A method for array beamforming

By using lightweight prior knowledge in AoA phased array antennas, AoA beamforming is used to use regular dense array excitation coefficients to solve the problem of long AoA beamforming time and achieve fast and real-time beam response .

CN118784042BActive Publication Date: 2024-11-26UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411273175.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-11-26
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

AoA phased array antenna lacks feasible analytical or quasi-analytical beamforming methods, resulting in a long beam switching time and cannot meet the real-time response requirements.

Method used

Using a lightweight neural network guided by prior knowledge, the real and imaginary parts of the excitation coefficient of the rules are used as inputs to output the real and imaginary parts of the AoA excitation coefficient, and AoA beam assignment is achieved through the forward transmission of a single parameter. simplifies the neural network structure and accelerates the beamforming process.

Benefits of technology

Real-time response of AoA beamforming is realized, which significantly reduces the calculation amount and beamforming time and meets real-time requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118784042B_ABST
    Figure CN118784042B_ABST
Patent Text Reader

Abstract

The present invention provides an array beamforming method. This method is based on a lightweight neural network guided by prior knowledge, and the method includes: step one: establishing an AoA beamforming sample data set; step two: building a neural network model based on prior knowledge, using the excitation coefficient of a regular dense array as prior knowledge; step three: using the sample data set to train the neural network model; step four: using the trained neural network model to perform AoA beamforming. The scheme of the present invention uses the excitation coefficient of a regular dense array as prior knowledge to reduce the cognitive difficulty of the input-output mapping relationship of the neural network, so that the neural network has a lightweight structure, thereby significantly reducing the amount of calculation for a forward parameter transfer. On the other hand, using a trained neural network to perform AoA beamforming only requires a forward parameter transfer, so the beamforming speed is fast and can meet real-time response requirements.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of phased array antennas, and in particular relates to phased array antenna beamforming, and specifically relates to an array of arrays (AoA) beamforming method. Background Art

[0002] Phased array antennas have been widely used in various wireless electronic systems due to their high gain and flexible beam scanning capabilities. Compared with traditional unit-level phased array antennas, AoA, a phased array antenna that uses a subarray array architecture, can significantly reduce the cost of phased array antennas and solve the problem of insufficient space for transceiver components in high-frequency phased array antennas. However, unlike unit-level phased array antennas, which can achieve rapid synthesis of scanning pencil beams through analytical methods such as Chebyshev and Taylor, AoA does not have a feasible analytical synthesis method. The beamforming of AoA relies on multiple rounds of iterative optimization of the optimization algorithm, resulting in a long beam switching time and unable to meet the real-time response requirements. Therefore, it is very necessary to propose an analytical or quasi-analytical beamforming method for AoA to meet the real-time beam response requirements of AoA. Summary of the invention

[0003] In order to meet the demand of AoA for analytical or quasi-analytical beamforming methods, the present invention adopts a lightweight neural network guided by prior knowledge, and proposes an AoA beamforming method based on a neural network guided by prior knowledge. The prior knowledge used is the real and imaginary parts of the regular dense array excitation coefficients, which are obtained by the Chebyshev amplitude and analytical scanning phase, and as the input of the neural network, the input of the neural network is no longer a beamforming index, and the output of the neural network is the real and imaginary parts of the AoA excitation coefficients. The AoA beamforming using the trained neural network is only a forward transfer of parameters once, so it can be regarded as a quasi-analytical beamforming method.

[0004] The present invention provides an array beamforming method, which is used for array AoA scanning pencil beamforming. The specific steps of the method are as follows:

[0005] Step 1: Create an AoA beamforming sample dataset

[0006] Considering an array composed of M antenna units, AoA effectively reduces the number of transceiver modules by adopting a subarray array architecture. The AoA beamforming sample data set consists of N pairs of input and output pairs of the neural network. The input of the neural network is the real and imaginary parts of the excitation coefficients of the M-element regular dense array, that is, [Re( oh ele 1 ),Im( oh ele 1 ),Re( oh ele 2 ),Im( oh ele 2 ),...,Re( oh ele M ),Im( oh ele M )] , where Re() and Im() represent the real part and imaginary part respectively. represents the excitation coefficient of the ith antenna unit, and the output of the neural network is the real and imaginary parts of the excitation coefficient of the M-element AoA, that is, [Re( oh sub 1 ),Im( oh sub 1 ),Re( oh sub 2 ),Im( oh sub 2 ),...,Re( oh sub L ),Im( oh sub L )] , where L represents the number of subarrays used by the M-element AoA, represents the excitation coefficient of the j-th subarray. N pencil beam scanning angles are randomly specified within the AoA beam scanning range. At the same time, the sidelobe level threshold is given. The regular dense array excitation coefficient is obtained by the Chebyshev amplitude and analytical scanning phase. The AoA excitation coefficient is obtained by the convex optimization algorithm model that maximizes the mainlobe gain. The two sets of excitation coefficients in the same sample pair correspond to the same beam scanning angle.

[0007] Step 2: Build a neural network model based on prior knowledge

[0008] A lightweight neural network guided by prior knowledge is used. The input of the neural network is not the beamforming index, but the prior knowledge of beamforming. The output of the neural network is the real part and the imaginary part of the AoA excitation coefficient. The prior knowledge used is the excitation coefficient of a regular dense array, which is obtained by Chebyshev amplitude and analytical scanning phase.

[0009] A neural network model is built with certain neural network layers and neurons, and an activation function is specified for the nonlinear network layer. Since the input of the neural network is the real and imaginary parts of the regular dense array excitation coefficients, the number of neurons in the neural network input layer is 2M. Since the output of the neural network is the real and imaginary parts of the AoA excitation coefficients, when the sparsity rate of the AoA transceiver module is P, the number of neurons in the neural network output layer is 2M×P.

[0010] Step 3: Train the neural network model using the sample data set

[0011] The AoA beamforming sample dataset established in step 1 is divided into a training dataset and a test dataset in a certain ratio, and the training dataset is divided into several batches; the neural network model is trained based on the above training dataset, the loss function is designed, and the optimizer, learning rate, and number of training epochs are specified.

[0012] Step 4: Use the trained neural network model to perform AoA beamforming

[0013] After the training of the neural network model is completed, the trained neural network model is used for real-time AoA beamforming; given an arbitrary pencil beam scanning angle and sidelobe level threshold, the real and imaginary parts of the corresponding regular dense array excitation coefficients are sent to the neural network, and the real and imaginary parts of the AoA excitation coefficients are obtained from the output of the neural network, thus completing the array AoA beamforming.

[0014] The beneficial technical effects of the present invention are:

[0015] First, the scheme of the present invention adopts the excitation coefficients of a regular dense array as prior knowledge to reduce the difficulty of recognizing the input-output mapping relationship of the neural network, so that the neural network has a lightweight structure, thereby significantly reducing the amount of calculation for a forward parameter transfer.

[0016] Second, using the trained neural network to perform AoA beamforming only requires one forward propagation of parameters, and the beamforming speed is fast, which can meet real-time response requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 is a flow chart of an array beamforming method provided by an embodiment of the present invention;

[0019] Figure 2 is a schematic diagram of a neural network provided by an embodiment of the present invention;

[0020] Figure 3 is a schematic diagram of the AoA subarray layout and the relative feeding phase within the subarray provided by an embodiment of the present invention;

[0021] Figure 4 : is a comparison diagram of the AoA excitation coefficients designed by the array beamforming method provided by an embodiment of the present invention and the convex optimization algorithm;

[0022] Figure 5 It is a comparison of the AoA power pattern achieved by the array beamforming method provided by the embodiment of the present invention and the convex optimization algorithm. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, the present invention provides an AoA beamforming method, the core of which is a neural network guided by prior knowledge, such as Figure 2 As shown, the steps of the method of the present invention are specifically as follows:

[0025] Step 1: Create an AoA beamforming sample data set. Consider a regular dense array consisting of 256 antenna elements. AoA divides the array into 64 sub-arrays, such as Figure 3 As shown in the figure, the sparseness rate of the transceiver module is 25%. Each subarray of AoA is a regular square subarray, including four antenna units with relative feeding phase differences of 0°, 60°, 60° and 120° respectively. Based on this AoA subarray layout, an AoA beamforming sample data set consisting of 3000 pairs of input and output pairs of the neural network is established. Each input sample is the real and imaginary part of the excitation coefficient of a 256-element regular dense array, that is, [Re( oh ele 1 ),Im( oh ele 1 ),Re( oh ele 2 ),Im( oh ele 2 ),...,Re( oh ele 256 ),Im( oh ele 256 )] , where Re() and Im() represent the real part and imaginary part respectively. Represents the excitation coefficient of the i-th antenna unit; each output sample is the real and imaginary part of the excitation coefficient of the 256-element AoA, that is, [Re( oh sub 1 ),Im( oh sub 1 ),Re( oh sub 2 ),Im( oh sub 2 ),...,Re( oh sub 64 ),Im( oh sub 64 )] ,in represents the excitation coefficient of the jth subarray. The two sets of excitation coefficients in the same sample pair correspond to the same scanning angle. The AoA beamforming sample data set considers 3000 randomly specified pencil beam scanning angles within the AoA beam scanning range, and the sidelobe level during beam scanning is required to be less than -10 dB. The regular dense array excitation coefficient is calculated by the Chebyshev amplitude and analytical scanning phase, that is, ,in Represents the vector composed of the excitation coefficients of 256 antenna elements in a regular dense array, is the magnitude vector, is the phase vector. The Chebyshev amplitude vector can be quickly obtained in MATLAB software by using the command reshape(chebwin(16,10)*chebwin(16,10).',1,256). The i-th element of the analytical scanning phase vector is obtained by Calculated, where k is the wave number corresponding to the working wavelength, (x i , y i ) is the position coordinate of the ith antenna unit, (θ 0 , φ 0 ) is the scanning angle. The AoA excitation coefficient is obtained by a convex optimization algorithm model that can maximize the main lobe gain. The mathematical description of the model is as follows:

[0026]

[0027] in, is the AoA excitation coefficient vector, which is the variable to be optimized in the model. is the vector composed of the radiation fields of 256 antenna elements at angles (θ, φ), represents the sidelobe angle domain, is a matrix that describes the relationship between antenna elements and subarrays. If the jth antenna element belongs to the i-th subarray, then ,otherwise .

[0028] Step 2: Build a neural network model based on prior knowledge. The neural network model is built with three fully connected layers. The number of neurons in the three layers is 512, 1024, and 128 respectively. The input layer and output layer are linear layers, and the activation function of the middle hidden layer is set to PReLU (Parametric Rectified Linear Unit). Since the regular and dense array of excitation coefficients are used as prior knowledge, the difficulty of recognizing the input-output mapping relationship of the neural network is reduced, and the structure of the neural network is very simple.

[0029] Step 3: Use the sample data set to train the neural network model. The AoA beamforming sample data set established in step 1 is divided into a training data set consisting of 2000 samples and a test data set consisting of 1000 samples. The training data set is evenly divided into 4 batches, each of which includes 500 groups of samples. The loss function of the neural network training is the mean square error loss function, the optimizer uses Adam, the learning rate is 0.001, and a total of 300 epochs are trained.

[0030] Step 4: Use the trained neural network model for AoA beamforming. After the neural network model is trained, it can be used for real-time AoA beamforming. Given an arbitrary pencil beam scanning angle and a sidelobe level threshold of -10 dB, the real and imaginary parts of the corresponding regular dense array excitation coefficients are sent to the neural network, and the real and imaginary parts of the AoA excitation coefficients can be obtained from the output of the neural network.

[0031] When AoA is to be scanned to (θ 0 = 30°, φ 0 = 0°), Figure 4 A comparison between the real and imaginary parts of the AoA excitation coefficients given by the AoA beamforming method proposed in the present invention and the real and imaginary parts of the AoA excitation coefficients given by the convex optimization algorithm is given. Figure 5 The AoA power patterns achieved by the above two sets of excitation coefficients are given. As can be seen from the figure, the real and imaginary parts of the AoA excitation coefficients designed by the present invention are very close to the real and imaginary parts of the AoA excitation coefficients given by the convex optimization algorithm, and the achieved AoA power patterns are also basically consistent. On the other hand, based on a personal computer configured with 32 GB of memory and a 3.7 GHz processor, the convex optimization algorithm takes 12 seconds to complete this beamforming, while the present invention only takes 0.5 milliseconds to complete this beamforming, demonstrating the advantages of the present invention in real-time beamforming.

[0032] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An array beamforming method, characterized in that: The method is used for array array AoA scanning pencil beamforming, comprising: Step 1: Establish an AoA beamforming sample data set, which consists of an array of M antenna units. AoA effectively reduces the number of transceiver modules by adopting a subarray array architecture. The AoA beamforming sample data set consists of N pairs of input and output pairs of the neural network. The input of the neural network is the real and imaginary parts of the excitation coefficients of the M-element regular dense array, that is, Where Re() and Im() represent the real part and imaginary part respectively, ω elei represents the excitation coefficient of the ith antenna unit, and the output of the neural network is the real and imaginary parts of the excitation coefficient of the M-element AoA, that is, Where L represents the number of subarrays used by the M-element AoA, represents the excitation coefficient of the j-th subarray. N pencil beam scanning angles are randomly specified within the AoA beam scanning range. At the same time, the sidelobe level threshold is given. The regular dense array excitation coefficient is obtained by the Chebyshev amplitude and analytical scanning phase. The calculation method is: ω ele =amp×e j×phase , where ω ele It represents the vector composed of the excitation coefficients of the M antenna units in a regular dense array, amp is the amplitude vector, phase is the phase vector, and the AoA excitation coefficient is obtained by a convex optimization algorithm model that maximizes the main lobe gain. The model is expressed as: Among them, ω sub is the AoA excitation coefficient vector, which is the variable to be optimized in the model. The M antenna elements are at an angle The vector of the radiation field on sidelobes represents the sidelobe angle domain, Γ is a matrix describing the affiliation between antenna units and subarrays. If the jth antenna unit belongs to the ith subarray, then Γ(i,j) = 1, otherwise Γ(i,j) = 0. The two sets of excitation coefficients in the same sample pair correspond to the same beam scanning angle. Step 2: Build a neural network model based on prior knowledge. Use a lightweight neural network guided by prior knowledge. The input of the neural network is not the beamforming index, but the prior knowledge of beamforming. The output of the neural network is the real and imaginary parts of the AoA excitation coefficient. The prior knowledge used is the excitation coefficient of a regular dense array. The excitation coefficient of the regular dense array is obtained by Chebyshev amplitude and analytical scanning phase. Step 3: Use the sample data set to train the neural network model. Divide the AoA beamforming sample data set established in step 1 into a training data set and a test data set according to a certain ratio, and divide the training data set into several batches. Train the built neural network model based on the above training data set, design the loss function and specify the optimizer, learning rate and number of training epochs. Step 4: Use the trained neural network model to perform AoA beamforming. After the neural network model training is completed, use the trained neural network model to perform real-time AoA beamforming. Given an arbitrary pencil beam scanning angle and sidelobe level threshold, the corresponding real and imaginary parts of the regular dense array excitation coefficients are sent to the neural network. The real and imaginary parts of the AoA excitation coefficients are obtained from the output of the neural network, thus completing the array AoA beamforming.

2. The array beamforming method according to claim 1, characterized in that: The step 2 specifically includes: building a neural network model with neural network layers and neurons, and specifying an activation function for the nonlinear network layer; since the input of the neural network is the real part and imaginary part of the regular dense array excitation coefficient, the number of neurons in the neural network input layer is 2M; since the output of the neural network is the real part and imaginary part of the AoA excitation coefficient, when the sparsity rate of the AoA transceiver module is P, the number of neurons in the neural network output layer is 2M×P.

Citation Information

Cited By

  • Method of performing beamforming and an apparatus thereof

    US20260128511A1