since most neural networks need to learn feature information from data through training, this invention requires pre-learning of the preferred fully connected network to construct the angle θ between interference and beam pointing and the amplitude weight vector w. norm The mapping relationship is implemented through the following steps:

[0056] Step 1: Build a fully connected network.

[0057] The ReLU function is used as the activation function of the intermediate layer, the Sigmoid function as the output layer function, and the mean squared error (MSE) function as the loss function. The network input is a single-valued angle θ, and the output is the array element amplitude weight w. norm That is, the input layer of the amplitude weight predictor has 1 neuron, the output layer has the same number of neurons as the number of array elements N, and the hidden layer dimension and the corresponding number of neurons are adjusted as needed. Let the total number of layers be L, forming a fully connected network, such as... Figure 2 As shown.

[0058] The functions mentioned above are represented as follows:

[0059] The ReLU function is expressed as: The ReLU function can improve the computational efficiency and gradient descent of neural networks, and reduce the problems of overfitting and gradient vanishing. Due to the linear and non-saturating nature of ReLU, it can converge quickly in SGD. In addition, it has low computational complexity and does not require exponential operations, which is why this function is used as the activation function of the intermediate layer in this invention.

[0060] The activation function includes the Sigmoid function, Tanh function, ReLU function, Softmax function, etc. This example uses, but is not limited to, the Sigmoid function, which is represented as follows: The Sigmoid function, used as the final function of the network output layer, can compress the network output to [0,1], thus satisfying the requirements for the magnitude of the weights. In addition, its gradient is smooth, which is convenient for differentiation and also prevents abrupt gradient changes during model training.

[0061] The Tanh function is represented as: In fact, the Tanh function is a variation of the Sigmoid function: unlike Sigmoid, Tanh is "zero-centered," so in some applications, Tanh is better than Sigmoid. However, in the case of saturated neurons, Tanh still does not solve the gradient vanishing problem, and this invention requires the output to be located in [0,1].

[0062] The Softmax function is represented as: The Softmax function is often used as an activation function in the output layer of a neural network. It maps the values ​​of the output layer to the 0-1 interval through the activation function, constructing a probability distribution of the neuron output. In multi-class classification problems, the larger the value mapped by the Softmax activation function, the greater the probability of the true class.

[0063] The MAE function is represented as: Where x represents the independent variable of the function, e represents the natural constant, and Net(x) i y represents the predicted value output by the predictor. i The target value is the element in each column of the real magnitude weight matrix A. MAE is robust to outliers because it calculates them with the same weights regardless of the error size (absolute error does not amplify differences). It is typically used in scenarios where outliers may represent important information or corrupt data.

[0064] The MSE function is represented as: This function is the most commonly used loss function in regression problems; it measures the difference between the model's predictions and the actual results. This example uses, but is not limited to, the MAE (Maximum Absolute Error) function as the loss function for fully connected networks. Compared to MAE, MSE is more suitable for scenarios requiring precise prediction.

[0065] Step 2: Construct data for network training to obtain the network magnitude weight predictor Net:

[0066] 2.1) Construct a linear array containing N array elements with a spacing of d between adjacent array elements, where d is the element spacing, d = λ / 2, λ is the signal wavelength, and N is an integer greater than or equal to 1;

[0067] 2.2) Based on the number of array elements N, a real magnitude weight matrix A in the range [0,1] is randomly generated as training data, and its representation is as follows:

[0068]

[0069] Among them, a n,mLet A be the element in the nth row and mth column, where n = 1, 2, 3, ..., N, m = 1, 2, 3, ..., M, N represents the number of array elements, M represents the size of the dataset, and elements in A satisfy: a n,m ∈[0,1];

[0070] 2.3) Use IFFT to calculate the antenna pattern f of each column in the real amplitude weighting matrix A. m :

[0071]

[0072] Among them, f m (θ) is the antenna pattern vector with angle θ calculated from the weight coefficients in the m-th column of A, g n (θ) represents the antenna angle θ gain of each element. Here, it is assumed that each element is an omnidirectional antenna, i.e., g n (θ) = 1; a n,m Let θ0 be the real amplitude weight of each array element, and θ0 be the angle between the array beam pointing direction and the normal.

[0073] 2.4) By f m The constitutive pattern matrix F:

[0074] F = [f1, f2, ..., f m ,…,f M ]

[0075] 2.5) The local minimum values ​​of each column of the search direction matrix F constitute the set V. m , where set V m The minimum value is the null position u. m , whose u m ≠±1; based on u m Calculate the null angle vector Θ corresponding to the pattern matrix F. null :

[0076] Θ null =[arcsin(u1),arcsin(u2),…,arcsin(u m ...,arcsin(u M )];

[0077] 2.6) The null angle Θ null As input values, each column of the real magnitude weight matrix A is used as the target output value of the network;

[0078] 2.7) By adjusting the network weights and biases of the magnitude weight predictor Net using backpropagation and the Adam algorithm, the magnitude weight predictor Net is fitted:

[0079] 2.7.1) Perform backpropagation on the loss function MSE to calculate the current gradient value g.t Then by g t Calculate the first moment estimate m of the current gradient. t :

[0080] m t =β1m t-1 +(1-β1)g t

[0081] Where β1 is an adjustable parameter for the first-order moment estimation;

[0082] 2.7.2) Based on the current gradient value g t Calculate the second moment estimate v of the current gradient. t :

[0083] v t =β2v t-1 +(1-β2)g t 2

[0084] Where β2 is an adjustable parameter for second-order moment estimation;

[0085] 2.7.3) Estimate m based on the first moment of the current gradient. t and second-order moment estimate v t The parameters of the fully connected network are updated to obtain the weight parameters W of the updated network. i And paranoia parameter B i :

[0086]

[0087] Among them, W′ i and B′ i The parameters before the update are i = 1, 2, 3, ..., L, where α is the learning rate and ε is the value of v. t The compensation parameter is used to prevent v t Approaching zero;

[0088] 2.7.4) Based on the updated amplitude weight predictor parameters W i and B i We obtain x1, x2, ..., x i ,…,x L-1 Data transferred between layers:

[0089] x1 = ReLU(W1θ + B1)

[0090] x2 = ReLU(W2x1 + B2)

[0091] …

[0092] x i =RelU(W i xi-1 +B i )

[0093] …

[0094] x L-1 =ReLU(W L-1 x L-2 +B L-1 )

[0095] Among them, W1, W2, ..., W i ,…,W L-1 Here are the weight parameters for each layer, B1, B2, ..., B i ,…,B L-1 Here, L represents the bias parameter for each layer, L is the number of network layers, and ReLU is the activation function.

[0096] 2.7.5) Based on the output data x of layer L-1 L-1 and the weight parameters W of the Lth layer L And paranoia parameter B L The following expression is obtained for the magnitude weight predictor Net:

[0097] Net(θ) = Sigmoid(W) L x L-1 +B L )

[0098] Where Sigmoid is the activation function, and Net(θ) is the output of the magnitude weight predictor when the input is θ.

[0099] (II) Forming the amplitude weight vector w in stages norm and phase weight vector w phase The beamforming is completed by the weight vector w of the synthesized array elements.

[0100] Step 3: Obtain the interference angle θ based on the mixed signal X(l) J The amplitude weight predictor Net outputs an amplitude weight vector w based on the interference angle difference θ. norm .

[0101] 3.1) The array antenna receives a mixed signal X(l) consisting of communication signals and interference signals transmitted by the transmitting end:

[0102]

[0103] in, This is the signal received by the nth array element at time l;

[0104] For the direction angle θ of the wave kThe changing direction vector elements, j is the imaginary unit, λ is the signal wavelength, and d n The spacing between array elements;

[0105] s k (l) represents the k-th complex signal s k The value at time l in the snapshot;

[0106] 3.2) The interference angle θ is calculated from the received mixed signal X(l) using the Multi-Signal Decomposition (MUSIC) algorithm. J ;

[0107] 3.3) Calculate the interference angle θ J The angle θ between the beam pointing angle θ0 and the beam direction angle θ0 is used as the network input for forward propagation, and the output amplitude weight w is used as the output. norm :

[0108] w norm =Net(θ)

[0109] Net is the magnitude weight predictor.

[0110] Step 4: Calculate the phase weight vector w phase Use it in conjunction with the magnitude weight vector w norm The composite array element weight vector w is used to finally complete beamforming.

[0111] 4.1) Calculate the corresponding phase weight vector w based on the beam pointing angle θ0. phase :

[0112]

[0113] Where λ is the signal wavelength, θ0 is the beam pointing angle, [·] T This indicates transpose, d is the element spacing, and j is the imaginary unit.

[0114] 4.2) Obtain the required element weight vector w by taking the Hadamard product of the amplitude weight vector and the phase weight vector:

[0115] w = w phase ⊙w norm

[0116] Where ⊙ represents the Hadama product.

[0117] 4.3) The mixed signal X(l) is weighted and superimposed using the element weight vector w to obtain the processed signal Y:

[0118] Y = w T X(l)

[0119] in,[·] T This indicates transpose.

[0120] 4.4) The antenna pattern f is calculated using IFFT from the element weight vector w:

[0121]

[0122] Where f(θ) is the antenna array gain in the direction of angle θ, w n θ represents the complex amplitude weight of each array element, and θ0 is the angle between the array beam pointing direction and the normal.

[0123] This completes the phased beamforming based on a fully connected network.

[0124] Example 2: A staged beamforming system based on a fully connected network.

[0125] Reference Figure 3 This example system includes: training data generation module 1, interference angle estimation module 2, amplitude weight prediction module 3, phase weight generation module 4, and array element weight synthesis module 5. Among them:

[0126] The training data generation module 1 is used to construct the data required for the fitting process of the amplitude weight predictor. The data construction method is to randomly generate the amplitude weight of the array elements and then search for the null angle of the antenna pattern under the amplitude weight, thereby forming the null angle and amplitude weight data. The generated training data is transmitted to the amplitude weight prediction module 3.

[0127] The interference angle estimation module 2 is used to estimate the direction of arrival of the interference signal based on the signal currently received by the array, and output the included angle. The core of this module is the DOA estimation algorithm, which estimates the direction of arrival of the interference signal from the mixed signal received by the array, then calculates the difference between the DOA and the beam pointing angle to output the required included angle, and transmits the included angle to the amplitude weight prediction module 3.

[0128] The amplitude weight prediction module 3 first fits the data generated by the training data generation module 1, and then predicts the amplitude weight vector of the array elements based on the fitted network parameters and the included angle output by the interference angle estimation module 2. The core of this module is a fully connected network, which forms an amplitude weight predictor after learning the data transmitted by the training data generation module 1. The amplitude weight predictor performs forward propagation based on the included angle input by the interference angle estimation module 2 and outputs an amplitude weight vector, which is then transmitted to the array element weight synthesis module 5.

[0129] The phase weight generation module 4 is used to construct a phase weight vector based on the known beam pointing angle and output it to the array element weight synthesis module 5.

[0130] The array element weight synthesis module 5 is used to synthesize the amplitude weight vector of the array element predicted by the amplitude weight prediction module 3 and the phase weight vector constructed by the phase weight generation module 4 into the final array element weight vector, that is, to obtain the Hadamard product of the amplitude weight vector and the phase weight vector.

[0131] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. For example, the activation function and loss function of the fully connected circuit can be of other types besides those used in this example; in the interference angle estimation, other DOA estimation algorithms can be used instead of the MUSIC algorithm used in this example. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

[0132] It should be noted that the step numbers in the specification and patent claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

Claims

1. A staged beamforming method based on a fully connected network, characterized in that, Includes the following steps: (1) Pre-learn the neural network to form the magnitude weight predictor Net; including the following: 2a) Construct a linear array containing N array elements with a spacing of d between adjacent array elements, where d is the element spacing, d = λ / 2, λ is the signal wavelength, and N is an integer greater than or equal to 1; 2b) Based on the number of array elements N, a real magnitude weight matrix A in the range [0,1] is randomly generated as training data; 2c) Use IFFT to calculate the antenna pattern f of each column of the real amplitude weighting matrix A. m This forms the pattern matrix F: F=[f1,f2,…,f m ,…,f M ] Where M is the dimension of the training dataset; 2d) The set V is formed by searching the local minima of each column of the direction graph matrix F. m , where set V m The minimum value is the null position u. m (u m ≠±1), according to u m Calculate the null angle vector Θ corresponding to the pattern matrix F. null : Θ null =[arcsin(u1),arcsin(u2),…,arcsin(u m )…,arcsin(u M )]; 2e) Using the ReLU function as the activation function of the intermediate layer, the Sigmoid function as the output layer function, and the mean squared error (MSE) function as the loss function, a magnitude weighted predictor Net with a single-dimensional input and an N-dimensional output is constructed. 2f) The null angle Θ null Using the columns of the real magnitude weight matrix A as input values, the network magnitude weight predictor Net is fitted. (2) The amplitude weight predictor outputs an amplitude weight vector w based on the interference angle difference. norm Including the following: 3a) Based on the received mixed signal X(l), the interference angle θ is calculated using the MUSIC algorithm. J ; 3b) Calculate the interference angle θ J The angle θ between the beam pointing angle θ0 and the beam direction angle θ0 is used as the network input for forward propagation, and the output amplitude weight w is used as the output. norm : In norm =Net(θ) Wherein, Net is the magnitude weight predictor; (3) Based on the beam pointing angle, the corresponding phase weight vector w is calculated. phase The formula is as follows: Where λ is the signal wavelength, θ0 is the beam pointing angle, [·] T This indicates transpose, where d is the element spacing and j is the imaginary unit; The desired weight vector w is obtained by multiplying the amplitude weight vector and the phase weight vector using the Hadamard product, as shown in the following formula: w=w phase ⊙in norm Where ⊙ represents the Hadama product; (4) Take the Hadamard product of the amplitude weight vector and the phase weight vector to obtain the required weight vector w; (5) Multiply the signals received or transmitted by each antenna with their corresponding weights and then superimpose them to complete beamforming.

2. The method according to claim 1, characterized in that, The real magnitude weight matrix A generated in step 2b) is represented as follows: Among them, a n,m Let A be the element in the nth row and mth column, where n = 1, 2, 3, ..., N, m = 1, 2, 3, ..., M, N represents the number of array elements, M represents the size of the dataset, and elements in A satisfy: a n,m ∈[0,1].

3. The method according to claim 1, characterized in that, In step 2c), the antenna pattern vectors of each column of matrix A are calculated using the following formula: Among them, f m (θ) is the antenna pattern vector calculated from the weighting coefficients in the m-th column of A, g n (θ) represents the antenna gain of each element. Here, it is assumed that each element is an omnidirectional antenna, i.e., g n (θ)=1, a n,m θ is the real amplitude weight of each array element, and θ0 is the angle between the maximum beam pointing of the array and the normal.

4. The method according to claim 1, characterized in that, The ReLU function, Sigmoid function, and mean squared error (MSE) function in step 2e) are represented as follows: Where x represents the independent variable of the function, e represents the natural constant, and Net(x) i y represents the predicted value output by the predictor. i The target value is the element of each column in the real magnitude weight matrix A.

5. The method according to claim 1, characterized in that, Step 2f) involves fitting the network magnitude weight predictor Net, which uses backpropagation and the Adam algorithm to adjust the network weights and biases. This includes: 2f1) Based on the current gradient value g t Calculate the first moment estimate m of the current gradient. t : m t =β1m t-1 +(1-β1)g t Where β1 is an adjustable parameter for the first-order moment estimation; 2f2) Based on the current gradient value g t Calculate the second moment estimate v of the current gradient. t : v t =β2v t-1 +(1-β2)g t 2 Where β2 is an adjustable parameter for second-order moment estimation; 2f3) Estimate m based on the first moment of the current gradient t and second-order moment estimate v t Update the network parameters to obtain the updated network parameters W. i and B i : Among them, W′ i and B′ i The parameters before the update are i = 1, 2, 3, ..., L, where α is the learning rate and ε is the value of v. t The compensation parameter is used to prevent v t Approaching zero; 2f4) Based on the fitted amplitude weight predictor parameters W i and B i We obtain x1, x2, ..., x i ,…,x L-1 Data transferred between layers: x1 = ReLU(W1θ + B1) x2 = ReLU(W2x1 + B2) … x i =ReLU(W i x i-1 +B i ) … x L-1 =ReLU(W L-1 x L-2 +B L-1 ) Among them, W1, W2, ..., W i ,…,W L-1 Here are the weight parameters for each layer, B1, B2, ..., B i ,…,B L-1 Here, L represents the bias parameter for each layer, L is the number of network layers, and ReLU is the activation function. 2f5) Using the output data x of layer L-1 L-1 and the weight parameters W of the Lth layer L And paranoia parameter B L The magnitude weighted predictor Net is represented as follows: Net(θ)=Sigmoid(W L x L-1 +B L ) Where Sigmoid is the activation function, and Net(θ) is the output of the magnitude weight predictor when the input is θ.

6. The method according to claim 1, characterized in that, The mixed signal X(l) received by the array antenna in step 3a) can be expressed as: in, This is the signal received by the nth array element at time l; For the direction angle θ of the wave k The changing direction vector elements, j is the imaginary unit, λ is the signal wavelength, and d n The spacing between array elements; s k (l) represents the k-th complex signal s k The value at time l in the snapshot.

7. A staged beamforming system for implementing the method of claim 1, characterized in that, include: The training data generation module is used to construct the data required for the magnitude weight predictor fitting process. The interference angle estimation module is used to estimate the direction of arrival of the interference signal based on the currently received signal and output the included angle. The amplitude weight prediction module first fits the data generated by the training data generation module, and then predicts the amplitude weight vector of the array element based on the fitted network parameters and the angle output by the interference angle estimation module. The phase weight generation module is used to construct a phase weight vector based on the beam direction; The array element weight synthesis module is used to synthesize the amplitude weight vector of the array element predicted by the amplitude weight prediction module and the phase weight vector constructed by the phase weight generation module into the final array element weight vector.

Images