RIS-assisted multi-user communication beam forming optimization method and system based on graph neural network

Abstract

The invention discloses an RIS-assisted multi-user communication beam forming optimization method and system based on a graph neural network, and belongs to the field of reconfigurable intelligent metasurface-assisted communication. According to the method, the scene that RIS assists in multi-user communication is considered, the graph neural network is used for optimizing active beam forming of the base station and passive beam forming of the RIS, and then the system transmission rate of a downlink is maximized. The graph neural network model comprises an RIS node and user nodes of which the number is the same as that of service users, the RIS node is responsible for obtaining passive beam forming for the RIS by utilizing graph learning based on channel state estimation information input by all the user nodes, and each user node corresponds to one user; and obtaining active beam forming of a base station corresponding to the user by utilizing graph learning based on channel state estimation information input by the corresponding user. According to the method, the graph neural network is used, perfect channel estimation is not needed, and the joint optimization problem of active beam forming and passive beam forming can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of reconfigurable intelligent surface (RIS)-assisted communication, and particularly to a method and system for optimizing beamforming in RIS-assisted multi-user communication based on graph neural networks. Background Art

[0002] The RIS consists of multiple low-cost passive reflectors and can intelligently adjust the incident signal through a controller, thereby reconfiguring the wireless propagation environment, which is beneficial to enhancing communication performance. Especially in occluded scenarios, the transmission rate of the communication system can be improved by placing the reconfigurable intelligent surface.

[0003] In recent years, RIS-assisted multi-user multi-input single-output (MU-MISO) systems have received extensive attention. Researchers have proposed to maximize the system transmission rate by optimizing the active beamforming of the base station and the passive beamforming of the RIS. However, most existing studies assume perfect channel state information and rarely consider imperfect channel state information. In addition, the complexity of traditional iterative optimization algorithms is relatively high. At present, there is a particular lack of technology for jointly optimizing active beamforming and passive beamforming with low complexity and high scalability for the problem of maximizing the transmission rate of RIS-assisted MU-MISO systems under the condition of considering imperfect channel state information. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a method and system for optimizing beamforming in RIS-assisted multi-user communication based on graph neural networks. For the RIS-assisted MU-MISO system, without considering perfect channel state information, by jointly optimizing the active beamforming of the base station and the passive beamforming of the RIS, a balance is achieved between complexity and performance. Considering uplink pilot transmission and downlink joint beamforming optimization, the downlink transmission rate of the RIS-assisted MU-MISO system can be maximized.

[0005] Technical Solution: To achieve the above object of the invention, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a method for optimizing beamforming in RIS-assisted multi-user communication based on graph neural networks, including the following steps:

[0007] Utilize the uplink-downlink channel reciprocity of the RIS-assisted multi-user multi-input single-output MU-MISO system to establish a pilot based on uplink transmission, and take the active beamforming of the base station and the passive beamforming of the RIS as optimization variables to optimize the problem with the goal of maximizing the downlink system rate;

[0008] Construct a graph neural network model including an initial layer, multiple update layers and an output layer; the graph neural network model includes a RIS node and user nodes equal in number to the number of served users. The RIS node is responsible for obtaining the passive beamforming for the RIS through graph learning based on the channel state estimation information input by all user nodes. Each user node corresponds to a user, and the active beamforming for the base station corresponding to the user is obtained through graph learning based on the channel state estimation information input by the corresponding user.

[0009] Define the loss function as the negative value of the system rate, and update the weight parameters of the graph neural network model.

[0010] In the execution phase, the base station uses the trained graph neural network model to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the acquired uplink pilots.

[0011] Furthermore, the active beamforming of the base station includes beamforming for all served users. The beamforming for each user k is represented by , where M is the number of antennas of the base station, represents a complex number. The active beamforming of the base station is represented by W = [w 1 , w 2 ,..., w K , where K is the number of all served users. The passive beamforming of the RIS includes the phase responses of all RIS elements, represented by θ = [θ 1 , θ 2 , …, θ N H , where N represents the number of reflection elements of the RIS, and the superscript H represents conjugate transpose. Among them, η n represents the amplitude response of the nth RIS reflection element, j represents the imaginary unit, represents the phase response of the nth RIS element.

[0012] Furthermore, in the initial layer of the graph neural network model, the input of user node k is used to initialize the feature of user node k, where represents the estimated value of , and represent the channels between the RIS and the base station, between the kth user and the RIS, and between the kth user and the base station respectively. diag(·) represents generating a diagonal matrix from a vector. M is the number of antennas of the base station, and N represents the number of reflection elements of the RIS. The input ​Extract the real part and the imaginary part, and splice the real part and the imaginary part. Take the element-wise average of the output of the initial feature extractor passing through the RIS node as the initial feature of the RIS node;

[0013] In the update layer of the graph neural network model, update the features of the RIS node and each user node. The feature of the RIS node is specifically updated as follows: Take the average of the features of all user nodes in the previous update layer and splice it with the node of the RIS node in the previous update layer. Use the spliced result as the input of the aggregator of the RIS node in the current update layer, and splice the output of the aggregator with the feature of the RIS node in the previous update layer as the node feature of the RIS node in the current update layer; The feature of the k-th user node is specifically updated as follows: Extract the maximum value of the features of the user nodes other than the k-th user node and the features of the RIS node. Splice the extracted maximum feature, the feature of the k-th user node in the previous update layer, and the feature of the RIS node in the previous update layer. Use the spliced result as the input of the aggregator of the k-th user node in the current update layer, and splice the output of the aggregator with the feature of the k-th user node in the previous update layer to obtain the feature of the k-th user node in the current update layer;

[0014] In the output layer of the graph neural network model, convert the feature of the RIS node in the last update layer into the passive beamforming of the RIS, and convert the feature of the user node k in the last update layer into the active beamforming of the base station for user k.

[0015] Furthermore, the optimization problem is expressed as:

[0016]

[0017] C2:η n = 1, n ∈ {1, 2,..., N}

[0018]

[0019] where indicates that the input of the optimization problem is The optimization variables are (W, θ), g represents the function to be solved, represents the estimated value of H obtained by channel estimation k The estimated value of, and respectively represent the channels between the RIS and the base station, between the k-th user and the RIS, and between the k-th user and the base station, is Gaussian noise, P d represents the maximum downlink transmission power of the base station, ‖ ‖ 2 represents the 2-norm of the vector, | | represents taking the modulus of the complex number, and diag(·) represents generating a diagonal matrix from a vector.

[0020] Furthermore, the initial feature of the k-th user node is expressed as:

[0021]

[0022] where is the initial feature extractor of the user node, q represents the feature dimension, and the superscript T represents transpose, represents a real number, vec represents vectorization, represents taking the real part, represents taking the imaginary part;

[0023] The initial feature of the RIS node is expressed as:

[0024]

[0025] where is the initial feature extractor of the RIS node, represents the m-th row of, represents the element-wise mean function.

[0026] Furthermore, the update method of the RIS node in the d-th update layer is as follows:

[0027]

[0028] where represents the feature of the RIS node in the d-th update layer, represents the feature of the k-th user node in the (d - 1)-th update layer, is the aggregator of the RIS node in the d-th update layer, q represents the feature dimension, represents the element-wise mean function;

[0029] The feature of the k-th user node is updated in the following way:

[0030]

[0031] where is the aggregator of the k-th user node in the d-th update layer, represents the maximum function.

[0032] Furthermore, there is an output layer after D update layers, and the output layer is expressed as follows:

[0033]

[0034] where is the feature updater of the RIS node in the output layer, and the feature of the RIS node in the D-th update layer Via the feature updater Convert to the features of the RIS node in the output layer The passive beamforming of the RIS is θ = [θ 1 , θ 2 , …, θ N H Obtained from the features of the output layer in the following way Obtained:

[0035]

[0036] where j represents the imaginary unit respectively represent the nth and (N + n)th elements of the features of the output layer ;

[0037] For the features of each user node k in the update layer D Via the output layer feature updater of user node k Convert to the features of user node k in the output layer Expressed as follows:

[0038]

[0039] The beamforming vector of the base station for user k is w k = [w k,m m＝1,…,M Given, and its derivation is as follows:

[0040]

[0041] where respectively represent the mth and (m + M)th elements of the features of user node k in the output layer ;

[0042] In a second aspect, the present invention provides a RIS-assisted multi-user communication beamforming optimization system based on a graph neural network, including:

[0043] A problem construction module, configured to utilize the uplink-downlink channel reciprocity of the RIS-assisted multi-user multi-output single-input MU-MISO system to establish a pilot based on uplink transmission, and an optimization problem with the active beamforming of the base station and the passive beamforming of the RIS as optimization variables and the maximization of the downlink system rate as the objective;

[0044] ​​The graph neural network model construction module is used to construct a graph neural network model including an initial layer, multiple update layers, and an output layer; the graph neural network model includes a RIS node and user nodes with the same number as the number of served users. The RIS node is responsible for obtaining the passive beamforming for the RIS through graph learning based on the channel state estimation information input by all user nodes. Each user node corresponds to a user, and the active beamforming for the base station corresponding to the user is obtained through graph learning based on the channel state estimation information input by the corresponding user.

[0045] The network training module is used to define the loss function as the negative value of the system rate and update the weight parameters of the graph neural network model.

[0046] And the beamforming optimization module is used to, in the execution phase, the base station uses the trained graph neural network model to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the acquired uplink pilots.

[0047] In a third aspect, the present invention provides a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on a graph neural network are implemented.

[0048] In a fourth aspect, the present invention provides a computer program product, including a computer program. When the computer program is executed by the processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on a graph neural network are implemented.

[0049] Beneficial effects: The RIS-assisted multi-user communication beamforming optimization method based on a graph neural network proposed by the present invention has the following beneficial effects: (1) In the RIS-assisted MU-MISO system, using a graph neural network to jointly optimize the active beamforming of the base station and the passive beamforming of the RIS can maximize the system downlink transmission rate. (2) In the RIS-assisted MU-MISO system, a reconfigurable intelligent surface-assisted multi-user communication beamforming optimization method based on a graph neural network proposed by the present invention can maximize the system downlink transmission rate without relying on perfect channel state information. Description of the Drawings

[0050] Figure 1 The flowchart of the method in the embodiment of the present invention;

[0051] Figure 2 The schematic diagram of the RIS-assisted MU-MISO system model provided according to the embodiment of the present invention;

[0052] Figure 3 The schematic diagram of the architecture of the graph neural network model provided according to the embodiment of the present invention. Detailed implementation manners

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments described with reference to the accompanying drawings are exemplary and are intended to explain the present invention, but should not be construed as a limitation to the present invention.

[0054] Figure 1 It is a flowchart of a RIS-assisted multi-user communication beamforming optimization method based on a graph neural network according to an embodiment of the present invention.

[0055] As Figure 1 shown, the RIS-assisted multi-user communication beamforming optimization method based on a graph neural network includes the following steps:

[0056] Step 1: Utilize the uplink-downlink channel reciprocity of the RIS-assisted multi-user multiple-output single-input (MU-MISO) system to establish a pilot based on uplink transmission, and take the active beamforming of the base station and the passive beamforming of the RIS as optimization variables to optimize the problem with the maximization of the downlink system rate as the objective;

[0057] Step 2: Construct a graph neural network model including an initial layer, multiple update layers, and an output layer; the graph neural network model includes a RIS node and user nodes with the same number as the number of served users. The RIS node is responsible for obtaining the passive beamforming for the RIS through graph learning based on the channel state estimation information input by all user nodes. Each user node corresponds to a user, and the active beamforming of the base station corresponding to the user is obtained through graph learning based on the channel state estimation information input by the corresponding user;

[0058] Step 3: Define the loss function as the negative value of the system rate, and update the weight parameters of the graph neural network model;

[0059] Step 4: In the execution phase, the base station uses the trained graph neural network model to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the obtained uplink pilot.

[0060] Specifically, in Step 1, a RIS-assisted multi-user multiple-input single-output (MU-MISO) system model is established. The system includes a base station equipped with M antennas and a RIS with N reflection units, and simultaneously serves K single-antenna users. The system model is as Figure 2 shown.

[0061] In this embodiment, the time-division duplex protocol is adopted, and the uplink-downlink channel reciprocity of the RIS-assisted MU-MISO system is utilized to construct an uplink pilot transmission model and a downlink system rate optimization model. A pilot based on uplink transmission is established, and an optimization problem with the active beamforming of the base station and the passive beamforming of the RIS as optimization variables and the maximization of the downlink system rate as the objective is specifically included in the following steps:

[0062] Step 1-1, the set of K single-antenna users is represented by , and the channels between the RIS and the base station, between the k-th user and the RIS, and between the k-th user and the base station are represented by and respectively. The RIS phase shift matrix is represented by , where θ = [θ 1 , θ 2 , …, θ N H represents the phase responses of all RIS reflection elements, N represents the number of RIS reflection elements, the superscript H represents the conjugate transpose, where η n represents the amplitude response of the n-th RIS reflection element, and j represents the imaginary unit, represents the phase response of the n-th RIS element.

[0063] Step 1-2, construct an uplink pilot transmission model. The channel from user k to the base station can be expressed as:

[0064]

[0065] where Therefore, estimating all channel state information is equivalent to estimating the matrix The RIS generates P different reflection patterns through P different phase shift matrices, and P ≥ N + 1. The P different phase shift matrices are represented as Φ = [Θ 1 , Θ 2 , …, Θ P T , where Θ P is the p-th phase shift matrix and p ∈ {1, 2, …, P}; for user k, a pilot sequence u k = [u k,1 , u k,2 , …, u k,L T is adopted, and L ≥ K. To facilitate the distinction of different users, every two pilot sequences should be orthogonal, that is and where P t ​​​is the power of each pilot symbol. Considering that a pilot sequence is assigned to each RIS reflection mode, the pilot sequence of user k in the p-th reflection mode is denoted as u p,k =[u p,k,1 ,u p,k,2 ,…,u p,k,L T , so, in the p-th subframe of the base station, the received signal vector of the l-th pilot (1 ≤ l ≤ L) is expressed as:

[0066]

[0067] where is the l-th sampling noise vector of the base station in the p subframes, and is the noise power of each antenna on the base station, I is the identity matrix. Stacking the L pilot signal vectors received by the base station in the p-th subframe into matrix form, we can get:

[0068]

[0069] where Using the orthogonality of different user pilot sequences, multiplying u p,k by S p to separate the received signal of the k-th user, that is:

[0070]

[0071] where is the signal vector received by the base station from the k-th user in the p-th subframe, and is the variance of each element in v p,k . Therefore, after the base station receives P subframes, we can get:

[0072] R k =H k P+V k

[0073] where To estimate In the case of no prior channel knowledge, channel estimation can be performed by applying the Least Squares (LS) method and the Linear Minimum Mean Square Error (LMMSE) method. In this embodiment, the LS method is taken as an example for illustration, and we can get where represents the estimated value of H k obtained by using the LS estimator, while ​is the pseudo-inverse of P;

[0074] Steps 1 - 3, considering time-division duplexing and utilizing the reciprocity of the uplink and downlink, the downlink transmission rate between the base station and the k-th user is:

[0075]

[0076] where is Gaussian noise, represents the active beamforming of the base station towards the k-th user. Furthermore, the system transmission rate of the downlink of the RIS-assisted MU-MISO system is defined as:

[0077]

[0078] Step 1 - 4, construct an optimization problem to maximize the system rate of the downlink of the MU-MISO system by utilizing to optimize the active and passive beamforming of the downlink, that is:

[0079]

[0080] C2: η n = 1, n ∈ {1, 2,..., N}

[0081]

[0082] where W = [w 1 , w 2 ,..., w K represents the active beamforming of the base station towards K users, θ represents the passive beamforming of the RIS, represents that the input of the optimization problem is the optimization variables are (W, θ), g represents the function to be solved, which can be mapped to output (W, θ) through the input of to maximize the system rate of the downlink of the system. C1 represents the constraint that the active beamforming of the base station does not exceed the maximum downlink transmission power P d of the base station, C2 represents that the amplitude response of each unit of the RIS is a unit response, and C3 represents that the phase response of each unit of the RIS is between 0 and 2π.

[0083] In step 2 of this embodiment, a graph neural network model with K + 1 nodes including an initial layer, D update layers, and an output layer is constructed. The graph neural network model is as Figure 3 shown.

[0084] Specifically, in this embodiment, a model based on a graph neural network is constructed, which includes one RIS node and K user nodes. The RIS node is responsible for obtaining the passive beamforming for the RIS by using graph learning based on the channel state estimation information input by all user nodes. Each user node corresponds to a user, and the base station beamforming strategy for its specific user is obtained by using graph learning based on the channel state estimation information input by the corresponding user. The proposed GNN model consists of an initial layer, D update layers, and an output layer.

[0085] The initial layer is to use the input of user node k to initialize the feature of user node k and define the initial feature of the k-th user node which is expressed as:

[0086]

[0087] where is the initial feature extractor of the user node. Exemplarily, a three-layer fully connected neural network can be adopted, with the number of neurons in the input layer being 2M(N + 1), the number of neurons in the hidden layer being 128, and the number of neurons in the output layer being q, where q is a preset parameter. The real part and the imaginary part of the input of each user node are extracted, and the real part and the imaginary part are concatenated. After taking the element-wise average of the output of the initial feature extractor of the RIS node, it is used as the initial feature of the RIS node. The initial feature of the RIS node is expressed as:

[0088]

[0089] where is the initial feature extractor of the RIS node. Exemplarily, a three-layer fully connected neural network can be adopted, with the number of neurons in the input layer being 2(N + 1), the number of neurons in the hidden layer being 64, and the number of neurons in the output layer being q, is the m-th row of

[0090] In the update layer of the graph neural network model, the features of the RIS nodes and each user node are updated. The features of the RIS nodes are specifically updated as follows: take the average of the features of all user nodes in the previous update layer and concatenate them with the features of the RIS nodes in the previous update layer. The concatenated result is used as the input to the aggregator of the RIS nodes in the current update layer. The output of the aggregator is concatenated with the features of the RIS nodes in the previous update layer to obtain the node features of the RIS nodes in the current update layer. The features of the k-th user node are specifically updated as follows: extract the maximum value of the features of the user nodes other than the k-th user node and the features of the RIS nodes. Concatenate the extracted maximum feature, the features of the k-th user node in the previous update layer, and the features of the RIS nodes in the previous update layer. The concatenated result is used as the input to the aggregator of the k-th user node in the current update layer. The output of the aggregator is concatenated with the features of the k-th user node in the previous update layer to obtain the features of the k-th user node in the current update layer.

[0091] Exemplarily, to consider that the proposed graph neural network model contains D update layers, the update method of the RIS nodes in the d-th update layer is defined as follows:

[0092]

[0093] where represents the features of the RIS nodes in the d-th update layer, represents the features of the k-th user node in the d-1-th update layer, is the aggregator of the RIS nodes in the d-th update layer. A three-layer fully connected neural network can be used, with the number of neurons in the input layer being 2qd, the number of neurons in the hidden layer being q, and the number of neurons in the output layer being q, to combine the information of the user nodes and the RIS nodes. The RIS nodes receive features from K user nodes through the element-wise mean function . To retain the previous information, the RIS nodes will concatenate the features from the previous layer .

[0094] The features of the k-th user node are updated as follows:

[0095]

[0096] where is the aggregator of the k-th user node in the d-th update layer. A three-layer fully connected neural network can be used, with the number of neurons in the input layer being 2qd, the number of neurons in the hidden layer being q, and the number of neurons in the output layer being q. The maximum value function is used to extract the maximum value of the features of the user nodes other than the k-th user node and the features of the RIS nodes. The aggregator is used to concatenate the extracted maximum feature, the features of the k-th user node in the previous update layer Features of the previous updated layer RIS node After aggregation and extraction, it is combined with the features of the k-th user node in the previous updated layer to obtain the features of the k-th user node in the d-th updated layer

[0097] There is an output layer after D updated layers, which converts the features of the RIS node in the D-th updated layer into the passive beamforming θ of the RIS, and converts the features of the k-th user node in the D-th updated layer into the active beamforming of the base station for user k. The output layer is expressed as follows

[0098]

[0099] where is the feature updater of the output layer RIS node. A three-layer fully connected neural network can be used. The number of neurons in the input layer is q(D + 1), the number of neurons in the hidden layer is 128, and the number of neurons in the output layer is 2N. The features of the RIS node in the D-th updated layer are converted by the feature updater into the features of the RIS node in the output layer The passive beamforming θ of the RIS = [θ 1 , θ 2 , …, θ N H is obtained from the features of the output layer in the following way

[0100]

[0101] Furthermore, for the features of each user node k in the D-th updated layer are converted by the output layer feature updater of user node k into the features of user node k in the output layer which is expressed as follows

[0102]

[0103] A three-layer fully connected neural network can be used. The number of neurons in the input layer is q(D + 1), the number of neurons in the hidden layer is 128, and the number of neurons in the output layer is 2M. The beamforming vector of the base station for user k is given by w k = [w k,m m＝1,…,M and its derivation is as follows

[0104]

[0105] ​​In step 3, the graph neural network is trained in an unsupervised manner. The loss function is defined as the negative value of the system rate, that is, the system rate is maximized by minimizing the loss function, and the stochastic gradient descent method is used to update the weight parameters of the graph neural network.

[0106] In step 4 during the execution phase, the base station uses the trained graph neural network to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the obtained uplink pilots, which specifically includes the following steps:

[0107] Step 4-1: Load the trained graph neural network.

[0108] Step 4-2: The graph neural network outputs the passive beamforming of the RIS nodes and the active beamforming of the users according to the obtained channel state estimation information.

[0109] According to the RIS-assisted multi-user communication beamforming optimization method proposed in the embodiment of the present invention, for the RIS-assisted multi-user multiple-input single-output system, by constructing an uplink pilot transmission model and a downlink beamforming optimization model, pilots transmitted only through the uplink are constructed without perfect channel estimation information, the active beamforming of the base station and the passive beamforming of the RIS are optimized, and for the optimization problem of maximizing the system downlink transmission rate, a graph neural network is adopted. By constructing the initial layer, update layer, and output layer of the graph neural network, a graph neural network for jointly optimizing the active beamforming of the base station and the passive beamforming of the RIS is built. The loss function is designed as the negative value of the system downlink rate, and the graph neural network is trained using the gradient descent method. Using the trained graph neural network, the optimal active beamforming of the base station and the passive beamforming of the RIS are obtained according to the pilots transmitted through the uplink and by applying the least squares method. Through this method, the active beamforming of the base station and the passive beamforming of the RIS are optimized with low complexity, the system downlink transmission rate is maximized, and better communication performance is provided for users.

[0110] Based on the same inventive concept, an embodiment of the present invention provides a RIS-assisted multi-user communication beamforming optimization system based on a graph neural network, including: a problem construction module, which is used to utilize the uplink-downlink channel reciprocity of the RIS-assisted multi-user multiple-input single-output (MU-MISO) system to establish a pilot based on uplink transmission, and take the active beamforming of the base station and the passive beamforming of the RIS as optimization variables to establish an optimization problem with the maximization of the downlink system rate as the objective; a graph neural network model construction module, which is used to construct a graph neural network model including an initial layer, multiple update layers, and an output layer; in the graph neural network model, there is a RIS node and user nodes with the same number as the number of served users. The RIS node is responsible for obtaining the passive beamforming for the RIS through graph learning based on the channel state estimation information input by all user nodes. Each user node corresponds to a user, and the active beamforming of the base station corresponding to the user is obtained through graph learning based on the channel state estimation information input by the corresponding user; a network training module, which is used to define the loss function as the negative value of the system rate and update the weight parameters of the graph neural network model; and a beamforming optimization module, which is used in the execution stage for the base station to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the obtained uplink pilot by using the trained graph neural network model.

[0111] An embodiment of the present invention also provides a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on the graph neural network are implemented.

[0112] An embodiment of the present invention also provides a computer program product, including a computer program. When the computer program is executed by the processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on the graph neural network are implemented.

[0113] Any process or method description shown in the flowchart or described in other ways herein can be understood as representing code modules, segments, or portions including one or more executable instructions for implementing a customized logical function or process. The scope of the preferred embodiments of the present invention includes additional implementations, where functions can be executed in a substantially simultaneous manner or in an opposite order according to the functions involved, rather than in the order shown or discussed, which should be understood by those skilled in the technical field to which the embodiments of the present invention belong.

Claims

1. A RIS-assisted multi-user communication beamforming optimization method based on graph neural network, characterized in that: The following steps are involved: By using the uplink and downlink channel reciprocity of the RIS-assisted multi-user multi-output single-input MU-MISO system, an optimization problem based on uplink transmission pilot is established, with the active beamforming of the base station and the passive beamforming of the RIS as optimization variables and the goal of maximizing the downlink system rate. Constructing a graph neural network model including an initial layer, multiple update layers and an output layer; the graph neural network model includes a RIS node and user nodes equal to the number of service users, the RIS node is responsible for obtaining passive beamforming for RIS based on the channel state estimation information input by all user nodes by graph learning, each user node corresponds to a user, and the active beamforming of the base station corresponding to the user is obtained by graph learning based on the channel state estimation information input by the corresponding user; Define the loss function as the negative value of the system rate and update the weight parameters of the graph neural network model; In the execution phase, the base station uses the trained graph neural network model to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the acquired uplink pilot.

2. According to claim 1, the RIS-assisted multi-user communication beamforming optimization method based on graph neural network is characterized in that: The active beamforming of the base station includes the beamforming of the base station to all service users, and the beamforming of each user k is performed by Indicates that M is the number of antennas of the base station, represents a complex number, and the active beamforming of the base station is represented by W = [w1,w2,…,w K ], where K is the number of all service users, and the passive beamforming of the RIS includes the phase response of all RIS units, expressed as θ = [θ1, θ2, …, θ N ] H represents, N represents the number of reflection units of RIS, and the superscript H represents the conjugate transpose. Among them, η n represents the amplitude response of the nth RIS reflection unit, j represents the imaginary unit, θ n ∈[0,2π] represents the phase response of the nth RIS unit.

3. The RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to claim 1 is characterized in that: In the initial layer of the graph neural network model, the input of user node k is used Initialize the features of user node k, where express The estimated value of and denote the channels between RIS and the base station, between the kth user and RIS, and between the kth user and the base station, respectively. diag(·) denotes the generation of a diagonal matrix from a vector. M is the number of antennas in the base station, and N is the number of reflection units in the RIS. The input of each user node is Extract the real part and the imaginary part, concatenate the real part and the imaginary part, and average the output of the initial feature extractor of the RIS node element by element as the initial feature of the RIS node; In the update layer of the graph neural network model, the features of the RIS node and each user node are updated. The RIS node features are specifically updated as follows: the features of all user nodes in the previous update layer are averaged and then concatenated with the nodes of the RIS node in the previous update layer, and the concatenated result is used as the input of the aggregator of the RIS node in the current update layer, and the output of the aggregator is concatenated with the features of the RIS node in the previous update layer as the node features of the RIS node in the current update layer; the k-th user node features are specifically updated as follows: the maximum value of the features of the user nodes other than the k-th user node and the features of the RIS node is extracted, and the extracted maximum feature, the feature of the k-th user node in the previous update layer, and the feature of the RIS node in the previous update layer are concatenated, and the concatenated result is used as the input of the aggregator of the k-th user node in the current update layer, and the output of the aggregator is concatenated with the feature of the k-th user node in the previous update layer to obtain the feature of the k-th user node in the current update layer; In the output layer of the graph neural network model, the features of the RIS node in the last update layer are converted into the passive beamforming of RIS, and the features of the user node k in the last update layer are converted into the active beamforming of the base station to user k.

4. The RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to claim 2 is characterized in that: The optimization problem is expressed as: C2:h n =1,n∈{1,2,…,N} C3:θ n ∈[0,2π],n∈{1,2,…,N} in The input of the optimization problem is represented as The optimization variables are (W,θ), g represents the function to be solved, Denotes H obtained by channel estimation k The estimated value of and They represent the channels between RIS and the base station, between the kth user and RIS, and between the kth user and the base station, respectively. is Gaussian noise, P d represents the maximum downlink transmission power of the base station, ‖‖2 represents the 2-norm of the vector, || represents taking the modulus of the complex number, and diag(·) represents generating a diagonal matrix from the vector.

5. The RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to claim 3 is characterized in that: Initial features of the kth user node It is expressed as: in is the initial feature extractor of the user node, q represents the feature dimension, and the superscript T represents the transposition. represents a real number, vec represents vectorization, represents the real part, It means taking the imaginary part; Initial characteristics of RIS nodes It is expressed as: in is the initial feature extractor of the RIS node, express The mth row of represents the element-wise mean function.

6. The RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to claim 3 is characterized in that: The update method of RIS nodes in the dth update layer is as follows: in represents the characteristics of the RIS node in the dth update layer, represents the feature of the kth user node in the d-1th update layer, is the aggregator of the RIS nodes in the dth update layer, q represents the feature dimension, represents the element-wise mean function; The features of the kth user node are updated in the following way: in is the aggregator of the kth user node in the dth update layer, Represents the maximum value function.

7. The RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to claim 3 is characterized in that: There is an output layer after the D update layers, which is represented as follows: in It is the feature updater of the RIS node in the output layer. The RIS node updates the features of layer D. Feature Updater Converted to the features of RIS nodes in the output layer Passive beamforming of RIS θ=[θ1,θ2,…,θ N ] H The features of the output layer are obtained by get: Where j represents the imaginary unit, Represent the features of the output layer The nth and N+nth elements of ; For each user node k, update the features of layer D Output layer feature updater via user node k Converted to the features of user node k in the output layer It is expressed as follows: The beamforming vector of the base station for user k is w k =[w k,m ] m＝1,…,M Given, the derivation is as follows: in Represent the features of user node k in the output layer The mth and m+Mth elements of .

8. A RIS-assisted multi-user communication beamforming optimization system based on graph neural network, characterized in that: include: A problem construction module is used to utilize the uplink and downlink channel reciprocity of the RIS-assisted multi-user multi-output single-input MU-MISO system to establish an optimization problem based on uplink transmission pilots, with the active beamforming of the base station and the passive beamforming of the RIS as optimization variables and the goal of maximizing the downlink system rate; A graph neural network model construction module is used to construct a graph neural network model including an initial layer, multiple update layers and an output layer; the graph neural network model includes a RIS node and user nodes with the same number of service users, the RIS node is responsible for obtaining passive beamforming for RIS based on the channel state estimation information input by all user nodes by graph learning, each user node corresponds to a user, and the active beamforming of the base station corresponding to the user is obtained by graph learning based on the channel state estimation information input by the corresponding user; The network training module is used to define the loss function as the negative value of the system rate and update the weight parameters of the graph neural network model; and a beamforming optimization module, which is used in the execution phase for the base station to adjust the active beamforming of the base station and the passive beamforming of the RIS according to the acquired uplink pilot using the trained graph neural network model.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the RIS-assisted multi-user communication beamforming optimization method based on graph neural network according to any one of claims 1 to 7 are implemented.

Images