Network large-scale MIMO precoding method based on deep learning

Abstract

The invention discloses a network large-scale MIMO (Multiple Input Multiple Output) precoding method based on deep learning, which is suitable for downlink precoding design that a plurality of base stations cooperatively transmit data to a group of users in network large-scale MIMO transmission, and equivalently converts the precoding design for maximizing a system sum rate into a weighted minimum mean square error (WMMSE) problem. A WMMSE precoder is obtained through two-layer iteration, the WMMSE precoder is represented through low-dimensional parameters, and then a precoding design problem is converted into a low-dimensional variable optimization problem; a low-dimensional parameter optimization problem is solved by using deep learning, and a precoding vector is calculated through a closed expression. According to the method, the problem of high-dimensional precoding design is solved by solving the problem of low-dimensional parameters, and the performance close to the optimal sum rate is achieved with lower calculation complexity and higher calculation efficiency.

Description

Technical Field

[0001] The present invention relates to wireless communication downlink precoding, and in particular to a network large-scale MIMO precoding method and a deep learning design thereof. Background Art

[0002] Massive multiple-input-multiple-output (MIMO) is a technology that deploys large-scale antenna arrays on the base station (BS) side to provide efficient communication services to a large number of users. It can provide high data transmission rate, high efficiency and signal robustness to meet the high data demand, low latency and high reliability requirements of immersive communication.

[0003] In network massive MIMO transmission, using all base station antennas in the service cluster to transmit data collaboratively can obtain more available degrees of freedom, more effectively use the spectrum, and improve the reliability and robustness of transmission, especially for users at the edge of the cell. However, the use of massive antenna arrays greatly increases the number of iterations required for the iterative algorithm of network massive MIMO system precoding to converge, and the computational complexity of each iteration also increases dramatically, so it is necessary to further reduce the computational complexity. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a network large-scale MIMO precoding method based on deep learning to overcome the shortcomings of the prior art, achieve optimal sum rate performance, and reduce computational complexity.

[0005] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0006] A network massive MIMO precoding method based on deep learning, wherein the network massive MIMO transmission involves multiple base stations, multiple single-antenna users are randomly distributed in the coverage area, each base station is equipped with multiple antennas, and multiple base stations cooperatively transmit data for a group of users; when iteratively solving the WMMSE precoder, the optimization goal is to maximize the system sum rate: in and are the downlink channels and precoding vectors from all base stations to the kth user, is the channel vector from the lth base station to the kth user; is the precoding vector sent by the lth base station to the kth user, is the covariance matrix of the virtual interference plus noise of the kth user, σ is the noise power, the superscript H represents the matrix conjugate transpose, the superscript T represents the matrix transpose, L is the number of base stations, K is the number of users, and M is the number of antennas equipped for each base station;

[0007] In the precoding method, a WMMSE precoder that maximizes the system and rate is obtained by two-layer iteration, the WMMSE precoder is characterized by low-dimensional parameters, and then the precoding design problem is converted into an optimization problem of low-dimensional variables; deep learning is used to solve the low-dimensional parameter optimization problem, and the precoding vector is calculated through a closed-form expression.

[0008] Furthermore, the problem of maximizing the network massive MIMO system and rate under the power constraint of each base station is expressed as:

[0009]

[0010] Where P l is the maximum transmission power of the lth base station, D l is a block diagonal matrix whose l-th diagonal block is the M-order unit matrix I M , the remaining diagonal blocks are M-order all-zero arrays 0 M×M .

[0011] Furthermore, in the two-layer iteration, the precoding design problem of maximizing the system and rate is first converted into a convex optimization problem of weighted minimum mean-square-error minimization (WMMSE); in the outer iteration, the block coordinate descent (BCD) method is used to alternately optimize the receiver, weights and precoding vector according to the first-order optimality condition, and the equivalent problem is solved iteratively; in each step of the outer iteration, the BCD method is used for the quantities related to each base station to obtain the inner iteration process, and the precoding problem jointly designed by each base station is decoupled into a problem to be solved independently by each base station, and the dual variable corresponding to the power constraint of the base station is solved in the sub-problem.

[0012] Furthermore, in the outer iteration, the iteration expression is:

[0013]

[0014] Where:

[0015]

[0016] The superscripts (t) and (t+1) of the parameters represent the t-th and t+1-th iterations. and B (t) is the intermediate variable in the iterative calculation process, v k is the receiver of the kth user, w k is the weight, is the dual variable.

[0017] Furthermore, in the inner iteration, the iteration expression is:

[0018]

[0019] Where the superscript (t,q) represents the qth inner iteration in the tth outer iteration, and (t,q+1) represents the q+1th inner iteration in the tth outer iteration. for The vector extracted from For B (t) Submatrix of

[0020]

[0021] Find the optimal dual variable through binary search

[0022] Furthermore, the precoding vector expression corresponding to the stable point when the two-layer iterative algorithm converges only depends on the low-dimensional feature parameters and the channel vector, converting the precoding problem into an optimization problem about low-dimensional variables; the precoding method based on low-dimensional variables includes: using low-dimensional variables to calculate the direction of the precoding vector and calculate the allocated power; constructing the precoding vector using the direction and power of the precoding vector.

[0023] Furthermore, the optimization problem of low-dimensional variables is expressed as:

[0024]

[0025] Where λ = [λ1, ..., λ K ] T , μ=[μ1,...,μ L ] T ,λ k , and μ l , is the decision variable, Ξ is the intermediate variable, is the direction of the precoding vector of user k, ρ κ is the total power allocated to user k.

[0026] Furthermore, a neural network is used to solve optimization problems related to low-dimensional variables; the neural network consists of a convolutional layer, a fully connected layer and a normalization layer; specifically, the real and imaginary parts of the channel matrix are used as inputs of the two channels of the convolutional layer respectively; the features extracted by the convolutional layer are vectorized and used together with the signal-to-noise ratio as inputs of the fully connected layer; finally, the output of the fully connected layer is normalized as the output of the network.

[0027] Furthermore, neural network training includes an offline phase and an online phase;

[0028] The dataset is generated in the offline stage. For each sample in the dataset, a pair of channel matrix and signal-to-noise ratio is first generated, and the normalized low-dimensional parameters are calculated using the WMMSE precoder. in, After normalization l =[ l 1, ..., l K ] T and m =[ m 1, ..., m L ] T Combined with a channel matrix H and a signal-to-noise ratio γ, the data set is used to train a neural network;

[0029] In the online stage, the trained neural network is used to calculate the low-dimensional parameters, which are substituted into the structure of the WMMSE precoder to calculate the precoding vector. The steps include:

[0030] 1) Using low-dimensional variables to calculate the direction of the precoding vector

[0031]

[0032] 2) Calculate the allocated power

[0033]

[0034] 3) Construct the precoding vector using the direction and power of the precoding vector

[0035] A computer program product comprises a computer program / instruction, wherein when the computer program / instruction is executed by a processor, the steps of the deep learning-based network large-scale MIMO precoding method are implemented.

[0036] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0037] (1) Aiming at the network massive MIMO system and rate maximization problem, a two-layer iterative WMMSE precoding method is proposed. The expression of the precoding vector corresponding to its stable point shows that it can be reconstructed by low-dimensional parameters, so the precoding design problem is reformulated as an optimization problem about low-dimensional parameters.

[0038] (2) A neural network design method is proposed to solve the low-dimensional parameter optimization problem using deep learning. The complexity of network training is effectively reduced by performing equivalent transformations on the learning objectives and network inputs. The precoding vector can be directly calculated based on the output of the neural network, which significantly reduces the complexity of online calculations while maintaining performance close to that of iterative algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Precoding transmission scenario diagram for network massive MIMO;

[0040] Figure 2 Schematic diagram of the neural network architecture used in deep learning design for network massive MIMO precoding. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] like Figure 1 As shown, in the network large-scale MIMO precoding method based on deep learning disclosed in an embodiment of the present invention, multiple base stations are involved in the network large-scale MIMO transmission, and multiple single-antenna users are randomly distributed in the coverage area. Each base station is equipped with multiple antennas, and multiple base stations cooperate to transmit data for a group of users. When iteratively solving the WMMSE precoder, the optimization goal is to maximize the system and rate; in the precoding method, a two-layer iteration is used to obtain the WMMSE precoder that maximizes the system and rate, the WMMSE precoder is characterized by low-dimensional parameters, and then the precoding design problem is converted into an optimization problem of low-dimensional variables; deep learning is used to solve the low-dimensional parameter optimization problem, and the precoding vector is calculated through a closed-form expression.

[0043] In the two-layer iteration, the precoding design problem of maximizing the system and rate is first converted into a convex optimization problem of minimizing the weighted mean square error (WMMSE); in the outer iteration, the block coordinate descent method is used to alternately optimize the receiver, weights and precoding vector according to the first-order optimality condition to iteratively solve the equivalent problem; in each step of the outer iteration, the block coordinate descent method is used for the quantities related to each base station to obtain the inner iteration process, and the precoding problem jointly designed by each base station is decoupled into a problem that each base station solves independently, and the dual variable corresponding to the power constraint of the base station is solved in the sub-problem.

[0044] The precoding vector expression corresponding to the stable point when the two-layer iterative algorithm converges only depends on the low-dimensional feature parameters and the channel vector, and the precoding problem is converted into an optimization problem about low-dimensional variables; the precoding method based on low-dimensional variables includes: using low-dimensional variables to calculate the direction of the precoding vector and calculate the allocated power; using the direction and power of the precoding vector to construct the precoding vector.

[0045] A neural network is used to solve optimization problems related to low-dimensional variables. The neural network consists of a convolutional layer, a fully connected layer, and a normalization layer. Specifically, the real and imaginary parts of the channel matrix are used as inputs of the two channels of the convolutional layer respectively; the features extracted by the convolutional layer are vectorized and used together with the signal-to-noise ratio as inputs of the fully connected layer; finally, the output of the fully connected layer is normalized as the output of the network.

[0046] The following is a detailed description of the embodiments of the present invention in combination with a specific system model, problem statement and conversion, and deep learning design. In this embodiment, L base stations are provided to cooperatively transmit signals for K users within the coverage area, each base station is equipped with M antennas, and the user is equipped with a single antenna.

[0047] 1) System model and problem statement

[0048] The downlink received signal y of the kth user k for:

[0049]

[0050] In the formula, is the channel vector from the lth base station to the kth user; is the precoding vector sent by the lth base station to the uth user, satisfying the power constraint P l is the maximum transmit power of the lth base station, the superscript H represents the matrix conjugate transpose; d k is the data stream of the kth user, d u is the data stream of the u-th user; n k is noise, and the downlink channels from all base stations to the kth user and the precoding vectors are concatenated as and The downlink precoding vectors from all base stations to the u-th user are concatenated as follows:

[0051] Assuming that the user side knows the accurate CSI, the sum of the received interference and noise is Consider the covariance as Gaussian noise, the sum rate of the system is:

[0052]

[0053] The problem of maximizing the network massive MIMO system sum rate under the power constraint of each base station is expressed as:

[0054]

[0055] in, is a block diagonal matrix whose l-th diagonal block is the M-order unit matrix I M, the remaining diagonal blocks are M-order all-zero arrays 0 M×M .

[0056] 2) Iterative WMMSE precoder

[0057] The receiver on the user side is denoted as v k , then the mean square error (MSE) of the received signal can be expressed as

[0058]

[0059] in, For a given precoder, the optimal receiver for the MSE minimization problem is

[0060]

[0061] in At this time, the minimum mean squared error (MMSE, minimum mean-squared-error) is obtained

[0062]

[0063] Problem P1 can be equivalently transformed into the following weighted MSE minimization problem

[0064]

[0065] in v=[v1,…,v K ] and w=[w1,…,w k ] represent the precoder, receiver and weight of all users respectively.

[0066] The objective function and constraints of problem P2 are convex with respect to P, v, and w, and can be solved by iteratively updating these variables in sequence using the BCD method.

[0067] For the tth iteration, according to the first-order optimality condition, the weight w k The calculation formula is

[0068]

[0069] The optimal receiver is the MMSE problem Solution:

[0070]

[0071] Substituting equations (8) and (9) into problem P2, the convex optimization problem for the precoder in the tth iteration becomes

[0072]

[0073] The Lagrangian function of problem P3 is

[0074]

[0075] Among them, {μ l ≥0} is the dual variable corresponding to the power constraint, B (t) , They are

[0076]

[0077]

[0078] about The first-order optimal condition is

[0079]

[0080] Combining equation (16) with equation (8) and equation (9), we get the iterative expression of the precoding vector:

[0081]

[0082] To determine the dual variable The problem needs to be further decomposed into units of each BS and searched one by one. (t) and w (t) , problem P3 can be decomposed into L independent problems:

[0083]

[0084] Problem P4 can be solved by decoupling the optimization for each base station using the BCD method.

[0085] According to the variable p k，l The first-order condition of

[0086]

[0087] in for The vector extracted from For B (t) Submatrix of

[0088]

[0089] The iterative formula of the precoding vector is

[0090]

[0091] The superscript (t,q) represents the qth inner iteration in the tth outer iteration. It can be found through binary search.

[0092] In summary, the WMMSE precoding method can be summarized as follows:

[0093] Step 1: Initialize the precoding vector and dual variables

[0094] Step 2: Calculate the receiver according to equation (9)

[0095] Step 3: Calculate the weight according to formula (8)

[0096] Step 4: Calculate according to formula (21) and

[0097] Step 5: Repeat steps 2-4 until the inner iteration converges;

[0098] Step 6: Calculate according to formula (17)

[0099] Step 5: Repeat steps 2-6 until the outer iteration converges.

[0100] The solution obtained by the above steps is called WMMSE precoder.

[0101] 3) Equivalent optimization problem of low-dimensional optimization variables

[0102] When the above algorithm converges, we can get At this time, formula (17) can be written as

[0103]

[0104] in, are all scalars. According to formula (22), we know In the case of Completely by α k , k , and μ l , Determine, where α k Determine its energy, λ k , and μ l , Decide on a direction.

[0105] Therefore, the precoder design problem can be transformed into the following optimization problem about low-dimensional variables:

[0106]

[0107] Where α=[α1,...,α k ] T ,λ=[λ1,...,λ K ] T , μ=[μ1,...,μ L ] T .

[0108] According to formula (22), we can get

[0109]

[0110] in, remember according to Computable

[0111]

[0112] Therefore, the total power allocated to the kth user is

[0113]

[0114] Therefore, only λ and μ are needed to find the precoding vector. The optimization problem of low-dimensional variables can be further equivalent to

[0115]

[0116] The precoding method based on low-dimensional variables λ and μ can be summarized as follows:

[0117] Step 1: Calculate the normalized precoding vector according to equations (27a)-(27b)

[0118] Step 2: Calculate the allocated power ρ according to (27d) k ,

[0119] Step 3: Construct the precoding vector according to (27e)

[0120] Construct the precoding vector above through λ and μ The steps are recorded as functions

[0121] P=f(λ,μ;H) (28)

[0122] in,

[0123] 4) Deep Learning Design

[0124] According to formula (28), the precoding vector can be determined by low-dimensional parameters λ and μ. Therefore, a neural network can be used to learn λ and μ obtained by the WMMSE precoder to solve problem P6.

[0125] In order to eliminate the fluctuations in the order of magnitude of λ and μ, consider scaling λ and μ proportionally, which will lead to the inverse scaling of the function value P, that is,

[0126]

[0127] Therefore, λ k , and μ l , Normalized to l k =τλ k , m l =τμ l ,in The precoding vector and the normalized parameters satisfy the following relationship:

[0128]

[0129] According to equation (33), the power allocated to all users is scaled proportionally by the parameter τ, so the value of τ can be obtained by summing equation (33) for all users No additional learning is required. l =[ l 1, ..., l K ] T and m =[ m 1, ..., m L ] T Only depends on the power constraint value P l and the noise variance σ 2 The ratio γ = [P1 / σ 2 , ..., P L / σ 2 ] T , without knowing their specific values. Therefore, γ can be used as the input of the neural network.

[0130] like Figure 2As shown in the figure, a convolutional neural network and a fully connected layer are used to construct a neural network. The real and imaginary parts of the channel matrix H are used as two input channels respectively. The channel features are first extracted through the convolution layer, then vectorized and concatenated with the vector γ and input into the fully connected layer. Each convolution layer uses maximum pooling and a rectified linear unit (ReLU) activation function.

[0131] The structure of the neural network can be expressed as:

[0132]

[0133] z (i) =Conv(z (i-1) ), i=1, ..., L C (35)

[0134] o (0) =[z (i) ,γ] (36)

[0135] o (i) =FC(o (i-1 )), i=1, ..., L F (37) Among them, L C With L F They represent the depth of the convolutional layer and the fully connected layer, respectively. Conv(·) represents a layer of convolutional calculation, FC(·) represents a layer of fully connected calculation, and z (i) With o (i) Represent the i-th layer output of the convolutional layer and the fully connected layer respectively.

[0136] A normalization layer is used after the fully connected layer to eliminate the amplitude fluctuation of the output value of the neural network, that is,

[0137]

[0138] Among them, k Output of the last fully connected layer The kth element in O K+l for The K+lth element in .

[0139] The neural network structure of (34)-(39) can be expressed as a function

[0140]

[0141] in,

[0142] The deep learning design of network massive MIMO precoding is divided into two stages: offline and online.

[0143] The offline stage is used to generate data sets and train neural networks. When generating data sets, each time a pair of channel matrices and signal-to-noise ratios is generated, the WMMSE precoder is used to calculate low-dimensional parameters as corresponding labels, and then combined with the channel matrix and signal-to-noise ratio as a sample of the data set. The specific steps for generating offline data sets are as follows:

[0144] Step 1: Generate channel vector Noise variance and power constraints The superscript (n) indicates the nth sample;

[0145] Step 2: Obtain the WMMSE precoder through iterative calculation and the corresponding dual variable

[0146] Step 3: Calculation

[0147] Step 4: Calculate the normalized parameters and in

[0148] Step 5: and Combine into the nth sample;

[0149] Step 6: Repeat steps 1-5 until a sufficient number of samples are generated.

[0150] The dataset generated by the above steps is recorded as The goal of deep learning training is to minimize the following MSE cost function

[0151]

[0152] in, Indicates the size of the dataset, vector β (n) =[( l (n) ) T , ( m (n) ) T ] T , is the output of the neural network.

[0153] In the online stage, the channel matrix and signal-to-noise ratio are used as inputs to the trained neural network to obtain the predicted values ​​of the low-dimensional parameters and calculate the precoding vector. The specific steps are as follows:

[0154] Step 1: Calculate through neural network

[0155] Step 2: Calculate the direction of the precoding vector according to equations (30)-(31)

[0156] Step 3: Calculate the allocated power ρ according to equations (32)-(33) k ,

[0157] Step 4: Calculate the precoding vector

[0158] Complete the above steps 2 to 4 by and Construct precoding vector The steps are recorded as functions

[0159]

[0160] A computer program product disclosed in an embodiment of the present invention includes a computer program / instruction, which, when executed by a processor, implements the steps of the deep learning-based network large-scale MIMO precoding method.

[0161] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A network massive MIMO precoding method based on deep learning, characterized in that: The network large-scale MIMO transmission involves multiple base stations, and multiple single-antenna users are randomly distributed in the coverage area. Each base station is equipped with multiple antennas, and multiple base stations cooperate to transmit data for a group of users. When iteratively solving the WMMSE precoder, the optimization goal is to maximize the system sum rate: in and are the downlink channels and precoding vectors from all base stations to the kth user, is the channel vector from the lth base station to the kth user; is the precoding vector sent by the lth base station to the kth user, is the covariance matrix of the virtual interference plus noise of the kth user, σ is the noise power, the superscript H represents the matrix conjugate transpose, the superscript T represents the matrix transpose, L is the number of base stations, K is the number of users, and M is the number of antennas equipped for each base station; In the precoding method, a WMMSE precoder that maximizes the system and rate is obtained by two-layer iteration, the WMMSE precoder is characterized by low-dimensional parameters, and then the precoding design problem is converted into an optimization problem of low-dimensional variables; deep learning is used to solve the low-dimensional parameter optimization problem, and the precoding vector is calculated through a closed-form expression.

2. The network massive MIMO precoding method based on deep learning according to claim 1, characterized in that: The problem of maximizing the network massive MIMO system sum rate under the power constraint of each base station is expressed as: Where P l is the maximum transmission power of the lth base station, D l is a block diagonal matrix whose l-th diagonal block is the M-order unit matrix I M , the remaining diagonal blocks are M-order all-zero arrays 0 M×M .

3. The network massive MIMO precoding method based on deep learning according to claim 2, characterized in that: In the two-layer iteration, the precoding design problem of maximizing the system and rate is first converted into a convex optimization problem of minimizing the weighted mean square error WMMSE; in the outer iteration, the block coordinate descent method is used to alternately optimize the receiver, weights and precoding vector according to the first-order optimality condition, and the equivalent problem is solved iteratively; in each step of the outer iteration, the block coordinate descent method is used for the quantities related to each base station to obtain the inner iteration process, and the precoding problem jointly designed by each base station is decoupled into a problem to be solved independently by each base station, and the dual variable corresponding to the power constraint of the base station is solved in the sub-problem.

4. The network massive MIMO precoding method based on deep learning according to claim 3, characterized in that: In the outer iteration, the iteration expression is: Where: The superscripts (t) and (t+1) of the parameters represent the t-th and t+1-th iterations, and the superscript * indicates conjugation. and B (t) is the intermediate variable in the iterative calculation process, v k is the receiver of the kth user, w k is the weight, is the dual variable.

5. The network massive MIMO precoding method based on deep learning according to claim 4, characterized in that: In the inner iteration, the iteration expression is: Where the superscript (t,q) represents the qth inner iteration in the tth outer iteration, and (t,q+1) represents the q+1th inner iteration in the tth outer iteration. for The vector extracted from For B (t) Submatrix of Find the optimal dual variable through binary search 6. The network massive MIMO precoding method based on deep learning according to claim 1, characterized in that: The precoding vector expression corresponding to the stable point when the two-layer iterative algorithm converges only depends on the low-dimensional feature parameters and the channel vector, which converts the precoding problem into an optimization problem about low-dimensional variables. The precoding method according to the low-dimensional variable includes: calculating the direction of the precoding vector using the low-dimensional variable and calculating the allocated power; The precoding vector is constructed using the direction and power of the precoding vector.

7. The network massive MIMO precoding method based on deep learning according to claim 2, characterized in that: The optimization problem with low-dimensional variables can be expressed as: Where λ = [λ1, ..., λ K ] T , μ=[μ1,...,μ L ] T , and is the decision variable, Ξ is the intermediate variable, is the direction of the precoding vector of user k, ρ k is the total power allocated to user k, ||·|| represents the 2-norm of the vector, and |·| represents the absolute value.

8. The network massive MIMO precoding method based on deep learning according to claim 1, characterized in that: Use neural networks to solve optimization problems with low-dimensional variables; The neural network consists of a convolutional layer, a fully connected layer and a normalization layer; specifically, the real and imaginary parts of the channel matrix are used as the input of the two channels of the convolutional layer respectively; the features extracted by the convolutional layer are vectorized and used together with the signal-to-noise ratio as the input of the fully connected layer; finally, the output of the fully connected layer is normalized as the output of the network.

9. The network massive MIMO precoding method based on deep learning according to claim 7, characterized in that: Neural network training includes offline and online phases; The dataset is generated in the offline stage. For each sample in the dataset, a pair of channel matrix and signal-to-noise ratio is first generated, and the normalized low-dimensional parameters are calculated using the WMMSE precoder. in, After normalization λ =[ λ 1, ..., λ K ] T and μ =[ μ 1, ..., μ L ] T Combined with a channel matrix H and a signal-to-noise ratio γ, the data set is used to train a neural network; In the online stage, the trained neural network is used to calculate the low-dimensional parameters, which are substituted into the structure of the WMMSE precoder to calculate the precoding vector. The steps include: 1) Using low-dimensional variables to calculate the direction of the precoding vector 2) Calculate the allocated power 3) Construct the precoding vector using the direction and power of the precoding vector 10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the deep learning-based network large-scale MIMO precoding method according to any one of claims 1 to 9 are implemented.

Images