Antenna array beamforming method and system based on radiation pattern

Abstract

The application provides an antenna array beamforming method and system based on a radiation pattern, comprising a self-encoder network and a beamforming matrix; the self-encoder network comprises a self-encoder module and a vector-precoding module. The self-encoder network directly learns the nonlinear mapping relationship between the beamforming matrix and the target pattern, the network is trained by using data and the pattern features are extracted, so that the model can adapt to different hardware architectures such as full-digital beamforming, analog beamforming and hybrid beamforming, the fast mapping from the normalized pattern to the beamforming matrix is realized, the real-time deployment demand in a dynamic environment is met, the real-time performance, adaptability and prediction accuracy of array comprehensive beamforming are effectively improved, and a new technical path is provided for the wireless communication system to realize efficient and low-latency beamforming in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a method and system for predicting beamforming weights of antenna arrays based on radiation patterns, and in particular, a beamforming scheme for large-scale MIMO planar arrays based on autoencoders. Background Technology

[0002] In wireless communication systems, beamforming technology achieves precise control over signal propagation direction through precoding of multiple-input multiple-output (MIMO) antenna arrays, thereby improving signal coverage quality and system spectral efficiency. With the increasing demand for high-frequency, large-scale antenna arrays in 5G and future communication technologies, existing beamforming methods face significant challenges. Digital beamforming (DBF) can flexibly adjust amplitude and phase, but its hardware complexity and power consumption increase exponentially with the number of antennas, making it difficult to apply to ultra-large-scale arrays. Analog beamforming (ABF) only adjusts phase, has simple hardware implementation and low power consumption, but its radiation pattern synthesis capability is limited, and it only supports single-stream transmission. Hybrid beamforming (HBF) combines the advantages of both, but in practical applications, it requires solving the joint optimization problem of the digital and analog domains, resulting in a complex solution space that is difficult to converge quickly to the global optimum.

[0003] Array synthesis, a key technology in beamforming, aims to generate antenna radiation patterns that meet specific requirements by optimizing the beamforming matrix in the array. Traditional array synthesis methods mainly include analytical methods and optimization algorithms. Analytical methods, such as Chebyshev weighted and Taylor weighted methods, directly design the desired radiation pattern using mathematical formulas. However, their applicability is limited to uniform linear arrays (ULA) and is difficult to extend to uniform rectangular arrays (URA) or irregular arrays. Furthermore, for scenarios with complex target radiation patterns, complex mathematical models are required, resulting in significant computational difficulty. Optimization algorithms, including genetic algorithms (GA), particle swarm optimization (PSO), and differential evolution algorithms (DEA), approximate the global optimum through iterative search and can handle relatively complex target radiation patterns with constraints. However, their computational complexity increases sharply with the number of array elements, making them unsuitable for large-scale MIMO antenna arrays. In addition, traditional methods lack sufficient support for hardware constraints in hybrid beamforming and are only applicable to single hardware constraints (such as DBF and ABF).

[0004] In recent years, deep learning technology has provided new solutions for array synthesis in beamforming. Convolutional neural networks (CNN), recurrent neural networks (RNN), generative neural networks and other networks have been used to solve complex nonlinear problems in array synthesis, but due to the existence of multiple solutions, there is a great challenge in how to effectively train these neural networks. Although such methods avoid the high computational complexity of traditional optimization algorithms, there are still some problems: (1) The model is highly dependent on specific antenna array configurations (such as the number of array elements, spacing and arrangement), and the generalization ability is insufficient, making it unable to adapt to dynamically changing communication scenarios or heterogeneous arrays; (2) End-to-end training requires repeated adjustment of all network parameters, and changing the antenna or usage scenario requires repeated training of the network. In communication scenarios, existing research mostly focuses on directly generating beamforming matrices from channel state information (CSI) or target parameters (such as angle and gain), and rarely uses target radiation patterns to support real-time communication needs.

[0005] The patent document "Three-Dimensional Beamforming Method and Apparatus" (CN111181613A) discloses a method to optimize the overall three-dimensional beamforming of an array antenna by shaping and combining the horizontal and vertical radiation patterns of the target's three-dimensional radiation pattern. However, this method requires first splitting the three-dimensional radiation pattern into horizontal and vertical planes, optimizing them separately, and then superimposing them, resulting in high computational complexity and large latency.

[0006] The patent document "Beamforming Method, Apparatus, Base Station and Storage Medium" (CN115622597A) discloses the construction of a radiation model using actual measurement data (such as broadband radiation pattern) of antenna elements and a fixed mathematical model. Although this can improve the accuracy of beamforming weights, it faces problems such as high computational complexity of large-scale MIMO arrays and difficulty in handling hardware constraints of hybrid beamforming.

[0007] Therefore, by utilizing deep neural network technology based on autoencoder models, a general and computationally efficient beamforming matrix prediction method needs to be designed, which enables the direct prediction of the MIMO antenna array beamforming matrix through a normalized two-dimensional radiation pattern. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide an antenna array beamforming method and system based on radiation pattern.

[0009] A training method for an antenna array beamforming system based on a radiation pattern, according to the present invention, includes:

[0010] Pre-training steps: Randomly generate digital beamforming weights, calculate the normalized radiation pattern, input it into the encoder, extract the feature vector of the radiation pattern matrix, reconstruct the radiation pattern through the decoder, and freeze the encoder parameters;

[0011] Online training steps: The encoder extracts feature vectors from the normalized radiation pattern, and the input vector is mapped to a beamforming matrix by the precoding module as the prediction output.

[0012] Preferably, in the pre-training step, the digital beamforming weights are calculated using the array factor formula of the set MIMO antenna to obtain the corresponding normalized radiation pattern;

[0013] The normalized radiation pattern is converted to dB units with a fixed size of 180×180.

[0014] An unsupervised learning strategy is used to train the encoder.

[0015] In the pre-training step, mean squared error is used as the loss function to measure the difference between the normalized radiation pattern and the reconstructed radiation pattern, optimize the encoder parameters, and freeze them.

[0016] Preferably, the online training step uses a shorter training cycle, combines the frozen encoder parameters with the untrained vector-precoder module for online training, and customizes the calculation functions of the vector-precoder module and the radiation pattern for different MIMO antenna array configurations;

[0017]

[0018] Among them, F ABF F DBF F RF F BB These matrices represent analog beamforming (ABF), all-digital beamforming (DBF), analog beamforming (RF) in hybrid beamforming, and digital beamforming (BB) in hybrid beamforming, respectively.

[0019] N t This indicates the number of elements in a uniform rectangular antenna array.

[0020] j represents the imaginary unit;

[0021] q represents the predicted output;

[0022] Represent real numbers;

[0023] N s This indicates the number of data streams transmitted.

[0024] Preferably, the customized adjustment includes adjusting the output dimension and activation function in the vector-precoding module, and replacing the corresponding array element position matrix and array factor calculation formula.

[0025] The predicted output is the phase of the beamforming matrix element, the real and imaginary parts of the beamforming matrix element, or the phase of the analog beamforming matrix and the real and imaginary parts of the digital beamforming matrix.

[0026] The array factor calculation formula is the AF formula, the AF formula for two-dimensional URA, or the AF formula for one-dimensional summation.

[0027] According to the present invention, an antenna array beamforming system based on radiation pattern is trained using the aforementioned antenna array beamforming system training method, which includes: an autoencoder network and a hierarchical improved mean square error loss function.

[0028] The autoencoder network includes an autoencoder module and a vector-precoder module.

[0029] The autoencoder module includes an encoder and a decoder, with mean squared error as the loss function.

[0030] The autoencoder network is iteratively optimized based on a hierarchically improved mean square error loss function. The autoencoder network outputs a predicted beamforming matrix based on the input normalized radiation pattern.

[0031] Preferably, the encoder is a MobileNetV2 architecture and the decoder is a deconvolution decoder.

[0032] The vector-precoding module consists of a four-layer fully connected multilayer perceptron network.

[0033] The output can be set to 128-dimensional real and imaginary part weights, 200 analog phases and 4 digital real and imaginary part weights, or 16 phases.

[0034] Preferably, the beamforming matrix Embed relevant constraints according to different architectures.

[0035] If it is ABF, then constant modulus constraint:

[0036]

[0037] If it is HBF, then the beamforming matrix F is:

[0038] F = F RF F BB

[0039] If it is a DBF, then F only embeds power constraints;

[0040] Among them, F RF This represents the simulated beamforming matrix, with embedded constant mode constraints;

[0041] F BB This represents the digital beamforming matrix.

[0042] Preferably, the loss function in the hierarchical improved mean square error loss function is... The initial structure is as follows:

[0043]

[0044] The first term represents the constraint on the main lobe region, and the second term represents the constraint on the side lobe region.

[0045] P indicates the target pattern;

[0046] O represents the orientation pattern predicted by the network;

[0047] θ and φ represent the pitch angle and azimuth angle, respectively;

[0048] ML represents the number of pitch and azimuth coordinates of the main lobe region;

[0049] SL represents the number of pitch and azimuth coordinates of the sidelobe region;

[0050] ml and sl both represent the corresponding angle index.

[0051] Calculate MSE separately for different levels, performing the calculation by level:

[0052]

[0053] α represents the weighting coefficient corresponding to the MSE;

[0054] i represents the main lobe region of the i-th level layer.

[0055] According to the present invention, an antenna array beamforming method based on radiation pattern is provided, which uses the aforementioned antenna array beamforming system based on radiation pattern to predict the radiation pattern, including:

[0056] Step S1: Set the target radiation pattern, input the normalized target radiation pattern into the antenna array beamforming system based on the radiation pattern, and obtain the beamforming matrix;

[0057] Step S2: Calculate and generate the actual radiation pattern based on the output beamforming matrix.

[0058] Preferably, in step S2, a radiation pattern is calculated based on the array factor to evaluate and optimize the system's prediction performance.

[0059]

[0060] P(θ,φ)=|AF(θ,φ)|

[0061]

[0062] Where AF represents the array factor;

[0063] k represents the wave number;

[0064] λ represents the wavelength, kλ = 2π;

[0065] N z =N y Indicates the number of rows and columns of the antenna array;

[0066] d represents the element spacing;

[0067] θ and φ represent the pitch angle and azimuth angle, respectively;

[0068] F mn This represents the beamforming weights of the corresponding array elements;

[0069] m and n represent the antenna indices on the z-axis and y-axis, respectively;

[0070] j represents the imaginary unit;

[0071] F represents the beamforming matrix;

[0072] N t This indicates the number of elements in a uniform rectangular antenna array.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] 1. This invention achieves rapid mapping from normalized radiation patterns to beamforming matrices by integrating the data-driven capabilities of deep learning with the prior knowledge of electromagnetic models, while also meeting the real-time deployment requirements in dynamic environments.

[0075] 2. This invention effectively improves the real-time performance, adaptability, and prediction accuracy of array-integrated beamforming through the designed loss function and trained deep neural network, providing a new technical path for wireless communication systems to achieve efficient and low-latency beamforming in complex environments.

[0076] 3. This invention effectively captures the geometric features of the radiation pattern through the design of a hierarchical loss function, and achieves a balance between main lobe shape constraints and side lobe suppression. At the same time, it solves the problem of complex radiation pattern synthesis by leveraging the nonlinear mapping capability of the deep learning model. Attached Figure Description

[0077] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0078] Figure 1 This is a schematic diagram of the beamforming training method based on autoencoder networks.

[0079] Figure 2 This is a schematic diagram of the relevant parameters in the array factor of a large-scale MIMO antenna array.

[0080] Figure 3 This is a schematic diagram illustrating the prediction effect of a square orientation pattern.

[0081] Figure 4 This is a schematic diagram illustrating the prediction effect of the triangle direction pattern.

[0082] Figure 5 This is a schematic diagram illustrating the prediction effect of the non-line-of-sight orientation pattern of a real scene A.

[0083] Figure 6 This is a schematic diagram illustrating the prediction effect of the non-line-of-sight orientation pattern of real scene B.

[0084] Figure 7 A comparative diagram showing the beam pattern calculated from the beamforming matrix obtained by different beamforming methods in a non-line-of-sight pattern.

[0085] Figure 8 A schematic diagram comparing the spectral efficiency calculated from the beamforming matrices obtained by different beamforming methods in a non-line-of-sight pattern. Detailed Implementation

[0086] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0087] The present invention provides an antenna array beamforming system based on radiation pattern, to... Figure 1 For example, a pre-trained autoencoder network architecture is adopted, including a pre-trained encoder-decoder module based on MobileNetV2 and a deconvolutional decoder, and a lightweight fully connected vector-precoder module. The autoencoder network directly learns the nonlinear mapping relationship between the beamforming matrix and the target radiation pattern. The network is trained using data and radiation pattern features are extracted, enabling the model to adapt to different hardware architectures such as all-digital beamforming, analog beamforming, and hybrid beamforming, achieving efficient calculation of beamforming weights. Specifically, this includes:

[0088] Training dataset generation and preprocessing: A training dataset of 50,000 randomly generated digital beamforming (DBF) weight matrices was prepared. The corresponding normalized radiation patterns were calculated using the array factor formula for the MIMO antenna. To ensure a uniform scale for the input data, the amplitude of all radiation patterns was converted to dB units and uniformly set to a fixed size of 180×180 to fully cover the radiation characteristics at various angles in space, preparing for subsequent model training.

[0089] Training the autoencoder network. An unsupervised learning strategy is used to train the autoencoder network, the purpose of which is to enable the network to automatically extract deep spatial features from the radiation pattern, thereby compressing complex two-dimensional radiation information into a low-dimensional feature representation.

[0090] Specifically, an autoencoder network is trained using the preprocessed dataset, enabling the encoder model within the network to learn the image features of the radiation pattern of the defined planar array and extract the complex radiation pattern matrix into corresponding feature vectors. Utilizing an autoencoder effectively improves the feature extraction quality of the encoder, allowing the feature vectors to better represent the key features of the radiation pattern.

[0091] The input encoder utilizes the MobileNetV2 architecture, employing depthwise separable convolutions and inverted residual modules for efficient feature extraction. This significantly reduces the number of model parameters while capturing key features such as the main lobe morphology and side lobe distribution in the radiation map. The decoder uses a deconvolutional network structure to progressively reconstruct the radiation map of its original size from the low-dimensional features extracted by the encoder. The lightweight MobileNetV2 encoder ensures real-time performance, and parameter freezing technology addresses computational latency.

[0092] The entire pre-training process uses the mean squared error (MSE) of the input and output as the loss function to measure the difference between the input orientation map and the reconstructed orientation map, thereby continuously optimizing the network parameters so that the encoder can obtain stable and representative feature extraction capabilities, providing a solid foundation for subsequent stages.

[0093] Online training. Global features of main lobe shape, side lobe distribution, and multi-beam coupling are extracted from the normalized 2D beam pattern in one go, and the beamforming matrix suitable for all-digital beamforming, analog beamforming, and hybrid beamforming architectures is directly predicted through online training.

[0094] Specifically, after the pre-training phase, the obtained encoder model possesses strong feature extraction capabilities. The encoder parameters obtained during the pre-training phase are frozen and combined with an untrained lightweight vector-precoding module for online training. During online training, the input is the target radiation pattern, while the model output is the precoding matrix generated by the vector-precoding module. Online training employs a shorter training cycle, enabling the model to converge quickly and significantly reducing the response time during deployment. Simultaneously, by customizing the vector-precoding module and radiation pattern calculation function for different MIMO antenna array configurations, different vector-precoding modules can be flexibly deployed to predict beamforming matrices under different beamforming architectures, achieving versatility across various antenna array structures.

[0095] For antenna arrays of different sizes and architectures, only the output dimension and activation function in the vector-precoding module need to be adjusted, and the corresponding element position matrix and array factor (AF) calculation formula need to be replaced to quickly adapt to various scenarios. In many preferred examples, modularity allows the same model to be adapted to arrays ranging from 8×8 to 10×10 or even larger with only a few parameter adjustments. For example, in an 8×8 all-digital configuration, the output of the vector-precoding module is set to 8×8×2=128-dimensional real and imaginary part weights, and the AF formula for a two-dimensional uniform rectangular array (URA) is used; in a 10×10 hybrid (HBF) configuration with 2 RF chains, the output is split into 10×10×2=200 analog phases and 2×2=4 digital real and imaginary part weights, and then the AF is calculated; while in a 16-element linear analog (ABF) configuration, the vector-precoding module only outputs 16 phases, and the AF calculation is simplified to a one-dimensional summation. In this way, without modifying the pre-trained encoder, universal beamforming for different antenna arrays can be achieved through modular output layers and customized AF functions.

[0096] The vector-precoding module consists of a four-layer fully connected multilayer perceptron (MLP) network. Its main function is to map the high-dimensional feature vectors output by the encoder into the precoding matrix required for beamforming, i.e., the beamforming matrix. The predicted matrix structure differs depending on the beamforming architecture. If configured with ABF, the network's predicted output is considered as the phase of the precoding matrix elements; if configured with DBF, the network's predicted output is considered as the real and imaginary parts of the precoding matrix elements; if configured with HBF, the network's predicted output is considered as the phase of the analog beamforming matrix and the real and imaginary parts of the digital beamforming matrix.

[0097] Specifically, assuming the network output is q, then:

[0098]

[0099] Among them, F ABF FDBF F RF F BB These represent the analog and digital beamforming matrices in analog beamforming, all-digital beamforming, and hybrid beamforming, respectively. The array factor is calculated from the beamforming matrix, converting it into the actual predicted radiation pattern. N s This indicates the number of data streams transmitted.

[0100] By combining offline pre-training with online training, offline large-scale data pre-training combined with online fine-tuning of inference latency of tens of seconds, without relying on the complete CSI matrix, the computational overhead of complex arrays is significantly reduced through the designed loss function and trained deep neural network. This not only effectively improves the real-time performance and prediction accuracy of array synthesis beamforming and achieves accurate prediction of complex beamforming weights, but also has the characteristics of strong versatility and lightweight deployment. It provides a new technical path for the array synthesis of large-scale MIMO planar arrays with high efficiency and low latency in complex environments for wireless communication systems.

[0101] The present invention provides an antenna array beamforming matrix based on radiation pattern. The beamforming matrix of a given array is obtained according to the normalized radiation pattern, which effectively reduces the channel estimation overhead required for beamforming and the computation and time cost required for large-scale MIMO array synthesis, improves system efficiency, and is suitable for dynamic large-scale antenna array scenarios.

[0102] Using a given normalized radiation pattern, find the most suitable beamforming matrix so that the calculated radiation pattern satisfies the target radiation pattern. Since the radiation pattern is a two-dimensional matrix, it is a regression problem from image to vector.

[0103] like Figure 2 As shown, N is located in the yoz plane. t A uniform rectangular antenna array with n elements, the elements being spaced d along both the z-axis and y-axis, assuming that the radiation of each element is isotropic, can be calculated using its array factor as follows:

[0104]

[0105] P(θ,φ)=|AF(θ,φ)|.

[0106] Where AF represents the array factor, k represents the wavenumber, and is related to the wavelength λ: kλ = 2π, N z =N y This indicates the number of rows and columns of the antenna array. d represents the element spacing, θ and φ are the elevation and azimuth angles respectively, and F... mnLet represent the beamforming weights of the corresponding array elements, m and n represent the antenna indices on the z and y axes, respectively, and j represent the imaginary element. The beamforming matrix F is used to represent this as:

[0107]

[0108] For different beamforming architectures, in addition to all architectures needing to satisfy power constraints, the constraints on the beamforming matrix F are also different. To be compatible with diverse hardware constraints, relevant constraints are embedded in the network architecture. For example, the phase constraints of the analog beamforming matrix in hybrid beamforming are transformed into the network output.

[0109] For ABF, F needs to satisfy the constant modulus constraint:

[0110]

[0111] For HBF, F is obtained by multiplying the two beamforming matrices:

[0112] F = F RF F BB

[0113] Among them, the simulated beamforming matrix F RF The constant modulus constraint must be satisfied. BB This represents the digital beamforming matrix. For DBF, F only needs to satisfy the power constraint.

[0114] To ensure the predicted radiation pattern meets physical constraints and target design requirements, a hierarchical improved mean squared error loss function is constructed. This function strictly constrains amplitude fluctuations in the main lobe region (e.g., limiting them to between -1dB and 0dB), while applying threshold penalties in the side lobe region. This effectively suppresses unnecessary side lobes while maintaining geometric positioning accuracy, ensuring that side lobe levels remain below predetermined values. The loss function is constructed differently for radiation patterns under different conditions. To balance the needs of different scenarios and cover radiation patterns ranging from special shapes to real non-line-of-sight multipath scenarios, P represents the target radiation pattern, and O represents the radiation pattern predicted by the network. For radiation patterns with special shapes, the loss function is designed as follows:

[0115]

[0116] The first term represents the constraint on the main lobe region, ensuring that the fluctuation of the predicted radiation pattern within the main lobe is within 1 dB. The second term represents the constraint on the side lobe region, ensuring that the predicted radiation pattern within the side lobe region does not exceed the target level. ML represents the number of pitch and azimuth coordinates in the main lobe region; SL represents the number of pitch and azimuth coordinates in the side lobe region; both ml and sl represent the corresponding angle indices. For radiation patterns in real-world scenarios, considering that the distinction between the main lobe and side lobe regions is not obvious, multiple levels of calculation are required. The loss function is designed as follows:

[0117]

[0118] Here, α represents the weight coefficient corresponding to the MSE; i represents the main lobe region of the i-th level layer. Unlike radiation patterns with special shapes, the loss function for radiation patterns in real-world scenarios refines the definition of the main lobe region, calculating the MSE separately for different levels. This ensures that the predicted radiation patterns can achieve good fit for beams at different levels. Furthermore, different target radiation pattern datasets are used for online training, including samples with special shapes (such as triangles and rectangles) as well as non-line-of-sight (NLoS) channel radiation pattern samples from real-world scenarios. The entire training process involves fewer iterations, significantly reducing computational overhead and enabling the model to converge efficiently.

[0119] by Figure 3 As a preferred example, the beamforming pattern of a square target is predicted. A uniform rectangular array with an HBF configuration is used, with an element spacing of half a wavelength and a total of 100 elements (10×10). The target beamforming pattern is set with a square main lobe distribution, the main lobe gain remaining flat within the azimuth and elevation angles of -20° to 20°, and sidelobes suppressed to below -25dB. By inputting a normalized square beamforming pattern, the model outputs a predicted hybrid beamforming matrix, which, after array factor calculation, generates an actual beamforming pattern that highly matches the target beamforming pattern. Amplitude fluctuations within the main lobe region are strictly controlled within 1.5dB, the overall sidelobe level is below -25dB, and no significant sidelobe rise is observed in the transition region at the square edge. The total online training time is approximately 30 seconds, far shorter than traditional array synthesis algorithms. Through the design of a hierarchical loss function, the geometric features of the beamforming pattern are effectively captured, achieving a balance between main lobe shape constraints and sidelobe suppression. Furthermore, the nonlinear mapping capability of the deep learning model solves the challenge of complex beamforming pattern synthesis.

[0120] by Figure 4As a preferred example, the radiation pattern of a triangular target is predicted to verify its adaptability to asymmetric target radiation patterns. The main lobe of the target is distributed in an isosceles triangle in the azimuth dimension, with the vertex located at (0°, 20°) and the base covering -30° to 30°, while maintaining a uniform distribution in the pitch dimension. The predicted radiation pattern shows that the main lobe shape is highly consistent with the contour of the target triangle, the vertex gain error is less than 2dB, the edge transition is smooth, and the sidelobe leakage phenomenon commonly seen in traditional optimization algorithms is not observed. In addition, the model has a significant sidelobe suppression effect in the triangle peak region, with a maximum sidelobe level of only -25dB, which is strictly less than the set sidelobe level. This performance improvement is due to the hierarchical constraint mechanism of the physics-driven loss function: the main lobe region adopts strict amplitude limitation, while the sidelobe region adjusts the network weights through threshold penalty, thereby achieving fine-grained control even under complex geometries.

[0121] According to the present invention, an antenna array beamforming method based on radiation pattern is provided, which uses the aforementioned antenna array beamforming system based on radiation pattern to predict the radiation pattern, including:

[0122] Step S1: Set the target radiation pattern, input the normalized target radiation pattern into the antenna array beamforming system based on the radiation pattern, and obtain the beamforming matrix;

[0123] Step S2: Calculate and generate the actual radiation pattern based on the output beamforming matrix.

[0124] In more preferred embodiments, step S2 involves calculating and generating a radiation pattern based on the array factor, evaluating the system's prediction performance, and optimizing it.

[0125]

[0126] P(θ,φ)=|AF(θ,φ)|

[0127]

[0128] Where AF represents the array factor; k represents the wavenumber; λ represents the wavelength, kλ = 2π; N z =N y The numbers represent the number of rows and columns of the antenna array; d represents the spacing between array elements; θ and φ represent the elevation and azimuth angles, respectively; F mn The beamforming weights of the corresponding array elements are represented by m and n, respectively, which represent the antenna indices on the z and y axes; j represents the imaginary element; F represents the beamforming matrix; N t This indicates the number of elements in a uniform rectangular antenna array.

[0129] by Figure 5 , Figure 6 For example, the prediction results for non-line-of-sight orientation patterns in two real-world scenarios A and B are shown respectively. Figure 7 , Figure 8 As shown, the beam pattern and spectral efficiency calculated by other beamforming methods are compared, verifying the robustness of main lobe localization and side lobe suppression in a real non-line-of-sight multipath environment.

[0130] The predictive performance of the model is evaluated and optimized by comparing the beamforming matrix obtained based on the autoencoder with that obtained through CSI in terms of spectral efficiency and radiation pattern.

[0131] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A training method for an antenna array beamforming system based on radiation pattern, characterized in that, include: Pre-training steps: Randomly generate digital beamforming weights, calculate the normalized radiation pattern, input it into the encoder, extract the feature vector of the radiation pattern matrix, reconstruct the radiation pattern through the decoder, and freeze the encoder parameters; Online training steps: The encoder extracts feature vectors from the normalized radiation pattern, and the input vector is mapped to a beamforming matrix by the precoding module as the prediction output; In the pre-training step, the digital beamforming weights are calculated using the array factor formula of the set MIMO antenna to obtain the corresponding normalized radiation pattern. The normalized radiation pattern is converted to dB units with a fixed size of 180×180. The encoder is trained using an unsupervised learning strategy; In the pre-training step, mean squared error is used as the loss function to measure the difference between the normalized radiation pattern and the reconstructed radiation pattern, optimize the encoder parameters, and freeze them.

2. The training method for an antenna array beamforming system based on radiation pattern according to claim 1, characterized in that, In the online training step, the frozen encoder parameters are combined with the untrained vector-precoder module for online training, and the calculation functions of the vector-precoder module and the radiation pattern are customized for different MIMO antenna array configurations. in, , , , These matrices represent analog beamforming (ABF), all-digital beamforming (DBF), analog beamforming (RF) in hybrid beamforming, and digital beamforming (BB) in hybrid beamforming, respectively. This indicates the number of elements in a uniform rectangular antenna array. j represents the imaginary unit; q represents the predicted output; Represent real numbers; This indicates the number of data streams transmitted.

3. The training method for an antenna array beamforming system based on radiation pattern according to claim 2, characterized in that, The customized adjustments include adjusting the output dimension and activation function in the vector-precoding module, and replacing the corresponding array element position matrix and array factor calculation formula; The predicted output is the phase of the beamforming matrix element, the real part and the imaginary part of the beamforming matrix element, or the phase of the analog beamforming matrix and the real part and the imaginary part of the digital beamforming matrix. The array factor calculation formula is the AF formula, the AF formula for two-dimensional URA, or the AF formula for one-dimensional summation.

4. A radiation pattern-based antenna array beamforming system, employing the radiation pattern-based antenna array beamforming system training method according to any one of claims 1-3, characterized in that, include: Autoencoder network and hierarchical improved mean squared error loss function; The autoencoder network includes an autoencoder module and a vector-precoder module; The autoencoder module includes an encoder and a decoder, with mean square error as the loss function. The autoencoder network is iteratively optimized based on a hierarchically improved mean square error loss function. The autoencoder network outputs a predicted beamforming matrix based on the input normalized radiation pattern.

5. The antenna array beamforming system based on radiation pattern according to claim 4, characterized in that, The encoder is a MobileNetV2 architecture, and the decoder is a deconvolutional decoder. The vector-precoding module consists of a four-layer fully connected multilayer perceptron network. The output can be set to 128-dimensional real and imaginary part weights, 200 analog phases and 4 digital real and imaginary part weights, or 16 phases.

6. The antenna array beamforming system based on radiation pattern according to claim 4, characterized in that, The beamforming matrix Embed relevant constraints according to different architectures; If it is ABF, then constant modulus constraint: If it is HBF, then the F-beamforming matrix Simulated beamforming matrix Embedded constant modulus constraints; If it is a DBF, then F only embeds power constraints; in, This represents the digital beamforming matrix.

7. The antenna array beamforming system based on radiation pattern according to claim 4, characterized in that, The loss function in the hierarchical improved mean squared error loss function The initial structure is as follows: The first term represents the constraint on the main lobe region, and the second term represents the constraint on the side lobe region. P indicates the target pattern; O represents the orientation pattern predicted by the network; θ and φ represent the pitch angle and azimuth angle, respectively; ML represents the number of pitch and azimuth coordinates of the main lobe region. SL represents the number of pitch and azimuth coordinates of the sidelobe region; ml and sl both represent the corresponding angular indices; Calculate MSE separately for different levels, performing the calculation by level: This represents the weighting coefficient corresponding to the MSE; i represents the main lobe region of the i-th level layer.

8. A radiation pattern-based antenna array beamforming method, applied to the radiation pattern-based antenna array beamforming system according to any one of claims 4-7, characterized in that, include: Step S1: Set the target radiation pattern, input the normalized target radiation pattern into the antenna array beamforming system based on the radiation pattern, and obtain the beamforming matrix; Step S2: Calculate and generate the actual radiation pattern based on the output beamforming matrix.

9. The antenna array beamforming method based on radiation pattern according to claim 8, characterized in that, In step S2, a radiation pattern is calculated based on the array factor to evaluate and optimize the system's prediction performance. Where AF represents the array factor; k represents the wave number; λ represents the wavelength, kλ = 2π; N z =N y Indicates the number of rows and columns of the antenna array; d represents the element spacing; θ and φ represent the pitch angle and azimuth angle, respectively; F mn This represents the beamforming weights of the corresponding array elements; m and n represent the antenna indices on the z-axis and y-axis, respectively; j represents the imaginary unit; F represents the beamforming matrix; This indicates the number of elements in a uniform rectangular antenna array.

Images