A low-complexity real-time beamforming method for array antenna based on feature extraction

Abstract

The application discloses a low-complexity real-time beamforming method for array antenna based on feature extraction. The method comprises the following steps: firstly, for any given array geometry, the profile projection method is used to extract the profile features of the expected power pattern, and the feature excitation distribution of different main lobe shapes and radiation performance indexes is obtained; based on the feature excitation, the accurate excitation distribution corresponding to the expected pattern profile is obtained by using the alternating projection algorithm; the feature excitation and the accurate excitation distribution are sorted out to construct a benchmark data set; the neural network is trained using the benchmark data set to obtain a real-time prediction model of the feature excitation distribution to the accurate excitation distribution; the feature excitation is extracted according to the target expected power pattern profile, and then input to the neural network to predict the accurate array element excitation, so that the real-time beamforming is realized. By extracting the profile features of the expected power pattern, the complex nonlinearity of the neural network is reduced, the excitation solving efficiency is improved, and the method is suitable for real-time beamforming of the array.

Description

Technical Field

[0001] This invention relates to the field of array antennas, and more specifically, to a low-complexity real-time beamforming method for array antennas based on feature extraction. Background Technology

[0002] In 5G communication systems, beamforming technology is typically used to focus signals on specific areas and users to improve communication quality, capacity, and coverage. If a user moves to another location, a rapid switch to a new beam pattern is required to maintain signal transmission efficiency and reliability. Furthermore, in modern electronic warfare, beamforming technology can enhance the anti-jamming capabilities of electronic systems, improving system performance and reliability. Rapidly adjusting the beam pattern helps to promptly counter enemy interference, thereby improving battlefield effectiveness. Therefore, real-time beamforming technology has received close attention in the fields of communications and electronic warfare.

[0003] To date, various fast beamforming methods have been proposed in academia to synthesize the radiation patterns of array antennas, including analytical methods such as Chebyshev synthesis, Taylor synthesis, and Woodward synthesis. While these methods are highly efficient, their applicability is limited, and they cannot achieve fast beamforming for complex, non-uniform arrays. To achieve beamforming of arbitrary array radiation patterns, stochastic optimization algorithms such as genetic algorithms and particle swarm optimization, mathematical programming algorithms such as quadratic programming and semidefinite programming, as well as matrix beamforming and iterative Fourier transform methods have been proposed. Although these methods have greatly improved the versatility of solving beamforming patterns, their efficiency still falls short of real-time requirements.

[0004] Machine learning, a discipline that emerged in the mid-20th century aimed at studying how to simulate human learning activities using computers, possesses the ability to fit complex nonlinear functions and exhibits strong generalization capabilities. This has led to its widespread application in fields such as speech recognition, image processing, recommendation algorithms, and smart healthcare. In the research of array antenna beamforming, the ability of neural networks to fit nonlinear functions can be utilized. By training the neural network model with the correct dataset and employing methods to improve its generalization ability, the trained neural network model can ultimately predict the array excitation required for the desired beam in real time based on design specifications. Combining machine learning with antenna arrays leverages the low computational resource consumption of machine learning, reducing the research cost of complex antenna array design and improving the solution efficiency of array beamforming synthesis.

[0005] Chinese Patent 201710657351.3 discloses a method for pattern synthesis of multi-beam satellite array antennas based on particle swarm optimization (PSO) algorithm. This method employs a heuristic PSO algorithm, simulating the calculation of array element amplitude and phase as a bird flock searching for food in nature. By setting the initial population size, number of particles, and number of iterations, excellent flat-top beams and cosecant square beams are obtained, converging to the optimal solution in only 300 iterations. Although this method improves the convergence speed of PSO algorithm-optimized beamforming, its efficiency still cannot achieve real-time beamforming.

[0006] Chinese Patent 202111382528.6 discloses a deep learning-based method for synthesizing radiation patterns of planar array antennas. This method utilizes a forward analytical approach to calculate a sample set of radiation patterns, using the radiation patterns and feature labels as input data to a neural network, and the excitation distribution of the array units as output. It trains the end-to-end deep convolutional neural network's structural parameters, ultimately obtaining a convolutional neural network with good performance. This method can quickly solve for the excitation distribution of the target radiation pattern. However, the forward analytical approach to solving the dataset has certain limitations, causing the trained deep convolutional neural network to be unable to predict the excitation distribution of complex shaped radiation patterns.

[0007] Chinese Patent 201610817682.4 discloses a method for array antenna pattern synthesis based on a neural network algorithm. This method trains a radial basis function neural network (RBN) using multiple sets of ideal pattern sample data, employing snapshots of the angles of arrival (OA) and the optimal weight values ​​of the array elements as input and output data. After training, the RBN can adaptively generate the pattern for the antenna array system when given different OA angles. While this method can quickly generate array weights based on the OA and produce low-level nulls, it is only suitable for predicting array weights for focused beams and cannot be applied to predicting array weights for complex patterned antennas.

[0008] Chinese Patent 202310202063.4 discloses a real-time beam synthesis method for large-scale conformal arrays based on a generalized regressive neural network. This method utilizes a convex optimization algorithm to obtain the amplitude distribution of the conformal array at specific angles as a sample library to train the generalized regressive neural network. The array element amplitude is predicted by inputting the desired sidelobe level and beam pointing. While this method can predict the focused beam pattern of the conformal array in real time, it can only predict the excitation amplitude and cannot predict the excitation phase, making it difficult to achieve real-time prediction of complex shaped beam patterns.

[0009] To address the technical problems encountered in the aforementioned background technology, this invention proposes a low-complexity real-time beamforming method for array antennas based on feature extraction. Summary of the Invention

[0010] To address the aforementioned problems, the present invention aims to propose a low-complexity real-time beamforming method for array antennas based on feature extraction. This method is applicable to arrays with arbitrary layouts and can predict the excitation distribution of the array in real time based on the target desired power pattern profile.

[0011] To achieve the above-mentioned technical objectives, the present invention includes the following steps:

[0012] Step 1: Collection and organization of benchmark datasets

[0013] Based on any given array geometry, the contour projection method is used to synthesize the characteristic excitation distributions of the desired power pattern contours for different main lobe waveforms and radiation performance indices. Using the characteristic excitation distributions obtained by the contour projection method as the initial solution, the alternating projection algorithm is used to synthesize the precise excitation distributions corresponding to the desired power pattern contours with different main lobe waveforms and radiation performance indices. The characteristic excitations and precise excitation distributions obtained by the above methods are collected and organized to construct a benchmark dataset.

[0014] Step 2: Build and train the machine learning model

[0015] A neural network model is constructed, where the input data consists of the characteristic excitation distributions corresponding to the expected power pattern contours of different main lobe waveforms and radiation performance indices, and the output data consists of the precise excitation distributions corresponding to the expected power pattern contours of different main lobe waveforms and radiation performance indices. The backpropagation algorithm is used to train the network and obtain its structural parameters.

[0016] Step 3: Perform stimulus prediction based on the trained machine learning model.

[0017] By inputting the characteristic excitation distribution corresponding to the target main lobe waveform and the desired power pattern contour into the trained neural network, the accurate excitation distribution of the corresponding array unit can be obtained in real time.

[0018] Compared with existing technologies, the advantages of this invention are: it is applicable to any given array geometry. It proposes using a contour projection method to extract features from the desired power pattern contour, transforming high-dimensional pattern sampling data into low-dimensional feature excitation data, thus reducing the high-dimensional and complex nonlinear characteristics of the training data. The extracted feature excitation is used to predict the precise excitation corresponding to the desired power pattern, thereby further reducing the required network depth, improving prediction accuracy and real-time performance, and achieving low-complexity real-time beamforming. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method of the present invention;

[0021] Figure 2 This is a schematic diagram of the 256-element non-uniform planar array layout in an embodiment of the present invention;

[0022] Figure 3 This refers to the BP neural network model designed in this embodiment of the invention;

[0023] Figure 4 The target desired power pattern in this embodiment of the invention has a flat-top beamforming width of 62° and a maximum sidelobe level of -20.91dB.

[0024] Figure 5 For the embodiments of the present invention Figure 4 The neural network predicts the direction pattern of the target. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments are not intended to limit the present invention. The present invention mainly introduces a low-complexity real-time beamforming method for array antennas based on feature extraction. The method includes the following steps:

[0026] First, given any array, let's take an N-ary planar array as an example. Its pattern function can be expressed as:

[0027]

[0028] in The observation angle in the global coordinate system, w n For the excitation of the nth array element, Let λ represent the wavenumber, and λ represent the operating wavelength. n ,y n ) represents the position of the nth array element. For numerical calculation purposes, ... Discretized Where m = 1, 2, ..., M, and the following matrix is ​​defined:

[0029] AF={AF(u1,v1),AF(u2,v2),…AF(um ,v m ),…,AF(u M ,v M )} T

[0030] w = {w1, w2, ..., w} n ,…,w N} T

[0031] S = {s1, s2, ..., s} m ,…,s M} T

[0032]

[0033] The above array pattern function can then be rewritten as a matrix product:

[0034] AF = Sw

[0035] Next, this patent proposes to use the contour projection method to extract the feature excitation distribution. Assume the desired orientation pattern contour is... Using the least squares method, the characteristic excitation corresponding to the desired radiation pattern can be obtained.

[0036]

[0037]

[0038] in It is the left inverse matrix of S; (·) H This indicates the conjugate transpose, (·) + δ represents the pseudo-inverse of the matrix; I is the identity matrix; δ (δ>0) is the regularization coefficient, which can improve computational stability.

[0039] Next, this patent proposes using the alternating projection method to collect accurate excitation distributions. The alternating projection algorithm transforms the array synthesis problem into a problem of searching for solutions to the intersection of two sets, which are feasible sets. and expected set In addition, feasible sets The expected set contains all radiation patterns generated by a given array. It consists of all radiation patterns that satisfy the desired performance, but it is not required that they can actually be generated by a given array.

[0040] For a given desired power pattern profile, it is divided into main lobe regions Ω. ML and the Ω sidelobe region SL The upper boundary of the main lobe region is The lower boundary is The sidelobe area only has an upper boundary The constraints on the desired radiation pattern profile can be described by the following formula:

[0041]

[0042] make Let be the projection operator from the feasible set to the desired set, and q be the iteration order. Then, in the main lobe region Ω... ML The expected pattern projection of the (q-1)th iteration It can be represented as follows:

[0043]

[0044] In the Ω region of the accessory lobe SL The expected pattern projection of the (q-1)th iteration It can be represented as:

[0045]

[0046] Where ξ is the overpressure factor, typically taking values ​​in the range 0 < ξ < 1. According to the pattern product theorem, the pattern function of the (q-1)th iteration can be expressed as:

[0047] AF( q-1 ) = Sw( q-1 )

[0048] w( can be obtained by the least squares method) q-1 ):

[0049] w( q-1 ) = RAF( q-1 )

[0050] R = (S H S+δI) + S H

[0051] make The stimulus for the qth iteration after correction is then:

[0052] w( q )=RAF'

[0053] make Let the projection operator be the projection from the desired set to the feasible set. Then the feasible array pattern function for the q-th iteration can be expressed as:

[0054]

[0055] Combining the above formulas, the iterative formula for the alternating projection algorithm can be derived:

[0056]

[0057] For a given array, the iterative process begins with an initial array pattern AF. (0) It is generated by a set of preset excitations w0, and the feature excitations obtained by contour projection method can be selected. As the initial excitation, the precise excitation of the desired power pattern profile is finally obtained after iterative convergence through alternating projections.

[0058] Next, we collect K desired radiated power pattern profiles given the main lobe shape and radiation performance indicators. in This represents the power pattern profile of the k-th desired shaped beam. For the above K desired radiated power pattern profiles, the profile projection method proposed in this patent is used to obtain the characteristic excitation distribution dataset. The characteristic excitation matrix can then be expressed as:

[0059]

[0060] in, This represents the characteristic excitation distribution corresponding to the k-th desired power pattern profile. Since the excitation consists of amplitude and phase, therefore... The number of variables is 2N, and its structure can be represented as:

[0061]

[0062] For the above K desired radiated power pattern profiles, using the alternating projection algorithm to obtain the accurate excitation distribution dataset, the accurate excitation matrix can be expressed as:

[0063]

[0064] in, The characteristic excitation distribution corresponding to the k-th desired power pattern profile can be represented as follows:

[0065]

[0066] Based on the collected data samples, we construct a benchmark dataset. Using this benchmark dataset, we will train a neural network-based machine learning model. This design uses a backpropagation (BP) neural network model, but this invention is not limited to any particular type of neural network; radial basis function (RBF) neural networks, generalized regression neural networks, etc., can also be used. The training process mainly consists of the following steps: dataset partitioning, neural network parameter design, network training, and parameter tuning.

[0067] Dataset partitioning: Based on the collected data samples, we will randomly divide the benchmark dataset into training, validation, and test sets according to a certain ratio. The training set is used to fit the network model; the validation set is used to tune the model's hyperparameters; and the test set is used to evaluate the generalization ability of the trained model.

[0068] Neural network parameter design: The BP neural network mainly consists of three layers, including one input layer, one output layer, and one hidden layer. The number of neurons in the input and output layers is determined by the number of array elements, generally set to 2N. The number of neurons in the hidden layer is set according to the actual model fitting. The activation function of the hidden layer is the hyperbolic tangent activation function.

[0069]

[0070] This function converges faster than the Sigmoid function, and its approximate saturation region covers a wider range. The activation function for the output layer is the purelin linear function:

[0071] purelin(x) = x

[0072] This function is often used for function approximation problems in neural networks.

[0073] Network Training and Parameter Tuning: The training of a BP neural network includes two processes: forward propagation and backpropagation. This patent selects the LM (Levenberg-Marquard) algorithm as the backpropagation algorithm. This algorithm significantly reduces the probability of the objective function getting trapped in local minima and requires fewer training iterations, effectively saving training time. In the training of the BP neural network, the weights and biases of the neural network are first randomly initialized. The output of the neural network is calculated through forward propagation using sample data from the training set. The mean squared error (MSE) is used as the loss function to measure the difference between the predicted and accurate activations of the neural network. The loss function LS can be expressed as:

[0074]

[0075] In the loss function LS, W Train To predict incentives, For precise excitation, the Learning Model (LM) algorithm is used to backpropagate errors and update and adjust weights and biases based on the calculated loss function. A validation set is used to monitor model performance, and multiple iterations are performed until the Learning Model (LS) converges to a level that meets the performance metrics. During neural network training, the maximum number of epochs is no less than 2000, and the learning rate is controlled between 0.001 and 0.005 to ensure system stability. After training, the performance of the obtained neural network model is tested using a validation set. Once the BP neural network model reaches the performance requirements, the feature excitation distribution is extracted using the contour projection method based on the desired power pattern contour, and this distribution is input into the neural network model, allowing for real-time prediction of the array excitation required to achieve the desired power pattern.

[0076] The specific implementation of the low-complexity real-time beamforming technology for array antennas based on feature extraction proposed in this invention can be further given through the following simulation examples and results:

[0077] In this simulation example, consider a 256-element non-uniform spiral planar array consisting of ideal point sources with a minimum element spacing of 0.5λ, operating at 5 GHz. Its basic layout is as follows: Figure 2 As shown. Next, a dataset was generated based on the expected power pattern profile and radiation characteristics of the flat-top shaped beam. The dataset covers a pattern-shaped region with a width ranging from [40°, 80°] and a sidelobe level ranging from [-30dB, -20dB]. The contour projection method was used to obtain feature excitations, and the alternating projection method was used to obtain precise excitations, resulting in 16261 sets of sample data. The dataset was randomly divided in a ratio of 0.81:0.09:0.1, with 13171 sets (81%) as the training set, 1464 sets (9%) as the validation set, and 1626 sets (10%) as the test set. A BP neural network was trained using an Intel i7-11700k@3.6GHz processor and 64GB RAM. After multiple experiments, the number of neurons in the hidden layer was finally determined to be 512, with a learning rate of 0.002. The structure of the neural network is shown below. Figure 3 As shown. The LS value of the BP neural network after training and validation is 5.82 × 10⁻⁶. -7 The LS value of the BP neural network on the test set is 4.68 × 10⁻⁶. -7 This shows that the model fits successfully.

[0078] Based on the pre-trained BP neural network model, providing the feature excitation corresponding to the pattern contour shaping region width range BW∈[40°, 80°] and the sidelobe level range SLL∈[-30dB, -20dB] allows for the real-time acquisition of accurate excitation for the desired power pattern. A set of test samples was selected, with a desired pattern contour shaping width of 62° and a maximum sidelobe level of -20.91dB. Figure 4 This is the desired direction map for this objective. Figure 5 This is the predicted radiation pattern obtained based on the BP neural network model. Comparison shows that the main lobe shape of the predicted radiation pattern is consistent with the expected radiation pattern, and the sample MSE value is 3.62 × 10⁻⁶. -6 The sidelobe level increased by 1.2 dB, and the prediction time was 0.017 s, indicating good prediction performance. Table 1 shows multiple test results, revealing that the main lobe alignment is good and the MSE of the test samples is less than 3 × 10⁻⁶. -5 The maximum sidelobe level error does not exceed 3dB, and the prediction time does not exceed 0.02s. These results verify the effectiveness of this technique. The trained BP neural network has low complexity and extremely high excitation prediction efficiency, making it suitable for real-time beamforming of array antennas.

[0079] Table 1. Prediction results of the BP neural network model

[0080]

Claims

1. A low-complexity real-time beamforming method for array antennas based on feature extraction, characterized in that... Includes the following steps: 1) Based on any given array geometry, the feature excitation distributions of the desired power pattern contours with different main lobe waveforms and radiation performance indices are synthesized using the contour projection method; the feature excitation distributions obtained by the contour projection method are used as the initial solution, and the precise excitation distributions corresponding to the desired power pattern contours with different main lobe waveforms and radiation performance indices are synthesized using the alternating projection algorithm; the feature excitations and precise excitation distributions obtained by the above methods are collected and organized to construct a benchmark dataset. 2) Construct a neural network model, where the input data is the feature excitation distribution corresponding to the expected power pattern contours of different main lobe waveforms and radiation performance indicators, and the output data is the precise excitation distribution corresponding to the expected power pattern contours of different main lobe waveforms and radiation performance indicators; train the network to obtain the structural parameters of the neural network; 3) Extract the characteristic excitation distribution of the target main lobe waveform and the desired power pattern contour, and input it into the trained neural network to obtain the accurate excitation distribution of the corresponding array unit in real time.

2. The low-complexity real-time beamforming technique for array antennas based on feature extraction according to claim 1, characterized in that... This method can predict and obtain the array excitation distribution in real time based on the target desired power pattern profile.

3. The low-complexity real-time beamforming technique for array antennas based on feature extraction according to claim 1, characterized in that... A contour projection method is proposed to extract features from the contour of the desired power pattern, reducing the complex nonlinear features of the training data. The formula is expressed as follows: in For the characteristic excitation distribution, Let S be the desired power pattern profile, and S be the phase factor of the array. H This indicates the conjugate transpose, (·) + δ represents the pseudo-inverse of the matrix; I is the identity matrix; δ (δ > 0) is the regularization coefficient, which can improve computational stability.

4. The low-complexity real-time beamforming technique for array antennas based on feature extraction according to claim 1, characterized in that... The alternating projection algorithm is used to obtain the accurate excitation distribution of the desired power pattern, thereby improving the efficiency of data acquisition. The iterative formula of the alternating projection method can be expressed as follows: OF (q) =P F P D OF (q-1) Where AF represents the pattern function, q represents the number of iterations, and P... F P is the projection operator from the expected set to the feasible set. D The projection operator from the feasible set to the desired set.

Images