A deep learning-based MIMO intelligent hybrid beamforming method

Abstract

A kind of MIMO intelligent hybrid beamforming method based on deep learning, comprising: constructing MIMO millimeter wave channel system, generating imperfect channel state information sample;Based on the pre-constructed DFT codebook, construct the beamforming dataset containing imperfect channel state information sample and its corresponding label, and then divide the training set and the validation set;Constructing double-branch multi-scale parallel convolutional neural network model, using training set for offline training, using validation set for dynamic evaluation, until model training converges;The optimal analog beamforming matrix satisfying the constant modulus constraint of analog phase shifter is output by the model;The optimal digital beamforming matrix is calculated by using the zero-forcing algorithm to suppress the interference between multiple users;MIMO intelligent hybrid beamforming is completed by cooperation. The present application can reduce the computational complexity, reduce the dependence on accurate channel state information, and improve the performance of beamforming method in complex environment.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network technology, and more specifically to a deep learning-based MIMO intelligent hybrid beamforming method. Background Technology

[0002] With the rapid development of wireless communication technology, modern communication systems are increasingly demanding higher bandwidth, lower latency, and higher reliability, driving related technologies to evolve towards higher performance and efficiency. In high-speed mobile scenarios such as high-speed railways and urban rail transit, the increasing speed of movement and user demands present even more severe challenges to communication systems in terms of bandwidth, transmission latency, and link reliability. Millimeter-wave communication, with its abundant spectrum resources, has become a core technology for meeting the demands of high-bandwidth, high-speed data transmission in high-speed mobile environments. Meanwhile, hybrid beamforming technology combines the advantages of analog and digital beamforming, effectively reducing RF links, lowering hardware costs and complexity, and maintaining high-quality transmission performance in high-speed dynamic environments. Therefore, hybrid beamforming technology has enormous application potential in massive MIMO (Multiple-Input Multiple-Output) systems, especially in millimeter-wave communication in high-speed mobile scenarios.

[0003] However, in high-speed mobile communication scenarios, the rapid movement of terminals leads to significant Doppler shift and time-varying multipath fading characteristics in millimeter-wave channels, increasing the complexity of beamforming design. Furthermore, the constant-mode constraints of analog phase shifters and the strong coupling between analog and digital beamforming make the hybrid beamforming optimization problem typically a non-convex optimization problem, difficult to solve precisely using traditional mathematical methods. To address this non-convex optimization challenge, existing technologies often employ traditional beamforming methods for approximate solutions. Specifically, traditional beamforming methods frequently use algorithms such as orthogonal matched pursuit and alternating minimization, but these methods generally suffer from high computational complexity and slow convergence speed, failing to meet the high real-time requirements of high-speed mobile communication scenarios. Moreover, traditional methods often rely on accurate channel state information, which in real-world environments is frequently affected by noise, estimation errors, and feedback delays, resulting in imperfect channel information. This significantly degrades the performance of traditional beamforming methods in complex environments, making them unsuitable for practical applications.

[0004] Furthermore, Chinese invention patent CN116405077A discloses a deep learning-based hybrid beamforming method for large-scale MIMO. This method primarily involves first generating perfect channel state information, then processing it using a channel estimation algorithm to obtain imperfect channel state information. The imperfect channel state information and noise power are then input into a neural network model, which outputs digital and analog beamforming matrices through unsupervised training. While this method can reduce the computational complexity of traditional hybrid beamforming algorithms to some extent, its imperfect channel state information originates from subsequent estimation processing of perfect channel state information, and the beamforming matrix learning is performed using unsupervised training. In high-speed mobile communication scenarios, the adaptability and robustness of hybrid beamforming under imperfect channel state information conditions are poor. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based MIMO intelligent hybrid beamforming method to reduce computational complexity, reduce dependence on accurate channel state information, and improve the performance of beamforming methods in complex environments.

[0006] A deep learning-based MIMO intelligent hybrid beamforming method includes: Step S1: In the downlink of a mobile communication scenario, equip the base station at the transmitting end with... One transmitting antenna, the user end is equipped with Root receiving antenna, and base station configuration The base station achieves this by transmitting a single data stream to each user via the root radio frequency link. Parallel transmission for each user; Step S2: Based on the actual channel characteristics of mobile communication scenarios, a MIMO millimeter-wave channel system is constructed. By superimposing independent noise interference terms, the channel estimation error and noise interference in actual communication are simulated to generate imperfect channel state information samples. A candidate simulated beam set is constructed based on the pre-constructed DFT codebook. A greedy search algorithm is used to select the codeword combination that minimizes the total mean square error of the system from the candidate simulated beam set. The simulated beamforming matrix corresponding to the codeword combination is used as a label to construct a beamforming dataset containing imperfect channel state information samples and their corresponding labels. The beamforming dataset is then divided into a training set and a validation set according to a preset ratio. Step S3: Construct a dual-branch, multi-scale parallel convolutional neural network model. Convert the imperfect channel state information into a three-channel input tensor consisting of the amplitude, real part, and imaginary part of the channel matrix, which serves as the input to the convolutional neural network model. Use the relative mean square error of the system and rate as the loss function, and train the convolutional neural network model offline using the training set. Update the network parameters in real time, and use the validation set to dynamically evaluate the network performance during the training process until the model training converges, thus obtaining a converged convolutional neural network model. Step S4: In online applications, the imperfect channel state information acquired in real-time under actual communication scenarios is converted into actual three-channel input tensors and input into the training and converged convolutional neural network model. The model outputs the optimal simulated beamforming matrix that satisfies the constant mode constraint of the simulated phase shifter. ; Step S5, in the optimal simulated beamforming matrix Based on this, the zero-forcing algorithm is used to suppress interference between multiple users, and the optimal digital beamforming matrix is ​​calculated. Finally, the optimal simulated beamforming matrix is... With the optimal digital beamforming matrix Working together, we can complete MIMO intelligent hybrid beamforming.

[0007] The deep learning-based MIMO intelligent hybrid beamforming method provided by the present invention has the following beneficial effects: 1. This invention improves upon traditional convolutional neural networks by using a dual-branch, multi-scale parallel convolutional neural network to predict simulated beamforming matrices. The convolutional neural network model of this invention includes an input layer, a first-layer feature extraction module, a second-layer feature extraction module, a fully connected mapping layer, and an output layer. It can better handle imperfect channel state information, avoid the complex iterative solution process in traditional mathematical optimization methods, and effectively reduce computational complexity and processing latency.

[0008] 2. This invention simulates channel estimation errors and noise interference in actual communication by superimposing independent noise interference terms, generating imperfect channel state information samples. Based on a pre-constructed DFT codebook, a greedy search algorithm is used to select the codeword combination that minimizes the total mean square error of the system from the candidate simulated beam set. The simulated beamforming matrix corresponding to this codeword combination is used as a label, making this invention more applicable to single-base station multi-user high-speed mobile communication scenarios under imperfect channel state information conditions. It can reduce the dependence on accurate channel state information, improve the robustness of the method, and enhance the performance of beamforming methods in complex environments.

[0009] 3. This invention employs a dual-branch, multi-scale parallel convolutional neural network to predict the simulated beamforming matrix and uses a zero-forcing algorithm to solve the digital beamforming matrix. This combines the fast mapping capability of deep learning methods with the ability of the zero-forcing algorithm to suppress multi-user interference, which is beneficial to improving the system and speed in multi-user scenarios, thereby enhancing the system's transmission performance and meeting the requirements of online real-time deployment at the base station.

[0010] 4. This invention simulates channel estimation errors and noise interference in actual communication by superimposing independent noise interference terms, generating imperfect channel state information samples. Then, it constructs a beamforming dataset containing the imperfect channel state information samples and their corresponding labels, resulting in a beamforming dataset directly addressing imperfect channel state information. Training is then performed based on this dataset. Furthermore, this invention uses the relative mean square error of the system and rate as the loss function, utilizes the training set to train the convolutional neural network model offline, updates the network parameters in real time, and uses the validation set to dynamically evaluate the network performance during training. This achieves supervised training for beamforming matrix learning, effectively improving the adaptability and robustness of hybrid beamforming in high-speed mobile communication scenarios. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the deep learning-based MIMO intelligent hybrid beamforming method provided by the present invention.

[0012] Figure 2 This is a block diagram of the architecture of the convolutional neural network model in this invention.

[0013] Figure 3 The diagram shows a comparison between the method provided by this invention and existing traditional beamforming algorithms at different signal-to-noise ratios. Detailed Implementation

[0014] To facilitate understanding of the present invention, a more complete description will be given below with reference to various embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0016] Please see Figure 1This invention provides a deep learning-based intelligent hybrid beamforming method for MIMO systems. This method is primarily designed for large-scale MIMO systems and includes steps S1-S5: Step S1: In the downlink of a mobile communication scenario, equip the base station at the transmitting end with... One transmitting antenna, the user end is equipped with Root receiving antenna, and base station configuration The base station achieves this by transmitting a single data stream to each user via the root radio frequency link. Parallel transmission for each user.

[0017] In this embodiment, the mobile communication scenario specifically refers to high-speed mobile communication scenarios such as high-speed railways and urban rail transit. To meet the requirements of multi-user reuse, the system must meet the following requirements: .

[0018] Step S2: Based on the actual channel characteristics of mobile communication scenarios, a MIMO millimeter-wave channel system is constructed. By superimposing independent noise interference terms, the channel estimation error and noise interference in actual communication are simulated to generate imperfect channel state information samples. A candidate simulated beam set is constructed based on the pre-constructed DFT codebook. A greedy search algorithm is used to select the codeword combination that minimizes the total mean square error of the system from the candidate simulated beam set. The simulated beamforming matrix corresponding to the codeword combination is used as a label to construct a beamforming dataset containing imperfect channel state information samples and their corresponding labels. The beamforming dataset is then divided into a training set and a validation set according to a preset ratio.

[0019] In this process, a candidate simulated beam set is constructed based on a pre-built DFT codebook. The candidate simulated beam set is a set of candidate simulated beamforming matrices formed by combining codewords in the DFT codebook according to a preset number of radio frequency links. Each candidate simulated beamforming matrix corresponds to a codeword combination.

[0020] Then, a greedy search algorithm is used to iteratively select codeword combinations from the candidate simulated beam set, and the codeword combination that minimizes the total mean square error of the MIMO millimeter-wave communication system is selected. The simulated beamforming matrix corresponding to the codeword combination is used as a label to construct a beamforming dataset containing imperfect channel state information samples and their corresponding labels. The beamforming dataset is then divided into a training set and a validation set according to a preset ratio. The training set and the validation set are used for the training and validation of the subsequent neural network model.

[0021] The formula for minimizing the total mean square error is expressed as:

[0022] in, Represents the matrix trace operation. Represents the identity matrix. Indicates the number of users. Represents the receiver merging matrix, superscript This indicates the conjugate transpose. Represents the total channel matrix. This represents the optimal simulated beamforming matrix. This represents the optimal digital beamforming matrix, with the superscript -1 indicating matrix inversion. , This represents the set of candidate simulated beams.

[0023] Step S3: Construct a dual-branch, multi-scale parallel convolutional neural network model. Convert the imperfect channel state information into a three-channel input tensor consisting of the amplitude, real part, and imaginary part of the channel matrix, which serves as the input to the convolutional neural network model. Use the relative mean square error of the system and rate as the loss function, and train the convolutional neural network model offline using the training set. Update the network parameters in real time, and use the validation set to dynamically evaluate the network performance during the training process until the model training converges, thus obtaining a converged convolutional neural network model.

[0024] Please refer to Figure 2 The convolutional neural network model includes an input layer, a first-layer feature extraction module, a second-layer feature extraction module, a fully connected mapping layer, and an output layer.

[0025] The input layer is used to receive dimensions of To preserve the complex-valued characteristics of the millimeter-wave channel, the imperfect channel state information is decomposed into a three-channel input tensor consisting of the magnitude, real part, and imaginary part of the channel matrix, and used as the input of a two-branch multi-scale parallel convolutional neural network.

[0026] The first-layer feature extraction module adopts a dual-branch parallel structure, including two parallel convolutional layers, one for extracting spatial features and the other for extracting spatial features. The first convolutional layer Conv and the layer used to extract global coupling features, The second convolutional layer Conv is connected to the first feature extraction module, which is then connected to an average pooling layer for feature compression.

[0027] The second-layer feature extraction module adopts a dual-branch parallel structure, including two parallel convolutional layers, one for extracting angular domain features and the other for extracting angular domain features. The third convolutional layer Conv and the function used to extract global coupling features The fourth convolutional layer Conv, and the second feature extraction module compresses the features into a low-dimensional feature vector through an average pooling layer.

[0028] The fully connected mapping layer is used to perform feature mapping through multiple fully connected layers. To suppress overfitting, a Dropout regularization layer is set after each fully connected layer, and the LeakyReLU activation function is used to stabilize feature propagation.

[0029] The output layer is used to output the optimal simulated beamforming matrix obtained from the prediction. .

[0030] In this embodiment, during the offline training phase, the relative mean square error (MSE) of the system and rate is used as the loss function. The formula for calculating the loss function is:

[0031] in, Represents the loss function. This indicates the total number of input samples each time the loss is calculated; Indicates the first The system and rate corresponding to the optimal simulated beamforming matrix predicted by the convolutional neural network for each input sample; Indicates the first The optimal system and rate corresponding to the dataset labels for each input sample.

[0032] Backpropagation is performed using the training set to update the network parameters, and the network performance is evaluated using the validation set until a convolutional neural network model that has been trained and converged is obtained.

[0033] Step S4: In online applications, the imperfect channel state information acquired in real-time under actual communication scenarios is converted into actual three-channel input tensors and input into the training and converged convolutional neural network model. The model outputs the optimal simulated beamforming matrix that satisfies the constant mode constraint of the simulated phase shifter. .

[0034] Step S5, in the optimal simulated beamforming matrix Based on this, the zero-forcing algorithm is used to suppress interference between multiple users, and the optimal digital beamforming matrix is ​​calculated. Finally, the optimal simulated beamforming matrix is... With the optimal digital beamforming matrix Working together, we can complete MIMO intelligent hybrid beamforming.

[0035] Specifically, step S5 includes: Step S5.1, based on the optimal simulated beamforming matrix F A Construct an equivalent channel matrix , ,in, This is the total channel matrix; Step S5.2: Obtain the unconstrained digital beamforming matrix using the zero-forcing algorithm. : ; Among them, superscript This indicates the conjugate transpose. The regularization coefficient is . It is the identity matrix; Step S5.3: Introduce a scaling factor that satisfies the total system transmit power constraint. The optimal digital beamforming matrix is ​​obtained. , Finally, by combining the optimal analog beamforming matrix and the optimal digital beamforming matrix, MIMO intelligent hybrid beamforming is completed.

[0036] Furthermore, as a specific example, after completing the intelligent hybrid beamforming method through steps S1-S5, in order to achieve system performance evaluation, the method of this embodiment also includes a process for calculating and evaluating the system transmission rate in high-speed mobile communication scenarios, specifically including: Step S6: The transmitter sends data symbol s, which is then processed by the optimal digital beamforming matrix F. D and the optimal simulated beamforming matrix F A After processing, the transmitting signal from the transmitting end is obtained. :

[0037] in, For the corresponding number in the optimal digital beamforming matrix A column vector of users, Is to the first Data stream transmitted by each user; Step S7, send signal After being transmitted through the channel to the receiving end, at the... Apply a digital merger to each user. Then, the final received signal is obtained. :

[0038] in, Indicates the first Millimeter-wave communication channel matrix for individual users It is the first The mean and variance of the corresponding users are zero. Additive white Gaussian noise, For the corresponding number in the optimal digital beamforming matrix A column vector of users, Is to the first Data stream transmitted by each user.

[0039] Based on the final received signal The first one can be calculated The signal-to-interference-plus-noise ratio (SIR) of each user is used to obtain the system speed, and the calculation formula is as follows:

[0040]

[0041]

[0042]

[0043] in, Indicates system and rate, Indicates the first The expected signal power of each user Indicates the first Multi-user interference power per user, Indicates the first Noise power per user.

[0044] At a frequency of 28GHz, the number of transmitting antennas at the base station For 64, the number of user-end receiving antennas For 4 wires, number of RF links The number of users is 5. Under the experimental condition of 3, after the model training converges, the sum rate performance of different hybrid beamforming schemes as a function of signal-to-noise ratio is shown in the following curves: Figure 3 As shown. From Figure 3 It can be seen that in the low signal-to-noise ratio (SNR) region, the performance of various methods is mainly limited by noise, and the convolutional neural network-based scheme proposed in this invention has shown certain performance advantages. As the SNR gradually increases, inter-user interference gradually replaces noise as the dominant factor affecting system performance, and the performance advantage of the proposed method becomes increasingly significant. Specifically, benchmark methods such as Orthogonal Matching Pursuit (OMP), Discrete Fourier Transform (DFT-BS), Multi-Objective Alternating Minimization (MO-AltMin), Multi-Objective Optimization (MO), and Random Hybrid Beamforming (RHBF) show a gradual increase in performance at high SNR, while the method proposed in this invention maintains a steeper performance improvement trend, with its sum and rate performance significantly outperforming all the benchmark methods compared. This performance improvement is mainly due to the synergistic effect of the analog beamforming module based on the convolutional neural network and the digital beamforming module based on the zero-forcing algorithm. The former can achieve more accurate analog beam prediction from actual channel observations, while the latter further suppresses inter-user interference in high-speed dynamic environments, thus enabling the proposed scheme to achieve better system and rate performance under different SNR conditions.

[0045] In summary, the deep learning-based MIMO intelligent hybrid beamforming method described above has the following beneficial effects: 1. This invention improves upon traditional convolutional neural networks by using a dual-branch, multi-scale parallel convolutional neural network to predict simulated beamforming matrices. The convolutional neural network model of this invention includes an input layer, a first-layer feature extraction module, a second-layer feature extraction module, a fully connected mapping layer, and an output layer. It can better handle imperfect channel state information, avoid the complex iterative solution process in traditional mathematical optimization methods, and effectively reduce computational complexity and processing latency.

[0046] 2. This invention simulates channel estimation errors and noise interference in actual communication by superimposing independent noise interference terms, generating imperfect channel state information samples. Based on a pre-constructed DFT codebook, a greedy search algorithm is used to select the codeword combination that minimizes the total mean square error of the system from the candidate simulated beam set. The simulated beamforming matrix corresponding to this codeword combination is used as a label, making this invention more applicable to single-base station multi-user high-speed mobile communication scenarios under imperfect channel state information conditions. It can reduce the dependence on accurate channel state information, improve the robustness of the method, and enhance the performance of beamforming methods in complex environments.

[0047] 3. This invention employs a dual-branch, multi-scale parallel convolutional neural network to predict the simulated beamforming matrix and uses a zero-forcing algorithm to solve the digital beamforming matrix. This combines the fast mapping capability of deep learning methods with the ability of the zero-forcing algorithm to suppress multi-user interference, which is beneficial to improving the system and speed in multi-user scenarios, thereby enhancing the system's transmission performance and meeting the requirements of online real-time deployment at the base station.

[0048] 4. This invention simulates channel estimation errors and noise interference in actual communication by superimposing independent noise interference terms, generating imperfect channel state information samples. Then, it constructs a beamforming dataset containing the imperfect channel state information samples and their corresponding labels, resulting in a beamforming dataset directly addressing imperfect channel state information. Training is then performed based on this dataset. Furthermore, this invention uses the relative mean square error of the system and rate as the loss function, utilizes the training set to train the convolutional neural network model offline, updates the network parameters in real time, and uses the validation set to dynamically evaluate the network performance during training. This achieves supervised training for beamforming matrix learning, effectively improving the adaptability and robustness of hybrid beamforming in high-speed mobile communication scenarios.

Claims

1. A deep learning-based MIMO intelligent hybrid beamforming method, characterized in that, include: Step S1: In the downlink of a mobile communication scenario, equip the base station at the transmitting end with... One transmitting antenna, the user end is equipped with Root receiving antenna, and base station configuration The base station achieves this by transmitting a single data stream to each user via the root radio frequency link. Parallel transmission for each user; Step S2: Based on the actual channel characteristics of the mobile communication scenario, a MIMO millimeter-wave channel system is constructed. By superimposing independent noise interference terms, the channel estimation error and noise interference in actual communication are simulated, and imperfect channel state information samples are generated. A candidate simulated beam set is constructed based on a pre-built DFT codebook. A greedy search algorithm is used to select the codeword combination that minimizes the total mean square error of the system from the candidate simulated beam set. The simulated beamforming matrix corresponding to the codeword combination is used as a label to construct a beamforming dataset containing imperfect channel state information samples and their corresponding labels. The beamforming dataset is then divided into a training set and a validation set according to a preset ratio. Step S3: Construct a dual-branch, multi-scale parallel convolutional neural network model. Convert the imperfect channel state information into a three-channel input tensor consisting of the amplitude, real part, and imaginary part of the channel matrix, which serves as the input to the convolutional neural network model. Use the relative mean square error of the system and rate as the loss function, and train the convolutional neural network model offline using the training set. Update the network parameters in real time, and use the validation set to dynamically evaluate the network performance during the training process until the model training converges, thus obtaining a converged convolutional neural network model. Step S4: In online applications, the imperfect channel state information acquired in real-time under actual communication scenarios is converted into actual three-channel input tensors and input into the training and converged convolutional neural network model. The model outputs the optimal simulated beamforming matrix that satisfies the constant mode constraint of the simulated phase shifter. ; Step S5, in the optimal simulated beamforming matrix Based on this, the zero-forcing algorithm is used to suppress interference between multiple users, and the optimal digital beamforming matrix is ​​calculated. Finally, the optimal simulated beamforming matrix is... With the optimal digital beamforming matrix Collaborative efforts are needed to complete MIMO intelligent hybrid beamforming; The convolutional neural network model includes an input layer, a first-layer feature extraction module, a second-layer feature extraction module, a fully connected mapping layer, and an output layer. The input layer is used to receive dimensions of The three-channel input tensor is obtained by decomposing the imperfect channel state information and is composed of the magnitude, real part and imaginary part of the channel matrix; The first-layer feature extraction module adopts a dual-branch parallel structure, including two parallel convolutional layers, one for extracting spatial features and the other for extracting spatial features. The first convolutional layer and the layer used to extract global coupling features. The second convolutional layer of the first layer, where the feature extraction module is connected to the average pooling layer for feature compression; The second-layer feature extraction module adopts a dual-branch parallel structure, including two parallel convolutional layers, one for extracting angular domain features and the other for extracting angular domain features. The third convolutional layer and the layer used to extract global coupling features The fourth convolutional layer, the second feature extraction module compresses the features into a low-dimensional feature vector through an average pooling layer; The fully connected mapping layer is used to perform feature mapping through multiple fully connected layers. A Dropout regularization layer is set after each fully connected layer, and the LeakyReLU activation function is used to stabilize feature propagation. The output layer is used to output the optimal simulated beamforming matrix obtained from the prediction. ; Step S5 specifically includes: Step S5.1, based on the optimal simulated beamforming matrix F A Construct an equivalent channel matrix , ,in, This is the total channel matrix; Step S5.2: Obtain the unconstrained digital beamforming matrix using the zero-forcing algorithm. : ； Among them, superscript This indicates the conjugate transpose. The regularization coefficient is . It is the identity matrix; Step S5.3: Introduce a scaling factor that satisfies the total system transmit power constraint. The optimal digital beamforming matrix is ​​obtained. , Finally, by combining the optimal analog beamforming matrix and the optimal digital beamforming matrix, MIMO intelligent hybrid beamforming is completed.

2. The deep learning-based MIMO intelligent hybrid beamforming method according to claim 1, characterized in that, The formula for calculating the loss function is as follows: in, Represents the loss function. This indicates the total number of input samples each time the loss is calculated; Indicates the first The system and rate corresponding to the optimal simulated beamforming matrix predicted by the convolutional neural network for each input sample; Indicates the first The optimal system and rate corresponding to the dataset labels for each input sample.

3. The deep learning-based MIMO intelligent hybrid beamforming method according to claim 2, characterized in that, The method further includes: Step S6: The transmitter sends data symbol s, which is then processed by the optimal digital beamforming matrix F. D and the optimal simulated beamforming matrix F A After processing, the transmitting signal from the transmitting end is obtained. : in, For the corresponding number in the optimal digital beamforming matrix A column vector of users, Is to the first Data stream transmitted by each user; Step S7, send signal After being transmitted through the channel to the receiving end, at the... Apply a digital merger to each user. Then, the final received signal is obtained. : in, Indicates the first Millimeter-wave communication channel matrix for individual users It is the first The mean and variance of the corresponding users are zero. Additive white Gaussian noise, For the corresponding number in the optimal digital beamforming matrix A column vector of users, Is to the first Data stream transmitted by each user.

4. The deep learning-based MIMO intelligent hybrid beamforming method according to claim 3, characterized in that, The formulas for calculating the system and rate are: in, Indicates system and rate, Indicates the first The expected signal power of each user Indicates the first Multi-user interference power of individual users Indicates the first Noise power per user.

Images