A Hybrid Beamforming Optimization Method Based on Deep Learning

Abstract

The present invention discloses a hybrid beamforming optimization method based on deep learning, belonging to the technical field of digital-analog optimization in communication systems. The technical problem to be solved is how to jointly optimize the digital beamformer and the analog beamformer based on deep learning while reducing complexity and improving performance. The method includes the following steps: formulating the hybrid beamforming optimization problem, which is equivalent to obtaining the beamforming solution by maximizing the spectral efficiency under the conditions of constant modulus constraint and power constraint; collecting different realizations of the channel as channel information; preprocessing the channel samples; constructing a hybrid beamforming network model for optimizing hybrid beamforming based on deep learning; performing offline training on the hybrid beamforming network model based on the training set; and performing online testing on the trained hybrid beamforming network model based on the test set to obtain a feasible beamformer.

Description

Technical Field

[0001] The present invention relates to the technical field of digital - analog optimization in communication systems, and more particularly to a hybrid beamforming optimization method based on deep learning. Background Art

[0002] With the rapid development of wireless communication, the demand for wireless communication is increasing. Existing communication methods have limitations in utilizing system structure information and processing a large amount of data, and new communication theories need to be proposed to meet the requirements of communication systems. Most wireless communication systems use multi - antenna technology to increase the stability of receivers, data throughput, and signal - to - noise ratio. Massive Multiple - Input Multiple - Output (MIMO) is a key technology in communication. MIMO technology uses multiple transmit antennas and receive antennas at both ends of a wireless link to utilize multipath propagation, and can make full use of spatial resources. Higher multiplexing and diversity gains can be obtained through MIMO, thus achieving higher spectral efficiency.

[0003] Beamforming is a key technology for MIMO transmission. Beamforming is to change the amplitude and phase of the transmission signal of each transmit antenna according to some prior information (such as channel information, etc.) to direct the wireless signal to the target. In massive MIMO, although full - digital beamforming can theoretically achieve optimal performance, it will face problems such as high overhead, high implementation complexity, and high power consumption. Pure - analog beamforming cannot provide multi - data - stream multiplexing. Although it is simple to implement, the system rate is too low. The hybrid beamforming architecture, which combines low - dimensional baseband digital beamforming and high - dimensional radio - frequency analog beamforming, provides a good balance between hardware complexity and system performance. It can not only obtain the advantages of traditional MIMO precoding but also provide high beamforming gain. However, due to differences in hardware implementation, the phase shifter also adds a constant - modulus constraint to the signal passing through the analog beamformer. Currently, many traditional algorithms for solving the hybrid beamforming design problem have been proposed and have achieved good performance. However, these traditional algorithms are generally iterative algorithms, which require a large amount of time for iterative operations, have a large amount of calculation, high computational complexity, and complex calculation methods that are difficult to understand, and are not suitable for large - scale learning and promotion.

[0004] Since deep learning is a powerful tool for dealing with complex non-convex optimization problems, the application of deep learning in the field of hybrid beamforming optimization problems has received extensive attention in recent years. Using supervised deep learning methods to solve hybrid beamforming optimization problems is usually a popular solution. However, for non-convex optimization problems, it is difficult to obtain the global optimal solution of the problem. Therefore, the supervised learning method based on the local optimal solution cannot achieve good performance. In addition, the performance of supervised learning depends to a large extent on a large amount of labeled data. However, in wireless communication, labeled data is not easy to obtain, so supervised learning is not feasible in practical applications. In addition, using a fully connected neural network to construct a hybrid beamforming network model has the problem of too many weight parameters, which may increase the computational complexity. Therefore, in terms of design complexity and spectral efficiency performance, hybrid beamforming design is a challenging task.

[0005] How to jointly optimize the digital beamformer and the analog beamformer based on deep learning while reducing complexity and improving performance is a technical problem to be solved. Summary of the Invention

[0006] The technical task of the present invention is to address the above deficiencies and provide a hybrid beamforming optimization method based on deep learning to solve the technical problem of how to jointly optimize the digital beamformer and the analog beamformer based on deep learning while reducing complexity and improving performance.

[0007] In a first aspect, a hybrid beamforming optimization method based on deep learning of the present invention is applied to a downlink single-user MISO hybrid beamforming optimization system. The method includes the following steps:

[0008] Formulate the hybrid beamforming optimization problem, which is equivalent to obtaining a beamforming solution by maximizing the spectral efficiency under the conditions of constant modulus constraint and power constraint;

[0009] Collect different realizations of the channel as channel information, construct a channel sample set with the channel information as channel samples. The channel sample is a complex channel matrix, the real part of the channel sample is the amplitude of the channel, and the imaginary part of the channel sample is the phase of the channel;

[0010] Preprocess the channel samples to obtain preprocessed channel samples in real number form, and divide the preprocessed channel samples into a training set and a test set;

[0011] Construct a hybrid beamforming network model for optimizing hybrid beamforming based on deep learning. The hybrid beamforming network model takes channel information as input, and predicts and outputs the negative value of the spectral efficiency by learning the channel information;

[0012] Offline train the hybrid beamforming network model based on the training set, and save the weight parameters of the hybrid beamforming network model during the offline training process to obtain the trained hybrid beamforming network model;

[0013] Online test the trained hybrid beamforming network model based on the test set to obtain a feasible beamformer.

[0014] Preferably, the system includes a transmitter, a channel, and a receiver;

[0015] The transmitter is a base station configured with a hybrid beamformer, and one radio frequency chain and multiple transmit antennas are configured at the base station;

[0016] The hybrid beamformer includes a low-dimensional digital beamformer and a high-dimensional analog beamformer. The digital beamformer is used to transmit a signal to the one radio frequency chain, and the analog beamformer is implemented by multiple phase shifters and is used to connect the one radio frequency chain to the multiple transmit antennas;

[0017] The receiver is configured with one user, and the user is configured with one receive antenna. The base station configured with multiple transmit antennas cooperates with the user configured with one receive antenna through the channel for signal transmission.

[0018] Preferably, the phase shifter cannot adjust the amplitude of the signal and can only change the phase of the signal, and the analog beamformer has a constant modulus constraint

[0019] The base station has a transmit power constraint ||v A v D || 2 ≤P max where P max represents the maximum transmit power at the base station;

[0020] For the system, after the input signal s is transmitted through the digital beamformer v D , it is frequency-converted to the analog phase shifter through the one radio frequency chain, and then the transmit signal is constructed through the analog beamforming vector v A The formula is: Formula:

[0021] x = v A v D s

[0022] The hybrid beamforming vector in the downlink can be expressed as:

[0023] v = v A v D

[0024] After the transmitted signal \(x\) passes through a channel \(h\), the signal \(y\) received by the user at the receiving end is expressed as:

[0025] \(y = h\) H \(v\) A \(v\) D \(s + n\)

[0026] where represents the complex vector of the downlink channel gain, \(N\) t represents the number of transmit antennas, and the superscript \(H\) represents the conjugate transpose, represents the analog beamformer, \(v\) D represents the digital beamformer, \(s\) represents the input signal at the base station, \(n\) represents the additive white Gaussian noise, which follows a complex Gaussian distribution with a mean of 0 and a variance of \(\sigma\) 2 i.e., \(n\sim CN(0,\sigma\) 2 ), and \(\sigma\) 2 represents the noise power;

[0027] The signal-to-noise ratio received by the user is:

[0028]

[0029] where represents the complex vector of the downlink channel gain, and the superscript \(H\) represents the conjugate transpose, represents the analog beamformer, \(v\) D represents the digital beamformer, and \(\sigma\) 2 represents the noise power;

[0030] Based on the signal-to-noise ratio received by the user, the calculation formula for the downlink spectral efficiency is: [[ID=SO]]

[0031]

[0032] where represents the complex vector of the downlink channel gain, and the superscript \(H\) represents the conjugate transpose, represents the analog beamformer at the base station, \(v\) D represents the digital beamformer at the base station, and \(\sigma\) 2 represents the noise power.

[0033] Preferably, the hybrid beamforming optimization problem is formulated, and the corresponding calculation formula is:

[0034]

[0035] s.t. \(\|v\) A \(v\) D \|\) 2 \(\leq P\)max

[0036]

[0037] Among them, max represents the maximum value, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose, represents the analog beamformer, v D represents the digital beamformer, v = v A v D represents the hybrid beamforming vector of the entire downlink, σ 2 represents the noise power, P max represents the maximum transmit power, |||| 2 represents the square of the norm, [v A i represents the i-th element of v A of, N t represents the number of transmit antennas at the base station side.

[0038] Preferably, different realizations of the channel are collected as channel information, and the channel environment respectively simulates a Rayleigh fading channel where the channel parameters follow a standard complex Gaussian distribution and a geometric millimeter-wave channel with very limited multipath components;

[0039] The channel information is a complex vector of size 1×N t where N t represents the number of transmit antennas at the base station side, and 1 represents that the number of receive antennas at the user side is 1.

[0040] Preferably, the channel samples are preprocessed by the following operations: after splitting the real and imaginary parts of the complex channel samples, they are rearranged element by element in order into a three-dimensional real matrix of size 1×N t ×2, where 1 represents a single antenna at the receiving end, N t represents the number of antennas at the transmitting end, and 2 represents splitting the real and imaginary components of the complex channel samples to form two channels.

[0041] Preferably, the hybrid beamforming network model includes:

[0042] An input layer for receiving the preprocessed channel samples;

[0043] A convolutional block for extracting features from the channel samples to obtain multi-dimensional feature data;

[0044] A flattening layer for converting the multi-dimensional feature data into a one-dimensional vector;

[0045] A fully connected layer configured with N​t A neuron, applied to the one-dimensional vector of the flat layer output, generates phase information related to the analog beamformer;

[0046] The first Lambda layer, which is used to satisfy the constant modulus constraint, and converts the real phase vector θ into a complex analog beamforming vector v through Euler's formula A = e iθ , where i 2 = -1;

[0047] The second Lambda layer, which is used to output the negative opposite of the spectral efficiency as the output value of the hybrid beamforming network model. The formula for calculating the negative opposite of the spectral efficiency is:

[0048]

[0049] where γ, h, v A respectively represent the signal-to-noise ratio, channel vector, and analog beamforming vector related to the channel samples. Among them, the analog beamforming vector is the value output by the first Lambda layer, and the superscript H represents conjugate transpose. N t represents the number of transmit antennas at the base station side.

[0050] Preferably, there are three convolution blocks in total, and each convolution block includes:

[0051] A convolution layer, which is used to perform convolution operations on the input data;

[0052] A batch normalization layer, which is used to normalize the output of the convolution layer to prevent overfitting;

[0053] An ELU activation function, which alleviates the problem of gradient disappearance through positive value identification, and has good robustness for negative value inputs. It is used to activate the output of the batch normalization layer, and the average value of the output of the activation function is close to zero to accelerate the convergence speed;

[0054] A Dropout layer, which is used to force the output of neurons to zero with a random probability to avoid overregularization.

[0055] Preferably, a Sigmoid function is configured in the fully connected layer, and the probability distribution of the class is obtained through the Sigmoid function, and the outputs of multiple neurons in the fully connected layer are mapped to the interval (0, 1);

[0056] A batch normalization layer is configured between the fully connected layer and the flat layer, and the batch normalization layer is used to normalize the data input to the fully connected layer to prevent overfitting.

[0057] Preferably, the optimization model is offline trained based on the training set, including the following steps:

[0058] Initialize the hyperparameters of the hybrid beamforming network model, where the hyperparameters include weight parameters, offsets, number of network layers, number of input neurons in each network layer, number of output neurons, number of input channels, number of output channels, learning rate, number of training epochs, and number of samples input to the hybrid beamforming network model for each training;

[0059] Input the preprocessed channel samples in the training set into the hybrid beamforming network model in batches for offline training;

[0060] During the offline training process, based on the Adam optimizer, minimize the loss function through the backpropagation algorithm, and perform iterative updates of the parameters according to the learning rate dynamic decay strategy until the loss tends to balance and stop the iteration. The calculation formula of the loss function is:

[0061]

[0062] where N represents the total number of training samples. Additionally, γ n , h n , respectively represent the signal-to-noise ratio, channel vector, and analog beamforming vector related to the nth sample. The superscript H represents conjugate transpose, and N t represents the number of transmit antennas at the base station side.

[0063] A hybrid beamforming optimization method based on deep learning of the present invention has the following advantages:

[0064] 1. A low-complexity hybrid beamforming network model is constructed based on a convolutional neural network to optimize the hybrid beamforming design. This hybrid beamforming network model can learn more features from the channel information and quickly predict the analog beamforming solution. Additionally, the convolutional neural network can achieve parameter sharing, thereby reducing the number of parameters and floating-point operations of the neural network. Therefore, this hybrid beamforming network model has a low complexity, and the hybrid beamforming network model of this application can achieve excellent performance at low complexity;

[0065] 2. It can achieve a hybrid beamforming design with non-convex constraints, where multiple convolutional blocks are used to effectively extract more features. The convolutional neural network has strong feature extraction capabilities and can learn more features from the channel information;

[0066] 3. A convolutional neural network can reduce the number of learning parameters by sharing weights and offsets, achieve excellent spectral efficiency performance at low complexity, and customize network layers in the network model to meet the constraint conditions to maximize spectral efficiency under the constraints.

[0067] 4. Offline training of the hybrid beamforming network model based on an unsupervised learning mechanism realizes unsupervised learning of the proposed network model, thus avoiding the problem that the network model requires labeled data while in fact it is difficult to obtain labeled data for hybrid beamforming. It can save a large amount of time and computing resources, perform unsupervised training on the network, and learn how to adaptively predict feasible beamforming solutions by maximizing spectral efficiency, thereby reducing system complexity and improving network training performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] The present invention will be further described below with reference to the drawings.

[0070] Figure 1 FIG. is a schematic diagram of the downlink single-user MISO system model structure in a hybrid beamforming optimization method based on deep learning according to an embodiment;

[0071] Figure 2 FIG. is a flowchart of a hybrid beamforming optimization method based on deep learning according to an embodiment;

[0072] Figure 3 FIG. is a schematic diagram of the structure of a hybrid beamforming network model in a hybrid beamforming optimization method based on deep learning according to an embodiment;

[0073] Figure 4 FIG. is a relationship diagram between signal-to-noise ratio and spectral efficiency in a hybrid beamforming optimization method based on deep learning according to an embodiment;

[0074] Figure 5 FIG. is a relationship diagram between signal-to-noise ratio and spectral efficiency in a hybrid beamforming optimization method based on deep learning according to an embodiment;

[0075] Figure 6 FIG. is a relationship diagram between signal-to-noise ratio and spectral efficiency in a hybrid beamforming optimization method based on deep learning according to an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it. However, the embodiments given are not intended to limit the present invention. Without conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0077] An embodiment of the present invention provides a hybrid beamforming optimization method based on deep learning, which is used to solve the technical problem of jointly optimizing a digital beamformer and an analog beamformer based on deep learning on the premise of reducing complexity and improving performance.

[0078] Embodiment:

[0079] A hybrid beamforming optimization method based on deep learning of the present invention is applied to a downlink single-user MISO hybrid beamforming system as shown in Figure 1 The system includes a transmitter, a channel, and a receiver. The transmitter is configured with a base station, and there is 1 radio frequency chain and N t transmitting antennas at the base station. The receiver is configured with a user, and there is 1 receiving antenna at the user.

[0080] The base station is equipped with a uniform linear array composed of N t antenna elements. Specifically, the transmitting antennas at the base station are arranged in a uniform linear manner. The present invention stipulates that the antenna spacing r at the base station is half of the transmission wavelength λ, and the formula is: r = 0.5λ.

[0081] The input signal s at the base station follows a complex Gaussian distribution with a mean of 0 and a variance of 1, that is, s ~ CN(0, 1).

[0082] The base station is used to transmit a data stream to a single-antenna user, and the user is used to receive the data stream.

[0083] A hybrid beamformer is equipped at the base station end, which is composed of a low-dimensional baseband digital beamformer and a high-dimensional radio frequency analog beamformer.

[0084] The digital beamformer v D is used to transmit the input signal s. Since there is only one radio frequency chain at the transmitter end, the digital beamforming v D is actually a scalar here.

[0085] The analog beamformer is used to transmit the input signal s to the transmitting antennas. The analog beamformer is implemented by phase shifters. It is realized to connect multiple antennas with one radio frequency chain. However, these phase shifters make it impossible to adjust the amplitude of the signal and can only change the phase of the signal. Therefore, the analog beamformer needs to have a constant modulus constraint Specifically, the modulus value of each element of the analog beamforming vector is 1.

[0086] The communication system needs to have a transmit power constraint to ensure the normal operation of the communication system. Therefore, there is a transmit power constraint at the base station side ||v A v D || 2 ≤P max , where P max represents the maximum transmit power at the base station.

[0087] In a hybrid beamforming system, the input signal s first passes through a digital beamformer v D , and then is frequency-converted to an analog phase shifter through one RF chain, and then passes through the analog beamforming vector v A to construct the transmit signal The formula is:

[0088] x = v A v D s

[0089] The hybrid beamforming vector for the entire downlink can be expressed as:

[0090] v = v A v D

[0091] The transmit signal x then passes through a channel h, and the received signal y can be obtained at the receiving end. The signal received at the user is:

[0092] y = h H v A v D s + n

[0093] where, represents the complex vector of the downlink channel gain, N t represents the number of transmit antennas, the superscript H represents the conjugate transpose, represents the analog beamformer, v D represents the digital beamformer, s represents the input signal at the base station, n represents the additive white Gaussian noise, which follows a complex Gaussian distribution with a mean of 0 and a variance of σ 2 , that is, n ~ CN(0, σ 2 ), and σ 2 represents the noise power.

[0094] The signal-to-noise ratio received by the user is:

[0095]

[0096] where, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose, denotes the analog beamformer, v D denotes the digital beamformer, σ 2 denotes the noise power.

[0097] The downlink spectral efficiency is obtained as follows, and the formula is:

[0098]

[0099] where denotes the complex vector of the downlink channel gain, and the superscript H denotes the conjugate transpose, denotes the analog beamformer at the base station side, v D denotes the digital beamformer at the base station side, σ 2 denotes the noise power.

[0100] The method includes the following steps:

[0101] S100. Formulate the hybrid beamforming optimization problem, and the hybrid beamforming optimization problem is equivalent to obtaining the beamforming solution by maximizing the spectral efficiency under the conditions of constant modulus constraint and power constraint;

[0102] S200. Collect different realizations of the channel as channel information, construct a channel sample set with the channel information as channel samples, the channel samples are complex channel matrices, the real part of the channel samples is the amplitude of the channel, and the imaginary part of the channel samples is the phase of the channel;

[0103] S300. Preprocess the channel samples to obtain the preprocessed channel samples in real number form, and divide the preprocessed channel samples into a training set and a test set;

[0104] S400. Construct a hybrid beamforming network model for optimizing hybrid beamforming based on deep learning. The hybrid beamforming network model takes channel information as input, and predicts and outputs the negative value of the spectral efficiency through learning the channel information;

[0105] S500. Offline train the hybrid beamforming network model based on the training set, and save the weight parameters of the hybrid beamforming network model during the offline training process to obtain the trained hybrid beamforming network model;

[0106] S600. Online test the trained hybrid beamforming network model based on the test set to obtain a feasible beamformer.

[0107] Step S100 constructs a formula for the hybrid beamforming optimization problem. As a specific implementation of Step S100, the optimization objective of this method is spectral efficiency, so the downlink spectral efficiency is selected as the objective function. Specifically, the signal-to-noise ratio received by the user is:

[0108]

[0109] where h represents an N t ×1-dimensional complex vector of downlink channel gains, and the superscript H represents conjugate transpose; v A represents the analog beamformer at the base station side, which is an N t ×1-dimensional complex vector; v D represents the digital beamformer at the base station side. Since there is only one radio frequency chain at the transmitting end, the digital beamformer v D is actually a scalar; σ 2 represents the noise power.

[0110] The spectral efficiency can be obtained as:

[0111]

[0112] where represents the complex vector of downlink channel gains, and the superscript H represents conjugate transpose, represents the analog beamformer at the base station side, v D represents the digital beamformer at the base station side, and σ 2 represents the noise power.

[0113] The hybrid beamforming optimization problem constructed by the present invention can be expressed as maximizing the spectral efficiency under the conditions of transmit power constraint and constant modulus constraint to obtain the optimal hybrid beamforming vector. The specific formula is:

[0114]

[0115] s.t. ||v A v D || 2 ≤P max

[0116]

[0117] where max represents the maximum value, represents the complex vector of downlink channel gains, and the superscript H represents conjugate transpose, represents the analog beamformer, v D represents the digital beamformer, v = v A v D represents the hybrid beamforming vector of the entire downlink, and σ2 represents the noise power, P max represents the maximum transmit power, |||| 2 represents the square of the norm, [v A i represents the i-th element of v A , N t represents the number of transmit antennas at the base station side.

[0118] In an actual communication system, since the analog beamformer is implemented by simple phase shifters, these phase shifters make it impossible to adjust the amplitude of the signal and can only change the phase of the signal. Therefore, the analog beamforming vector v A needs to have a constant modulus constraint. Specifically, the modulus of each element of the analog beamforming vector is 1, that is It can be deduced that From the transmit power constraint ||v A v D || 2 ≤ P max it can be seen that to obtain the maximum spectral efficiency, the transmission power needs to be utilized to the maximum extent. Therefore, the present invention assumes that ||v A v D || 2 = P max , then the optimal wherein, N t represents the number of transmit antennas at the base station side, represents the analog beamformer, v D represents the digital beamformer, P max represents the maximum transmit power, |||| 2 represents the square of the norm, max represents the maximum value.

[0119] Thus, it can be known that the final hybrid beamforming optimization problem can be formulated as finding the optimal analog beamforming vector that maximizes the spectral efficiency under the constant modulus constraint. More specifically, the hybrid beamforming design corresponds to the following formula:

[0120]

[0121]

[0122] wherein, max represents the maximum value, P max represents the maximum transmit power, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose, represents the analog beamformer, N t represents the number of transmit antennas at the base station side, σ 2 represents the noise power, [v​A i denote the i-th element of v A .

[0123] Step S200 acquires channel information to construct a channel sample set. In an actual communication system, channel parameters are generally obtained in real time. The acquired channel information is a complex channel matrix composed of information such as the number of transmitting antennas, the number of receiving antennas, and the number of users. The real part of the complex channel matrix is the amplitude information of the channel, and the imaginary part of the complex channel matrix is the phase information of the channel.

[0124] In this example, a Rayleigh fading channel environment where channel parameters follow a standard complex Gaussian distribution and a geometric millimeter-wave channel environment with very limited multipath components are respectively simulated, and the channel h is acquired.

[0125] To ensure the generality of the network, this embodiment respectively obtains N different channel realizations for training and testing the network. Each channel realization is a complex vector of size 1×N t . The acquired channel sample set is divided into a training sample set and a testing sample set. Among them, N t represents the number of transmitting antennas at the base station side, and 1 represents that the number of receiving antennas at the user side is 1.

[0126] Step S300 performs preprocessing operations on the channel samples. Currently popular neural network frameworks such as TensorFlow and PyTorch do not support complex numbers as network inputs and can only receive real number inputs. However, channels and beamforming matrices are essentially complex numbers. Among them, the real part of the channel is the amplitude, and the imaginary part is the phase. And real-valued neural networks are easier to implement and can achieve better performance. Therefore, the acquired channel data samples need to be processed into a form that can be recognized and processed by the network model. Specifically, the complex channel samples input to the network need to be converted into a real number form. After splitting the real part and the imaginary part of each complex channel sample, they are rearranged element by element in order into a three-dimensional real number matrix of size 1×N t ×2, where 1 represents a single receiving antenna, N t represents the number of transmitting antennas, and 2 represents splitting the real part component and the imaginary part component of the complex channel to form two channels. The samples are input to the neural network module batch by batch during the training process.

[0127] Step S400 constructs a hybrid beamforming network model for optimizing hybrid beamforming based on deep learning.

[0128] As a specific implementation, the network model includes an input layer, three groups of convolutional blocks, a flatten layer, a fully connected layer, a Lambda-1 layer, a Lambda-2 layer, and an output layer. Among them, each group of convolutional blocks includes a convolutional layer, a batch normalization layer, an activation layer, and a Dropout layer.​

[0129] The input layer receives a channel sample, which is a three-dimensional real matrix of size 1×N t ×2, where 1 represents the single antenna at the receiving end, N t represents the number of antennas at the transmitting end, and 2 represents splitting the real and imaginary components of the complex channel to form two channels.

[0130] Three convolutional blocks are used to extract features from the input data. Specifically, the first convolutional block includes the following steps: using a convolutional layer with 16 convolutional kernels of size 1×3 to perform a convolution operation on the input channel matrix with a stride of 1. The convolutional layer performs spatial (height and width) and inter-channel feature fusion on the local area; then using batch normalization to normalize the output of the convolutional layer. Batch normalization is a regularization technique to prevent overfitting, which can achieve a higher learning rate and thus accelerate convergence. Batch normalization accelerates network learning by reducing internal covariate shift; after batch normalization, using the ELU activation function for activation processing. It can alleviate the problem of vanishing gradients through positive identifications and has good robustness for negative inputs. It attempts to make the average value of the output of the activation function close to zero, thereby accelerating the convergence speed; then adding a Dropout layer to randomly discard a certain proportion of neurons. Dropout is a technique that forces the output of neurons to zero with a random probability. The random probability of Dropout is set to a very small value, i.e., 0.05, to avoid over-regularization problems. In the second convolutional block, the convolutional layer uses 8 convolutional kernels of size 1×3 to perform a convolution operation on the input response with a stride of 1. Then, similarly, using a batch normalization layer for normalization, using the ELU activation function for activation, and using the Dropout layer to randomly reset the output of each hidden neuron to zero with a probability of 0.05. In the third convolutional block, the convolutional layer uses 4 convolutional kernels of size 1×3 to perform a convolution operation on the input response with a stride of 1. Then, similarly, using a batch normalization layer for normalization, using the ELU activation function for activation, and using the Dropout layer to randomly reset the output of each hidden neuron to zero with a probability of 0.05 to reduce the impact caused by the initial weight selection.

[0131] After the feature extraction operation of the three convolutional blocks, a flatten layer is used to convert the multi-dimensional feature data of size 1×N t ×4 output by the convolutional block into a one-dimensional vector.

[0132] With N tA fully connected layer of neurons is applied to the output of the flatten layer to generate phase information related to the analog beamformer. The Sigmoid function is used to process the data of the fully connected layer to obtain the probability distribution of classes, and the outputs of multiple neurons are mapped to the interval (0, 1). The formula is: P = Sigmoid, where P represents the probability value of the class, and a batch normalization layer is used before the fully connected layer to prevent overfitting.

[0133] The custom layer, i.e., the Lambda layer, in the neural network framework is added to the network to satisfy the constant modulus constraint. Specifically, the Euler formula is used to convert the real phase vector θ into the complex analog beamforming vector v A = e iθ , where i 2 = -1.

[0134] Finally, another custom Lambda layer is added to obtain the network output value of the sample. Among them, the output value of the final network is actually the negative value of the spectral efficiency value. The formula is:

[0135]

[0136] Among them, γ, h, v A respectively represent the signal-to-noise ratio, channel vector, and analog beamforming vector related to the sample. Among them, the analog beamforming vector is the value output by the first Lambda layer, and the superscript H represents the conjugate transpose. N t represents the number of transmit antennas at the base station side.

[0137] The above is to obtain the output value of a channel sample, and the average value of the outputs of all channel samples is taken to obtain the final output value of this hybrid beamforming network model.

[0138] Step S500 uses the training sample set to perform offline training on this hybrid beamforming network model.

[0139] First, initialize the hyperparameters of the network model. Hyperparameters usually include weight parameters, offsets, the number of network layers, the number of input neurons, output neurons, input channels, output channels, learning rate, number of training epochs, and the number of samples input to the network for each training, etc.

[0140] Then, perform offline training on this hybrid beamforming network model based on the training set and save the trained weight parameters.

[0141] The goal of training is to find a feasible analog beamformer by maximizing the spectral efficiency. The training sample set preprocessed in step S3 is input into the neural network model constructed in step S4 in batches for offline training. The number of training epochs is set to 1000, and 16 channel samples are input into the neural network model in each training epoch. The Adam optimizer is used to optimize the parameters of the loss function, and the initial learning rate is set to 0.01. Further, a learning rate dynamic decay strategy is adopted. Specifically, if no improvement in the performance of the neural network model is observed within 20 training epochs, the learning rate is reduced. The learning rate will be reduced in the form of lr = lr * factor, where lr represents the current learning rate and factor represents the factor for reducing the learning rate each time. In the present invention, factor = 0.2 is set.

[0142] The constructed loss function is a custom objective function. In the present invention, the loss function of the neural network model is set to the negative of the system spectral efficiency. The calculation formula of the loss function is:

[0143]

[0144] where N represents the total number of training samples. In addition, γ n , h n , respectively represent the signal-to-noise ratio value, channel vector, and analog beamforming vector related to the nth sample. The superscript H represents conjugate transpose, and N t represents the number of transmit antennas at the base station side. To improve the robustness of the neural network to different signal-to-noise ratios, the present invention inputs the signal-to-noise ratio values of different samples into the network. When training the network model and minimizing the loss function, it is accompanied by the maximization of the system spectral efficiency.

[0145] During the training process, the loss function is used to perform backpropagation on the neural network model, continuously update the neural network parameters to optimize the neural network model, and save the training weight parameters during the training process. Further, the offline training method is an unsupervised learning method, which does not require a large amount of labeled data and only requires channel state information for network training. The proposed unsupervised learning mechanism can train the neural network to learn how to adaptively predict a feasible beamforming solution according to the input information without prior knowledge of the optimal target.

[0146] Step S600 uses the test sample set to perform online testing on the trained hybrid beamforming network model to obtain a feasible beamformer when the spectral efficiency takes the maximum value.

[0147] During the online testing phase, all the parameters of the hybrid beamforming network model have been trained and fixed. In the online testing phase, first, the trained weight parameters of the hybrid beamforming network model saved in the previous offline training phase are loaded. Then, the trained hybrid beamforming network model is used for online prediction to obtain the output result, where the output result is actually the negative value of the spectral efficiency value, that is, the loss function value.

[0148] This embodiment considers a downlink MISO system model where a base station equipped with N t transmit antennas and 1 radio frequency chain serves 1 single-antenna user for hybrid beamforming design. Simulation experiments are used to verify the performance of the MISO hybrid beamforming algorithm based on unsupervised convolutional neural network proposed in the present invention, and comparisons are made with the full-digital beamforming algorithm, traditional hybrid beamforming algorithm I, and traditional hybrid beamforming algorithm II. Traditional hybrid beamforming algorithm I is the prior art disclosed in the paper "Alternating minimization algorithms for hybrid precoding in millimeter wave MIMO systems", and traditional hybrid beamforming algorithm II is the prior art disclosed in the paper "Hybrid digital and analog beamforming design for large-scale antenna arrays". Further, the present invention also makes a comparison with the network architecture disclosed in the paper "Beamforming design for large-scale antenna arrays using deep learning" and the hybrid beamforming algorithm based on fully connected neural network.

[0149] As Figure 4 shown, it shows the comparison of the spectral efficiency performance of different hybrid beamforming schemes in a large-scale Rayleigh fading channel when N t = 64. As the signal-to-noise ratio increases, the spectral efficiency values of all beamforming algorithms gradually increase. Specifically, compared with the hybrid beamforming scheme, the full-digital beamforming method provides higher spectral efficiency performance. Further, it can be seen that under the same channel samples, the hybrid beamforming algorithm based on convolutional neural network proposed in the present invention achieves better performance than the traditional hybrid beamforming iterative algorithm and obtains higher spectral efficiency performance than the hybrid beamforming algorithm based on fully connected neural network.

[0150] In addition, as Figure 5 shown, it shows when N tWhen = 128, comparison of spectral efficiency performance achieved by different hybrid beamforming schemes in a large-scale Rayleigh fading channel. Compared with N t = 64, the number of transmit antennas at the base station has doubled. As can be seen from the figure, due to the increase in the number of antennas, the spectral efficiency values of all hybrid beamforming schemes have been significantly improved. Compared with the traditional hybrid beamforming algorithm and the hybrid beamforming algorithm based on a fully connected neural network, the hybrid beamforming algorithm based on a convolutional neural network proposed by the present invention still has higher spectral efficiency performance. And, similar to when N t = 64, the full-digital beamforming algorithm also obtains the highest spectral efficiency value, which is higher than the spectral efficiency values of all other hybrid beamforming algorithms.

[0151] In addition, as Figure 6 shown, it shows the comparison of spectral efficiency performance achieved by different hybrid beamforming schemes in a large-scale geometric millimeter-wave channel with finite clusters when N t = 64. As the signal-to-noise ratio increases, the performance of the algorithm proposed by the present invention is second only to the full-digital beamforming algorithm, and has a higher spectral efficiency value compared with other hybrid beamforming algorithms. As mentioned above, in two classic channel scenarios, the hybrid beamforming algorithm based on a convolutional neural network proposed by the present invention has higher spectral efficiency performance than the traditional hybrid beamforming algorithm and the hybrid beamforming algorithm based on a fully connected neural network.

[0152] Since the fully connected neural network has the ability of global perception, therefore, the fully connected neural network has a serious problem, that is, there are too many parameters. While the convolutional neural network can achieve local perception, and the weights of different neurons in the convolutional layer are shared, which greatly reduces the parameters and improves the training performance of the entire neural network model, and can extract features more effectively.

[0153] It can be seen that compared with the traditional hybrid beamforming algorithm and the hybrid beamforming algorithm based on a fully connected neural network, the hybrid beamforming algorithm based on a convolutional neural network proposed by the present invention not only obtains more excellent spectral efficiency performance, but also achieves lower system complexity.

[0154] The above has detailedly demonstrated and explained the present invention through the drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above-mentioned multiple embodiments, those skilled in the art can know that more embodiments of the present invention can be obtained by combining the code review means in the above different embodiments, and these embodiments are also within the protection scope of the present invention.

Claims

1. A hybrid beamforming optimization method based on deep learning, characterized in that Applied to a downlink single-user MISO hybrid beamforming optimization system, the method includes the following steps: Formulate the hybrid beamforming optimization problem, which is equivalent to obtaining a beamforming solution by maximizing the spectral efficiency under the conditions of constant modulus constraint and power constraint; Collect different realizations of the channel as channel information, construct a channel sample set with the channel information as channel samples, the channel samples are complex channel matrices, the real part of the channel samples is the amplitude of the channel, and the imaginary part of the channel samples is the phase of the channel; Preprocess the channel samples to obtain preprocessed channel samples in real number form, and divide the preprocessed channel samples into a training set and a test set; Construct a hybrid beamforming network model for optimizing hybrid beamforming based on deep learning. The hybrid beamforming network model takes channel information as input, predicts and outputs the negative value of the spectral efficiency by learning the channel information; Offline train the hybrid beamforming network model based on the training set, and save the weight parameters of the hybrid beamforming network model during the offline training process to obtain a trained hybrid beamforming network model; Online test the trained hybrid beamforming network model based on the test set to obtain a feasible beamformer; The hybrid beamforming network model includes: An input layer for receiving the preprocessed channel samples; A convolutional block for extracting features from the channel samples to obtain multi-dimensional feature data; A flattening layer for converting the multi-dimensional feature data into a one-dimensional vector; Fully connected layer, the fully connected layer is configured with N t neurons, applied to the one-dimensional vector output by the flat layer to generate phase information related to the analog beamformer; The first Lambda layer, which is used to satisfy the constant modulus constraint, converts the real phase vector θ into a complex analog beamforming vector v through Euler's formula A = e iθ , where i 2 = -1; A second Lambda layer for outputting the negative value of the spectral efficiency as the output value of the hybrid beamforming network model. The formula for calculating the negative value of the spectral efficiency is: where γ, h, and v A represent the signal-to-noise ratio, channel vector, and analog beamforming vector related to the channel sample, respectively. Among them, the analog beamforming vector is the value output by the first Lambda layer, and the superscript H represents the conjugate transpose. N t represents the number of transmit antennas at the base station side.

2. The hybrid beamforming optimization method based on deep learning according to claim 1, wherein The system includes a transmitter, a channel, and a receiver; The transmitter is a base station configured with a hybrid beamformer, and one radio frequency chain and multiple transmit antennas are configured at the base station; The hybrid beamformer includes a low-dimensional digital beamformer and a high-dimensional analog beamformer. The digital beamformer is used to transmit the signal to the one radio frequency chain, and the analog beamformer is implemented by multiple phase shifters and is used to connect the one radio frequency chain to the multiple transmit antennas; The receiver is configured with one user, and the user is configured with one receive antenna. The base station configured with multiple transmit antennas cooperates with the user configured with one receive antenna through the channel for signal transmission.

3. The hybrid beamforming optimization method based on deep learning according to claim 2, characterized in that The phase shifter makes it impossible to adjust the amplitude of the signal and can only change the phase of the signal. The analog beamformer has a constant modulus constraint The base station side has a transmit power constraint ‖v A v D ‖ 2 ≤P max , where P max represents the maximum transmit power at the base station; For the system, the input signal s passes through the digital beamformer v D After transmission, it is frequency-converted to the analog phase shifter through the one RF chain, and then through the analog beamforming vector v A Constructed to transmit a signal The formula is: x = v A v D s Downlink hybrid beamforming vector can be expressed as: v = v A v D After the transmitted signal x passes through a channel h, the signal y received by the user at the receiver is expressed as: y = h H v A v D s + n wherein, represents the downlink channel gain complex vector, N t represents the number of transmit antennas, and the superscript H represents the conjugate transpose, represents the analog beamformer, v D represents the digital beamformer, s represents the input signal at the base station, n represents the additive white Gaussian noise, which follows a complex Gaussian distribution with a mean of 0 and a variance of σ 2 , that is, n ~ CN(0, σ 2 ), and σ 2 represents the noise power; The signal-to-noise ratio received by the user is: Among them, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose. represents the analog beamformer, and v D represents the digital beamformer, and σ 2 represents the noise power. Based on the signal-to-noise ratio received by the user, the formula for calculating the downlink spectral efficiency is: Among them, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose, represents the analog beamformer at the base station side, v D represents the digital beamformer at the base station side, σ 2 represents the noise power.

4. The hybrid beamforming optimization method based on deep learning according to claim 1, 2 or 3, characterized in that Formulate the hybrid beamforming optimization problem, and the corresponding calculation formula is: s.t.‖v A v D ‖ 2 ≤P max where max represents the maximum value, represents the complex vector of the downlink channel gain, and the superscript H represents the conjugate transpose, represents the analog beamformer, v D represents the digital beamformer, v = v A v D represents the hybrid beamforming vector of the entire downlink, σ 2 represents the noise power, P max represents the maximum transmit power, ‖‖ 2 represents the square of the norm, [v A i represents the i-th element of v A , N t represents the number of transmit antennas at the base station side.​ 5. The hybrid beamforming optimization method based on deep learning according to claim 4, wherein Collect different realizations of the channel as channel information, and the channel environment simulates a Rayleigh fading channel where the channel parameters follow a standard complex Gaussian distribution and a geometric millimeter wave channel with very limited multipath components respectively; The channel information is a 1×N t complex vector, where N t represents the number of transmit antennas at the base station side, and 1 represents that the number of receive antennas at the user side is 1.

6. The hybrid beamforming optimization method based on deep learning according to claim 5, wherein The channel samples are preprocessed as follows: After splitting the real and imaginary parts of the complex channel samples, they are rearranged element by element in order into a three-dimensional real matrix of size 1×N t ×2, where 1 represents a single antenna at the receiving end, N t represents the number of antennas at the transmitting end, and 2 represents splitting the real and imaginary components of the complex channel samples to form two channels.

7. The hybrid beamforming optimization method based on deep learning according to claim 5, wherein There are three convolutional blocks in total, and each convolutional block includes: Convolutional layer, which is used to perform convolution operations on the input data; Batch normalization layer, which is used to normalize the output of the convolutional layer to prevent overfitting; ELU activation function, which alleviates the problem of vanishing gradients through positive identification and has good robustness for negative inputs. It is used to activate the output of the batch normalization layer and make the average value of the output of the activation function close to zero to accelerate the convergence speed; Dropout layer, which is used to force the output of neurons to zero with a random probability to avoid over-regularization.

8. The hybrid beamforming optimization method based on deep learning according to claim 5, characterized in that The fully connected layer is configured with a Sigmoid function to obtain the probability distribution of classes and map the outputs of multiple neurons in the fully connected layer to the interval (0, 1); A batch normalization layer is configured between the fully connected layer and the flattening layer, and the batch normalization layer is used to normalize the data input to the fully connected layer to prevent overfitting.

9. The hybrid beamforming optimization method based on deep learning according to claim 5, wherein Offline training of the optimization model based on the training set includes the following steps: Initialize the hyperparameters of the hybrid beamforming network model, and the hyperparameters include weight parameters, offsets, number of network layers, number of input neurons in each network layer, number of output neurons, number of input channels, number of output channels, learning rate, number of training epochs, and number of samples input to the hybrid beamforming network model for each training; Input the preprocessed channel samples in the training set into the hybrid beamforming network model in batches for offline training; During the offline training process, based on the Adam optimizer, minimize the loss function through the backpropagation algorithm, and perform iterative updates of the parameters according to the learning rate dynamic decay strategy until the loss tends to balance and stop the iteration. The calculation formula of the loss function is: Among them, N represents the total number of training samples. In addition, γ n , h n , respectively represent the signal-to-noise ratio, channel vector, and analog beamforming vector related to the nth sample. The superscript H represents the conjugate transpose. N t represents the number of transmit antennas at the base station side.

Images