A Deep Learning-Based Large-Scale MIMO Hybrid Beamforming Method

Abstract

This invention discloses a deep learning-based large-scale MIMO hybrid beamforming method that can achieve hybrid beamforming using imperfect channel state information. This method is applicable to single-base station multi-user scenarios. It employs an SV millimeter-wave channel model to generate a large amount of perfect channel state information as a dataset. The perfect channel state information is then used to perform channel estimation using a sparse reconstruction channel estimation algorithm to obtain imperfect channel state information. A neural network model is constructed, with imperfect channel state information and noise power as inputs, and digital beamforming matrices and analog beamforming matrices as outputs. Unsupervised training is performed using the negative total downlink rate as a loss function. The converged neural network model is then tested with imperfect channel state information and noise power as inputs. Compared with the traditional OMP algorithm, this method exhibits lower time complexity and better performance.

Description

Technical Field

[0001] This invention belongs to the field of hybrid beamforming technology, specifically relating to a large-scale MIMO hybrid beamforming method based on deep learning. Background Technology

[0002] To meet the demands of the information age for ultra-high transmission rates, massive MIMO (Multiple Input Multiple Output) technology has become a key technology in 5G. By increasing the number of antennas at the transceiver end, it improves spatial multiplexing gain and diversity gain, making outstanding contributions to system capacity and reducing system power consumption, and is applied in all aspects of life.

[0003] Beamforming (BF) is a crucial technique for improving spectral efficiency in massive MIMO (Multi-User Machines). By controlling the amplitude and phase of the transmitted signal from the MIMO antenna array, the beam signal is converged and enhanced in a specified direction, reducing energy consumption in other directions. This improves the signal quality received at the receiver, reduces interference between users, and increases cell coverage. Therefore, researchers have begun to study how to design beamformers at the transmitter end to achieve both low system power consumption and increased spectral rate at a lower cost. Traditional purely digital beamforming methods require each antenna at the transceiver end to have its own radio frequency (RF) link, resulting in significant hardware costs. Traditional analog beamforming methods share a single RF link across all antennas, reducing hardware costs, but only one RF link allows for single-data-stream transmission, hindering multi-user scenarios. Combining the advantages of digital and analog beamforming, hybrid digital-analog beamforming has gradually become the mainstream research approach.

[0004] However, most existing hybrid beamforming algorithms require extensive iterative computation, resulting in high complexity, and largely rely on perfect Channel State Information (CSI), which is difficult to achieve in reality. With the development of deep learning, it has gradually found widespread application in the field of communications. Because offline-trained neural network models have lower complexity compared to traditional iterative algorithms, deep learning-based large-scale MIMO beamforming methods are gradually replacing traditional optimization methods. Currently, most deep learning-based beamforming methods are supervised training, requiring large amounts of labeled data, which is costly to obtain. Furthermore, the accuracy of existing traditional algorithms is significantly affected by the imperfect channel state information obtained through channel estimation. Summary of the Invention

[0005] Purpose of the invention: To achieve higher spectral efficiency for the system under imperfect channel state information, a deep learning-based method for large-scale MIMO hybrid beamforming is proposed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a large-scale MIMO hybrid beamforming method based on deep learning, comprising the following steps:

[0007] Step 1: Configure the base station at the transmitting end to be equipped with N t There are 1 antenna and N antennas. RF The number of data streams in the root transmit RF link is N. s The receiving end has K users, each equipped with a single antenna, while the transmitting end uses a uniform linear array, and each user has the same priority.

[0008] Step 2: At the transmitting end, a millimeter-wave channel estimation algorithm based on sparse reconstruction is used to estimate the channel of the millimeter-wave narrowband channel to obtain imperfect channel state information. and noise power σ 2 ;

[0009] Step 3: Obtain the imperfect channel state information after channel estimation. and noise power σ 2 Inputting a deep neural network beamforming model yields the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D .

[0010] Furthermore, the deep neural network beamforming model includes:

[0011] The input layer is used to input imperfect channel state information. and noise power σ 2 Input the first fully connected layer;

[0012] The first fully connected layer is used to process the imperfect channel state information of the input. and noise power σ 2 Perform forward computation, and then input the forward computation result of the first fully connected layer into the second fully connected layer;

[0013] The second fully connected layer is used to further perform forward computation on the forward computation result of the first fully connected layer, and then input the forward computation result of the second fully connected layer into the third fully connected layer;

[0014] The third fully connected layer is used to further perform forward computation on the forward computation result of the input second fully connected layer, and then input the forward computation result of the third fully connected layer into the fourth fully connected layer;

[0015] The fourth fully connected layer is used to perform forward computation on the result of the forward computation of the third fully connected layer, and then perform complex-value transformation on the result of the forward computation of the fourth fully connected layer to obtain the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D and input / output layers;

[0016] And the output layer, used to convert the analog beamforming matrix V RF and digital beamforming matrix V D Output.

[0017] Furthermore, the training steps of the deep neural network beamforming model specifically include:

[0018] Step S1: The millimeter-wave narrowband channel generates a large amount of perfect channel state information through the millimeter-wave SV channel model. and noise power σ 2 ;

[0019] Step S2: Use a sparse reconstruction-based millimeter-wave channel estimation algorithm to estimate the narrowband millimeter-wave channel from step S1, obtaining imperfect channel state information. and noise power σ 2 ;

[0020] Step S3: Obtain the imperfect channel state information obtained in step S2. and noise power σ 2 Input a deep neural network beamforming model, and output a simulated beamforming matrix V′. RF and digital beamforming matrix V′ D The corresponding perfect channel state information obtained in step S2 and noise power σ 2 Together, we calculate the loss function to train the deep neural network beamforming model, where the loss function is the negative value of the total user rate, i.e., -R.

[0021] Furthermore, the formula for the loss function is as follows:

[0022]

[0023] Where N is the total number of samples in one training iteration. V′ represents the perfect channel state information of the k-th user in the n-th sample. RF,n This represents the simulated beamforming matrix output by the deep neural network beamforming model for the nth sample. V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The kth column, V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The i-th column.

[0024] Furthermore, the training of the deep neural network beamforming model also includes testing the total downlink user reachability rate, which is based on the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D The specific steps for calculating the total downlink reachable rate for users include:

[0025] Step A1: The transmitting end sends data symbol s, which is then processed by the final digital beamforming matrix V. D and the final simulated beamforming matrix V RF The information sent by the sender to the k-th user is as follows:

[0026]

[0027] in It is the i-th column of the final digital beamforming matrix, s i It is the data stream vector of the i-th user;

[0028] Step A2: The information sent by the transmitter to the k-th user is transmitted to the k-th user at the receiver via a millimeter-wave narrowband channel, resulting in the signal received by the k-th user. The information received by the k-th user is:

[0029]

[0030] in(·) H Indicates conjugate transpose, n k It is additive white Gaussian noise at the k-th user location. For the perfect channel state information corresponding to the k-th user, s k Let k be the data stream vector of the k-th user;

[0031] Step A3: Calculate the total downlink achievable rate using the information sent to K users by the transmitter and the signals received by K users at the receiver. The formula for calculating the total downlink achievable rate is:

[0032]

[0033] in V represents the final digital beamforming matrix. D The kth column, This represents the i-th column of the final digital beamforming matrix.

[0034] Beneficial effects:

[0035] 1. This invention employs a deep learning-based hybrid beamforming method, which can be trained offline and has lower computational complexity than traditional digital, analog, and hybrid beamforming algorithms.

[0036] 2. This invention can achieve a hybrid beamforming scheme that provides near-perfect channel state information even with imperfect channel state information, and has higher accuracy than traditional hybrid beamforming algorithms.

[0037] 3. This invention uses an unsupervised approach to train neural network models, which requires lower label generation costs than supervised learning neural network models. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0039] Figure 2 This is a schematic diagram of a large-scale MIMO system with multiple users on a single base station according to the present invention;

[0040] Figure 3 This is a schematic diagram of the neural network model structure of the present invention;

[0041] Figure 4 This is a graph showing the test results comparing the present invention with existing OMP algorithms on test data.

[0042] Figure 3 In the diagram: 1. Input layer, 2. First fully connected layer, 3. Second fully connected layer, 4. Third fully connected layer, 5. Fourth fully connected layer, 6. Output layer. Detailed Implementation

[0043] The invention will now be further explained with reference to the accompanying drawings.

[0044] This invention provides a deep learning-based method for large-scale MIMO hybrid beamforming, comprising the following steps:

[0045] Step 1: Configure the base station at the transmitting end to be equipped with N t There are 1 antenna and N antennas. RF The number of data streams in the root transmit RF link is N. s The receiving end has K users, each equipped with a single antenna, while the transmitting end uses a uniform linear array, and each user has the same priority.

[0046] Step 2: At the transmitting end, a millimeter-wave channel estimation algorithm based on sparse reconstruction is used to estimate the channel of the millimeter-wave narrowband channel to obtain imperfect channel state information. and noise power σ 2 ;

[0047] Step 3: Obtain the imperfect channel state information after channel estimation. and noise power σ 2 Inputting a deep neural network beamforming model yields the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D .

[0048] See Figures 1-4 This implementation provides a large-scale MIMO hybrid beamforming method based on deep learning, specifically as follows: Figure 1 As shown, the method includes the following steps:

[0049] Step 1: Construct as follows Figure 2 The sending and receiving ends are shown below;

[0050] More specifically, step 1 includes: first, setting up the scenario, configuring the sending end to be equipped with N t There are 1 antenna and N antennas. RF The number of data streams in the root transmit RF link is N. s There are a total of K users at the receiving end, each equipped with a single antenna. The base station uses a uniform linear array, and each user has the same priority. K is a constant, which depends on the application environment.

[0051] In step 2, a millimeter-wave channel estimation algorithm based on sparse reconstruction is used to estimate the channel of the millimeter-wave narrowband channel. The transmitter transmits T pilot signals in T consecutive time periods, using T beamforming vectors f. p The user receiver splices the detected signals to obtain Y. The formula for this process is as follows:

[0052]

[0053] Where F = [f1, f2, ..., f T [ ] is the concatenation matrix of T beamforming vectors, and N is the noise matrix obtained by concatenating T noise vectors. During the pilot training phase, taking advantage of the sparsity of the millimeter-wave channel, Y is quantized and then, after a series of derivations, the formula is shown below:

[0054]

[0055] Where y v Let ρ be the quantization vector of Y, ρ be the average power of the pilot training, and A be the quantization vector of N. t The antenna response matrix is ​​given by the L paths of the root transmitting antenna, where α represents the complex gain vector of the L paths.

[0056] Now assume that N points are taken from the departure angle using a uniform grid, A D It is N tThe dictionary matrix is ​​×N dimensional, ignoring grid quantization error, and the formula is as follows:

[0057]

[0058] Define the perception matrix Ψ = F T A D , through y v And the perception matrix Ψ = F T A D Channel estimation based on sparse reconstruction is performed on millimeter-wave narrowband channels to obtain imperfect channel state information. and noise power σ 2 The estimated channel vector of the k-th user

[0059] After the deep neural network beamforming model training in step 3 converges, the imperfect channel state information obtained from the channel estimation in step 2 is input. and noise power σ 2 The model outputs the final digital beamforming matrix V. D and the final simulated beamforming matrix V RF To achieve hybrid digital-analog beamforming.

[0060] See Figure 3 The deep neural network beamforming model in step 3 of this embodiment includes: input layer 1, first fully connected layer 2, second fully connected layer 3, third fully connected layer 4, fourth fully connected layer 5, and output layer 6.

[0061] Input layer 1 is used to input imperfect channel state information. and noise power σ 2 Input the first fully connected layer 2;

[0062] The first fully connected layer 2 is used to process the imperfect channel state information of the input. and noise power σ 2 Perform forward computation, and then input the forward computation result of the first fully connected layer 2 into the second fully connected layer 3.

[0063] The second fully connected layer 3 is used to further perform forward computation on the forward computation result of the first fully connected layer 2, and then input the forward computation result of the second fully connected layer 3 into the third fully connected layer 4.

[0064] The third fully connected layer 4 is used to further perform forward computation on the forward computation result of the second fully connected layer 3, and then input the forward computation result of the third fully connected layer 4 into the fourth fully connected layer 5.

[0065] The fourth fully connected layer 5 is used to perform forward computation on the forward computation results of the input third fully connected layer 4, and then perform complex value transformation on the forward computation results of the fourth fully connected layer 5 to obtain the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D And input / output layer 6.

[0066] And output layer 6, used to convert the analog beamforming matrix V RF and digital beamforming matrix V D Output.

[0067] The training steps of the deep neural network beamforming model in this embodiment include:

[0068] Step S1: The SV millimeter-wave channel model used in millimeter-wave narrowband channels includes one line-of-sight (LoS) path and L-1 non-line-of-sight (NLoS) paths. The formula for the k-th user channel vector is as follows:

[0069]

[0070] Where α l Let α represent the complex gain of the l-th path, satisfying independent and identically distributed α. l ~CN(0,1), a(θ) l ) represents the angle θ under the l-th path. l The array response vector.

[0071] Based on this SV millimeter-wave channel model, simulations generate a large dataset of perfect channel state information, including the channel vector of the k-th user. and noise power σ 2 .

[0072] Step S2: Perform channel estimation on the millimeter-wave narrowband channel from step S1, using the same estimation steps as in step 2 above, to obtain imperfect channel state information. and noise power σ 2 .

[0073] Step S3: As Figure 3 The deep neural network beamforming model shown uses the imperfect channel state information obtained in step S2. Decomposed into real and imaginary parts and noise power σ 2 Together, they serve as input to a deep neural network beamforming model, with an input dimension of 2×K×N. t +1, the second to fifth layers of the network are fully connected layers, with neurons set to 2048, 1024, 512, and 256 respectively. Each layer uses the Rectified Luminous Activation Function (ReLU). The output Nt ×N RF Real value Representing the phase, to satisfy the constant mode constraint of simulated beamforming, the simulated beamforming matrix V′ is obtained through complex value transformation. RF The formula is shown below:

[0074]

[0075] Another part of the output of the deep neural network beamforming model is 2×N s ×N RF The N-dimensional real value is obtained by complex value transformation. s ×N RF Digital beamforming matrix V″ D The power normalization constraint is then applied, as shown in the following formula:

[0076] V′ D =V″ D / ||V′ RF V″ D || F (6)

[0077] Among them, ||·|| F This represents the F-norm of a matrix.

[0078] The simulated beamforming matrix V′ RF and digital beamforming matrix V′ D The corresponding real channel state information obtained in step S2 and noise power σ 2 The loss function is calculated together to train the deep neural network beamforming model, enabling it to converge. The optimization objective is to maximize the total downlink reachable rate of the system. The loss function uses the negative value of the total downlink reachable rate of the system, as shown in the following formula:

[0079]

[0080] Where N is the total number of samples in one training iteration. V′ represents the perfect channel state information of the k-th user in the n-th sample. RF,n This represents the simulated beamforming matrix output by the deep neural network beamforming model for the nth sample. V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The kth column, V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The i-th column.

[0081] In this embodiment, the training of the deep neural network beamforming model also includes testing the total downlink user reachability rate, which is based on the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D The specific steps for calculating the total downlink reachable rate for users include:

[0082] Step A1: The transmitting end sends data symbol s, which is then processed by the final digital beamforming matrix V. D and the final simulated beamforming matrix V RF The signal sent to K users at the transmitting end is calculated using the following formula:

[0083]

[0084] in It is the i-th column of the final digital beamforming matrix, s i This is the data stream vector of the i-th user. Step A2: The information sent by the transmitter to the k-th user is transmitted to the k-th user at the receiver via a millimeter-wave narrowband channel, resulting in the signal received by the k-th user. The formula for calculating the information received by the k-th user is as follows:

[0085]

[0086] in(·) H Indicates conjugate transpose, n k It is additive white Gaussian noise at the k-th user location. For the perfect channel state information corresponding to the k-th user, s k Let be the data stream vector of the k-th user.

[0087] Step A3: Construct the achievable downlink rate for the k-th user using the signals sent to the k users by the transmitter and the signals received by the k-th user by the receiver, as shown in the following formula:

[0088]

[0089] in, V represents the final digital beamforming matrix. D The kth column, V represents the final digital beamforming matrix. D The i-th column, σ 2 Indicates noise power.

[0090] The system spectral efficiency is equivalent to the total downlink achievable rate of the system, as shown in the following formula:

[0091]

[0092] in, V represents the final digital beamforming matrix. D The kth column, σ represents the i-th column of the final digital beamforming matrix. 2 Indicates noise power.

[0093] Final simulated beamforming matrix V RF The constant modulus constraint condition must be met |[V] RF ] i,j | = 1, the transmitting end must satisfy the maximum transmit power normalization constraint condition ||V RF V D || F ≤1.

[0094] from Figure 4 It can be seen that the frequency is 28GHz, and the number of transmitting antennas N at the base station is... t =64, number of RF links N RF =8, channel path number L=3, after model training convergence, compared with the traditional OMP algorithm, it can be seen that under imperfect channel state information, our invented model method has better performance and is closer to the results obtained by perfect channel state information.

[0095] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A large-scale MIMO hybrid beamforming method based on deep learning, characterized in that, Includes the following steps: Step 1: Configure the base station at the transmitting end to be equipped with N t There are 1 antenna and N antennas. RF The number of data streams in the root transmit RF link is N. s The receiving end has K users, each equipped with a single antenna, while the transmitting end uses a uniform linear array, and each user has the same priority. Step 2: At the transmitting end, a millimeter-wave channel estimation algorithm based on sparse reconstruction is used to estimate the channel of the millimeter-wave narrowband channel to obtain imperfect channel state information. and noise power σ 2 ; Step 3: Calculate the imperfect channel state information obtained after channel estimation. and noise power σ 2 Inputting a deep neural network beamforming model yields the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D The training steps for the deep neural network beamforming model specifically include: Step S1: The millimeter-wave narrowband channel generates a large amount of perfect channel state information through the millimeter-wave SV channel model. and noise power σ 2 ; Step S2: Use a sparse reconstruction-based millimeter-wave channel estimation algorithm to estimate the narrowband millimeter-wave channel from step S1, obtaining imperfect channel state information. and noise power σ 2 ; Step S3: Obtain the imperfect channel state information obtained in step S2. and noise power σ 2 Input a deep neural network beamforming model, and output a simulated beamforming matrix V′. RF and digital beamforming matrix V′ D The corresponding perfect channel state information obtained in step S2 and noise power σ 2 Together, we calculate the loss function to train the deep neural network beamforming model, where the loss function is the negative value of the total user rate, i.e., -R.

2. The deep learning-based large-scale MIMO hybrid beamforming method according to claim 1, characterized in that, The deep neural network beamforming model includes: Input layer (1) is used to input imperfect channel state information. and noise power σ 2 Input the first fully connected layer (2); The first fully connected layer (2) is used to process the input imperfect channel state information. and noise power σ 2 Perform forward computation, and then input the forward computation result of the first fully connected layer (2) into the second fully connected layer (3); The second fully connected layer (3) is used to further perform forward calculations on the forward calculation results of the first fully connected layer (2) and then input the forward calculation results of the second fully connected layer (3) into the third fully connected layer (4); The third fully connected layer (4) is used to further perform forward calculations on the forward calculation results of the input second fully connected layer (3), and then input the forward calculation results of the third fully connected layer (4) into the fourth fully connected layer (5); The fourth fully connected layer (5) is used to perform forward computation on the forward computation result of the input third fully connected layer (4), and then perform complex value transformation on the forward computation result of the fourth fully connected layer (5) to obtain the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D , and input to the output layer (6); And the output layer (6), used to convert the analog beamforming matrix V RF and digital beamforming matrix V D Output.

3. The deep learning-based large-scale MIMO hybrid beamforming method according to claim 1, characterized in that, The formula for the loss function is as follows: Where N is the total number of samples in one training iteration. V′ represents the perfect channel state information of the k-th user in the n-th sample. RF,n This represents the simulated beamforming matrix output by the deep neural network beamforming model for the nth sample. V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The kth column, V′ represents the digital beamforming matrix output by the deep neural network beamforming model for the nth sample. D The i-th column.

4. The deep learning-based large-scale MIMO hybrid beamforming method according to claim 3, characterized in that, The training of the deep neural network beamforming model also includes testing the total downlink user reachability rate, which is based on the final simulated beamforming matrix V. RF and the final digital beamforming matrix V D The specific steps for calculating the total downlink reachable rate for users include: Step A1: The transmitting end sends data symbol s, which is then processed by the final digital beamforming matrix V. D and the final simulated beamforming matrix V RF The information sent by the sender to K users is as follows: in, It is the i-th column of the final digital beamforming matrix, s i It is the data stream vector of the i-th user; Step A2: The information sent by the transmitter to the k-th user is transmitted to the k-th user at the receiver via a millimeter-wave narrowband channel, resulting in the signal received by the k-th user. The information received by the k-th user is: in,(·) H Indicates conjugate transpose, n k It is additive white Gaussian noise at the k-th user location. For the perfect channel state information corresponding to the k-th user, s k Let k be the data stream vector of the k-th user; Step A3: Calculate the total downlink achievable rate using the information sent to K users by the transmitter and the signals received by K users at the receiver. The formula for calculating the total downlink achievable rate is: in, V represents the final digital beamforming matrix. D The kth column, This represents the i-th column of the final digital beamforming matrix.

Images