Multi-cell wireless network distributed resource optimization method based on graph neural network

Abstract

The invention discloses a multi-cell wireless network distributed resource optimization method based on a graph neural network, and the method comprises the steps: building a total rate and an optimization problem of all user terminal equipment according to a communication model of a multi-cell wireless network; a message passing graph neural network model containing edge feature embedding is designed, and various kinds of channel information are mapped to nodes and edges of the graph; and a graph neural network model is trained in a centralized manner, the optimization problem is maximized by using a parameterization scheme of the graph neural network, the trained graph neural network is deployed on each cell in a distributed manner, and multi-cell wireless network distributed resource optimization based on the graph neural network is realized.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to a method for optimizing distributed resources in a multi-cell wireless network based on a graph neural network. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] Since its inception, the cellular network architecture has gradually become the core architecture of commercial communications with its advantages of simple implementation and high-frequency band resource utilization. With the development of mobile communication technology, the growth of large-scale traffic has been coped with by technologies such as wider bandwidth, improved radio interface and network encryption. Considering different cost-effectiveness, the contribution of large-scale multiple input multiple output technology is particularly significant. As a key 5G technology, massive MIMO (Multiple Input Multiple Output) constitutes a centralized solution for encrypted networks. It provides extremely high beamforming gain by using large-scale antenna arrays to aim high-gain antenna beams at one or more users; in addition, highly directional beams can not only increase the received power level, but also reduce interference from adjacent users and transmission points in space, and achieve spatial reuse of scarce time / frequency resources through spatial multiplexing operations, thereby improving the spectrum efficiency and energy efficiency of the system.

[0004] Multi-cell cooperative communication refers to the collaboration between multiple cells in the 5G communication system to achieve higher network capacity and better user experience. In traditional wireless communications, different cells are often independent of each other, but in multi-cell cooperative communication, resources can be shared and optimized between cells, thereby improving communication efficiency and reliability. The centralized optimization solution currently widely used in multi-cell networks has high computational complexity in large-scale networks and cannot adapt to the dynamic changes in the number of cells or base stations and the number of users. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a multi-cell wireless network distributed resource optimization method based on graph neural network in view of the shortcomings of the prior art.

[0006] Graph neural network is a neural network model specially designed for processing graph structured data. It aims to learn the representation of nodes or graphs and perform complex graph data analysis and reasoning. Unlike traditional neural networks that mainly process vector or matrix data, graph neural networks can effectively capture the relationship and topological structure between nodes and are suitable for various complex graph data analysis tasks. Unlike conventional neural network algorithms, graph neural networks extend existing neural network methods to the field of non-Euclidean data, that is, the field of graph data processing. As a neural network model designed for graph data, graph neural networks have three advantages: strong expressiveness, good generalization, and distributed operation. They can meet the challenges of intelligent wireless network design in 5G.

[0007] In a graph neural network, each node represents an entity or object in the graph, and each edge represents the association between nodes. The goal of a graph neural network is to learn the representation of each node so that information can be propagated, features can be aggregated, and predicted across the entire graph structure. By learning the interactions and structural information between nodes, graph neural networks can achieve various tasks such as node classification, link prediction, and graph generation.

[0008] Graph neural networks usually include two main aspects: node representation learning and graph representation learning. Node representation learning aims to learn a low-dimensional representation of each node so that the node can retain its neighbor relationship and feature information in the graph structure in the representation space. Graph representation learning maps the entire graph structure into a low-dimensional representation space in order to analyze and infer the entire graph. Its core idea is to use the connection relationship and local structural information between nodes to propagate information and aggregate features through multi-layer neural networks to obtain a more global graph representation. The advantage of graph neural networks is that they can process various complex graph data structures, including data in social networks, recommendation systems, bioinformatics and other fields. By learning the interactions and structural information between nodes, graph neural networks can achieve more accurate and effective graph data analysis and mining.

[0009] In the field of wireless communications, the application of graph neural networks has broad prospects and potential. Graph neural networks can be used to model and analyze the topology of communication networks, realize the optimal configuration of network resources and the necessity and advantages of applying graph neural networks in communication networks; by learning the spectrum utilization of wireless channels through graph neural networks, intelligent spectrum perception and spectrum allocation can be realized; using graph neural networks to model and optimize the transmission, reception and demodulation of wireless signals can improve the performance and efficiency of communication systems; using graph neural networks to model and analyze users, devices and base stations in mobile networks can realize intelligent management and scheduling of network resources.

[0010] A method for optimizing distributed resources in a multi-cell wireless network based on a graph neural network comprises the following steps:

[0011] Step 1, taking maximizing the total transmission rate of user terminal equipment in multiple cells as the optimization goal, combined with the maximum power constraint of the cell base station, establish an optimization problem;

[0012] Step 2: Design a graph model as an input of the graph neural network for the optimization problem, where the graph model includes base station information of neighboring cells and base station and user information of the current cell;

[0013] Step 3, establish a message passing graph neural network learning framework including edge embedding;

[0014] Step 4: Taking the maximization of the sum of the system user transmission rates as the loss function, train and optimize the learning parameters of the graph neural network designed in step 3;

[0015] Step 5: Distributedly deploy graph neural networks on base stations in each cell to design beamforming matrices, and implement distributed resource optimization of multi-cell wireless networks based on graph neural networks.

[0016] In the present invention, uppercase bold letters are used to represent matrices, lowercase bold letters are used to represent vectors, and lowercase letters are used to represent scalars. i represents the i-th column of matrix H, H i，j represents the i-th column and j-th row element of matrix H. i represents the i-th element of vector h. E(·) represents the expectation, represents a complex field of dimension A*B, diag(A) represents a vector consisting of the diagonal elements of A, and I K represents the K-dimensional identity matrix, the superscripts T and H represent the transpose and conjugate transpose of the matrix respectively, the superscripts * and t represent the optimal solution and the t-th iteration value of the variable respectively. The backslash \ indicates exclusion from the set.

[0017] Further, step 1 includes: assuming that a network has L cells, each cell is equipped with a base station, and each base station serves N l (l≤L) users, where the base station in each cell transmits signals to the user terminal equipment through the downlink; each base station has N t antennas, and the user terminal equipment has a single antenna.

[0018] The data symbol s of all user terminal equipment is expressed as in The channel from the base station of cell l to user i in the cell is Represents, where i≤N l , the channel from the base station of cell j to user i outside the cell is indicates; w lirepresents the beam vector when the base station of cell l sends a signal to the terminal device of user i in cell l, then the signal y received by the terminal device of user i in cell l is li It can be expressed as:

[0019]

[0020] SINR received by the terminal device of user i in cell l li It is expressed as:

[0021]

[0022] in represents the additive Gaussian white noise of the ith user in the lth cell, represents the variance of the noise.

[0023] The overall optimization problem of this network model is specifically manifested as the total rate and maximization problem of all users:

[0024]

[0025] Where st represents the set of constraints on the optimization problem. is the maximum transmit power that the base station of cell l can use, SUNR li is the signal-to-interference-noise ratio received by the terminal equipment of user i in cell l, is the beamforming matrix of the base station in cell l.

[0026] Step 2 includes: using graph neural network to model the original distributed problem as follows:

[0027]

[0028] st(N TX ,N RX,l ,E l )=φ(X TX ,X RX,l ,Y l )

[0029] The heterogeneous graph model G is established based on the communication topology model of cell l l = {N TX , N RX，l , E l}, where N TX = {N l |l≤L} represents the transmitting nodes of the entire multi-cell communication system, including base stations within the cell and base stations outside the cell, N TX =[n TX，1 , ..., n TX，L ] TRepresents the variable on the TX node, Represents the feature matrix of the TX node; N RX,l = {N i |i≤N l ,l≤L} represents the receiving node in cell l, that is, the user node, M RX，l =[n RX，l，1 , ..., n RX，l，L ] T Represents the variable on the RX node, Represents the feature matrix of the RX node; E l represents the edges between the nodes of the graph represented by the current cell l, represents the variable on the edge, Represents the edge features of each edge; d TX d RX and d E They represent feature dimensions respectively, where φ is the mapping function from features to variables.

[0030] Step 3 includes: building a graph neural network model with the following structure:

[0031] The graph neural network model consists of three parts: initialization module, node update module and normalization module. The initialization module first converts the complex features (X TX , X RX，l , Y l ) is converted to real features

[0032]

[0033] in represents the real part, represents the imaginary part; the initial representations of TX node features, RX node features, and edge features are transformed by a single-layer MLP with a rectified linear unit as the activation function:

[0034]

[0035] in are learnable network parameters;

[0036] The node update module consists of T message passing layers. Each message passing layer includes an aggregation layer and a combination layer. The aggregation layer collects the features of adjacent nodes and edge features. The combination layer combines the features obtained by the aggregation layer and the node features to obtain and update the current node features.

[0037] In the node update module, the update strategy for the TX node features in the t-th layer of the neural network is:

[0038]

[0039] in yes The mth row of yes The kth row of represents the set of neighboring RX nodes of the mth TX node, It is the value of the aggregation layer output after aggregating adjacent RX nodes and edge features. AGG represents the aggregation function. are two different MLPs.

[0040] The update strategy for the RX node features in the t-th layer of the neural network is:

[0041]

[0042] in represents the set of neighboring RX nodes of the mth TX node, It is the value of the aggregation layer output after aggregating adjacent TX nodes and edge features. AGG represents the aggregation function. are two different MLPs.

[0043] The update strategy for edge features in the t-th layer of the neural network is:

[0044]

[0045] in yes The (m, k)th element of It is the value of the node features of the adjacent TX nodes and RX nodes output by the aggregation layer. They are three MLPs with different parameters.

[0046] In the normalization module, the features of each node and edge in the graph are first represented by a single-layer MLP. Mapping to output variables The single-layer MLP uses the rectified linear unit as the activation function, specifically:

[0047]

[0048] is a trainable parameter;

[0049] Then we need to output the variable Convert to the required complex variable Specifically:

[0050]

[0051] Finally, the output variable Normalization is performed so that its power meets the constraints, and finally the beamforming matrix of the current cell base station is obtained as follows:

[0052]

[0053] in is the power of the signal transmitted by the base station in cell l according to the current beamforming matrix, P max It is the maximum power that each base station can use to transmit signals.

[0054] Step 4 includes:

[0055] Generate a large amount of random sample data as training sets and test sets. The input of the data set is the channel information of the user terminal and each base station in the current cell, and the transmission power constraints of each base station. The output of the graph neural network is the beamforming matrix of the base station in the current cell, and the loss function is the sum of the total rates of all users.

[0056] Applying the graph neural network framework to the problem of maximizing the number of users and the rate, we obtain the following problem statement:

[0057]

[0058] The TX node feature is set to the base station transmit power, the RX node feature is set to the user-side noise, and the edge feature is set to the channel information between antennas; the beamforming matrix output by the network of each cell is applied to the overall optimization problem, and the beamforming matrix of each base station is output.

[0059] With the continuous development of mobile communication technology, 5G network brings faster data transmission speed, lower latency and greater network capacity. In this new generation of mobile communication network, multi-cell coordination technology is one of the key factors to realize its potential.

[0060] Multi-cell collaborative communication refers to the collaboration and cooperation between multiple cells in a 5G communication system to achieve higher network capacity and higher user experience. In traditional wireless communications, different cells are often independent of each other, but in multi-cell collaborative communication, resources can be shared and optimized between cells, thereby improving the efficiency and reliability of communication. The present invention proposes a distributed optimization scheme based on graph neural networks, which solves the beamforming design problem in downlink transmission by adjusting the graph neural network parameter output of each cell, greatly reducing the calculation time and network overhead while ensuring approximate performance. In the present invention, a machine learning method based on graph neural networks is proposed. Through simulation verification, the method of the present invention has good performance and can maximize users and rates under power constraints.

[0061] Beneficial effects: The centralized optimization scheme currently widely used in multi-cell networks requires high computational complexity in large-scale networks. Compared with the traditional centralized optimization algorithm, the graph neural network method proposed in the present invention can bring better performance. The learning method based on the graph neural network has excellent computing speed. Although the training time of the network is relatively long, the trained network can quickly output the beamforming matrix in different channel environments.

[0062] At the same time, this method has good generalization performance, and the trained model can adapt well to changes in the number of cells or base stations and the number of users. For a network structure with good generalization performance, a small change in the number of base stations BS (Base Station) and user equipment UE (User Equipment) will not cause obvious performance loss. In a relatively stable communication environment, it is closer to the production environment.

[0063] In summary, the contribution of this invention aims to solve some key challenges in implementing multi-cell MISO (Multiple Input Single Output) systems, such as collaborative optimization of power allocation. By using advanced graph neural network algorithms and considering base station power constraints, the scheme proposed in this invention can provide a practical solution for realizing high-performance multi-cell large-scale MISO networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0065] Figure 1 It is a simulation diagram of the usage scenario of the method of the present invention.

[0066] Figure 2 It is a schematic diagram of the model of the method of the present invention.

[0067] Figure 3 It is a schematic diagram of the overall architecture of the GNN method of the present invention.

[0068] Figure 4 It is a schematic diagram for comparing the cumulative distribution functions of the method of the present invention.

[0069] Figure 5 It is a schematic diagram of the generalization capability of the method of the present invention for the number of users in a cell.

[0070] Figure 6 It is a schematic diagram of the generalization capability of the method of the present invention with respect to the number of cells. DETAILED DESCRIPTION

[0071] The present invention mainly studies the problem of maximizing user transmission rate in a multi-cell communication system, improves the communication performance of the network by optimizing the beamforming matrix of the base station, and adopts a low-complexity algorithm based on a graph neural network to distribute and maximize the performance of the network, including the following steps:

[0072] Step 1, Problem Description:

[0073] Consider Figure 1 The multi-cell wireless communication network shown in FIG. 1 has L cells, each cell is equipped with a base station, and each base station serves N l (l≤L) users, the base station transmits signals to the user terminal equipment through the downlink; each base station has N t antennas, and the user terminal equipment has a single antenna.

[0074] The data symbol s of all user terminal equipment is expressed as Where E(s·s H ) = IN; the channel from the base station of cell l to user i in the cell is Represents, where i≤N l , the channel from the base station of cell j to user i outside the cell is indicates; w li represents the beam vector when the base station of cell l sends a signal to the terminal device of user i in cell l, then the signal received by the terminal device of user i in cell l can be expressed as:

[0075]

[0076] SINR received by the terminal device of user i in cell l li It is expressed as:

[0077]

[0078] in represents the additive Gaussian white noise of the ith user in the lth cell, represents the variance of the noise.

[0079] Establishing an optimization problem: In order for the base station to improve the total rate of users in all cells under limited power constraints, the method of the present invention formulates the following problem:

[0080]

[0081] Where st represents the set of constraints on the optimization problem. is the maximum transmit power that can be used by the base station in cell 1. The constraint ensures that the transmit power of each base station does not exceed its maximum power limit.

[0082] Step 2: This step proposes a graph structure model corresponding to the communication model of each cell.

[0083] like Figure 2 As shown in the figure, a heterogeneous graph model is established based on the communication network of a single cell. The types of nodes include transmitting nodes representing the base stations in the cell and the base stations in the adjacent cells, and receiving nodes representing the users in the cell. The symbols in the heterogeneous graph neural network are as follows: the TX node corresponds to the transmitting node in the heterogeneous graph, and the RX node corresponds to the receiving node in the heterogeneous graph. The feature vectors on the i-th TX node and the k-th RX node are Among them, d TX and d RX Respectively represent the corresponding feature dimensions. Therefore, the feature matrices of the TX node and the RX node can be expressed as and Similarly, a tensor can be used to represent the characteristics of the edge where d E is the edge feature dimension. On the other hand, the variables on the TX node and the RX node can be represented as N TX =[n TX，1 , ..., n TX，L ] T and in represents the variables on the i-th TX node and the k-th RX node, d′ TX ,d′ RX Respectively represent the corresponding variable feature dimensions. Similarly, the edge variables are represented as: d′ E Indicates the dimension of the edge variables.

[0084] For a distributed heterogeneous graph problem, it is modeled as follows:

[0085]

[0086] st(N TX ,N RX,l , E l )=φ(X TX ,X RX,l ,Y l )

[0087] where φ is the mapping function from features to variables.

[0088] Step 3 includes: The overall GNN architecture of this algorithm is as follows Figure 3 As shown;

[0089] The initialization module converts the input features into initial node and edge feature representations, which are updated according to the node and edge update mechanisms in the T node update modules. The dimensions of each feature remain unchanged during the update process. Finally, the normalization module converts the feature representations into the required variables and outputs them.

[0090] In the initialization module, since the current neural network is not suitable for processing complex numbers, the features that may be complex numbers (X TX , X RX，l , Y l ) is converted to real features Specifically, the real part and imaginary part of the complex number are extracted and concatenated to obtain the following representation:

[0091]

[0092] in represents the real part, Represents the imaginary part. Then a single-layer MLP with a rectified linear unit (ReLU) as the activation function is used to transform the initial representations of TX nodes, RX nodes, and edges:

[0093]

[0094]

[0095] These are all learnable parameters.

[0096] In the node update module, both node features and edge features need to be updated based on edge convolution. At the same time, based on the message passing model of graph neural network, the update of node features and edge features requires two steps of aggregation and combination.

[0097] In the node update module, the update strategy for the TX node features in the t-th layer of the neural network is:

[0098]

[0099] in yes The mth row of yes The kth row of represents the set of neighboring RX nodes of the mth TX node, It is the value of the aggregation layer output after aggregating adjacent RX nodes and edge features. AGG represents the aggregation function. are two different MLPs.

[0100] The update strategy for the RX node features in the t-th layer of the neural network is:

[0101]

[0102] in represents the set of neighboring RX nodes of the mth TX node, It is the value of the aggregation layer output after aggregating adjacent TX nodes and edge features. AGG represents the aggregation function. are two different MLPs.

[0103] The update strategy for edge features in the t-th layer of the neural network is:

[0104]

[0105] in yes The (m, k)th element of It is the value of the node features of the adjacent TX nodes and RX nodes output by the aggregation layer. They are three MLPs with different parameters.

[0106] The update strategies in the above three graph neural networks have been proven by existing work to be permutation equivariant and can greatly reduce the number of training samples.

[0107] In the normalization module, a single-layer MLP is first used to represent the features of each node and edge in the graph (X TX , X RX，l , Y l ) is mapped to the output variable The single-layer MLP uses the rectified linear unit as the activation function, specifically:

[0108]

[0109] is a trainable parameter;

[0110] Then we need to output the variable Convert to the required complex variable Specifically:

[0111]

[0112] Then the output variable Normalization is performed so that its power meets the constraints, and finally the beamforming matrix of the current cell base station is obtained as follows:

[0113]

[0114] in is the power of the signal transmitted by the base station in cell l according to the current beamforming matrix, Pmax It is the maximum power that each base station can use to transmit signals.

[0115] Step 4 uses a graph neural network-based method to optimize the beamforming matrix of the base station.

[0116] Generate sample data as training sets and tests. The input of the data set is the channel information of the user terminals and each base station in the current cell, and the transmission power constraints of each base station. The output of the graph neural network is the beamforming matrix of the base station in the current cell, and the loss function is the sum of the total rates of all users.

[0117] Applying the graph neural network framework to the problem of maximizing the number of users and the rate, we obtain the following problem statement:

[0118]

[0119] The TX node feature is set to the base station transmit power, the RX node feature is set to the user-side noise, and the edge feature is set to the channel information between antennas; the beamforming matrix output by the network of each cell is applied to the overall optimization problem, and the beamforming matrix of each base station is output.

[0120] Performance evaluation,Numerical results are used to demonstrate the effectiveness of the proposed multi-cell MISO communication system.

[0121] The specific simulation parameters are as follows: Consider a cellular network with uniform distribution of cell locations and user locations, and the base station is located at the center of each cell. The path loss is 30.5+36.7log 10 (d) dB, where d is the distance between the base station antenna and the user antenna. The small-scale channel follows Rayleigh fading.

[0122] When no modification of parameters is mentioned, network training and simulation testing are performed by default according to the parameter settings in Table 1 below:

[0123] Table 1 Simulation parameter setting table

[0124] parameter value Area <![CDATA[1.5×1.5km 2 ]]> Cell radius 250m Base station height 30m User receiving antenna height 1.5m Noise Power - -90dBm Maximum transmit power of base station 30dBm Number of cells 4 Number of users per cell 4 Number of antennas per base station 10

[0125] For the graph neural network method, all aggregation functions are implemented using the maximum aggregation function, and all MLPs are implemented by three linear layers, each followed by a ReLU activation function. The node update module is set to have three layers. During the training process, the number of epochs is set to 500, the batch size is set to 256, the training set size of the graph neural network is 10,000, and the learning rate is set to 10 -4, and the Adam optimizer is used. All experiments are implemented using Pytorch on Intel Core i7-11800H (2.30GHz) and NVIDIA GeForce RTX 3060 GPU (12GB).

[0126] Furthermore, in order to evaluate the proposed distributed edge convolution graph neural network algorithm (DistributedEGNN), the present invention sets up three baseline methods as typical for comparison. They are: a distributed graph neural network algorithm (Distributed GNN) without edge convolution and a centralized graph neural algorithm (Centralized GNN). The difference between the distributed graph neural network algorithm (Distributed GNN) without edge convolution and the algorithm proposed in the present invention is that the node update module aims to perform message passing and aggregation on the nodes instead of on the edges, that is, only update the node features, not the edge features; the difference between the centralized graph neural algorithm and the algorithm proposed in the present invention is that the entire multi-cell communication model is abstracted into a graph model as the input of the graph neural network for training, and the graph model of a non-single cell is input.

[0127] Figure 4The overall performance comparison of three graph neural network-based methods (Distributed EGNN, Distributed GNN, Centralized GNN) is shown. Taking the Weighted Minimum Mean-Square Error (WMMSE) algorithm as the benchmark algorithm, the cumulative distribution curves of the three graph neural network-based algorithms for 200 random realizations are plotted, where the horizontal axis is the sum of all user rates and the vertical axis is the cumulative distribution probability. It can be seen from the figure that the three graph neural network-based algorithms can produce higher user rates and rates than the WMMSE algorithm. The average user sum rate of the Distributed EGNN algorithm is improved by about 5.6% compared with the WMMSE algorithm, the average user sum rate of the Distributed GNN algorithm is improved by about 2.7% compared with the WMMSE algorithm, and the Centralized GNN algorithm is improved by about 4.1% compared with the WMMSE algorithm. Among them, the average number of users and rate of the Distributed EGNN algorithm are also significantly improved compared with the Distributed GNN algorithm. This is because compared with the message passing mode of the traditional graph neural network algorithm, the introduction of edge convolution is more conducive to user nodes obtaining additional feature information of the base station serving them, such as the interference channel information from the base station to other users. This message passing architecture can better adapt to the distributed multi-cell interference channel model, so that the message passing process can simultaneously transmit useful features and interference features, which can better fit the mathematical modeling form of the proposed problem, and therefore can also show better performance.

[0128] Figure 5 The generalization performance of three graph neural network-based methods (Distributed EGNN, Distributed GNN, Centralized GNN) for changes in the number of users in a cell is demonstrated. The neural network is trained in a training set with 4 users in the cell, and the performance is tested in a test set with 3 / 5 / 6 / 7 users in the cell. Figure 4 In the figure, the horizontal axis is the number of users in the cell, and the vertical axis is the number of users and the rate. Figure 5 It can be seen that the three graph neural network algorithms can adapt well to the increase or decrease in the number of users in the cell, reflecting good generalization performance.

[0129] Figure 6The generalization performance of three graph neural network-based methods (Distributed EGNN, Distributed GNN, Centralized GNN) for changes in the number of cells is demonstrated. The neural network is trained on a training set with 4 cells, and the test performance is tested on a test set with 3 / 5 / 6 / 7 cells. Figure 6 In the figure, the horizontal axis is the number of cells, and the vertical axis is the number of users and the rate. Figure 6 It can be seen that the three graph neural network algorithms can adapt well to the increase or decrease in the number of cells, showing good generalization performance. Compared with Centralized GNN, Distributed GNN performs better when the number of cells is 7. This is because as the number of cells increases, the increase in the scale of network nodes of the centralized algorithm is much greater than that of the distributed algorithm, which will lead to performance degradation.

[0130] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention are essentially or partly contributed to the prior art can be embodied in the form of a computer program, i.e., a software product, which can be stored in a storage medium and includes several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, a MUU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0131] The present invention provides a multi-cell wireless network distributed resource optimization method based on graph neural network. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A multi-cell wireless network distributed resource optimization method based on graph neural network, characterized in that: The following steps are involved: Step 1, taking maximizing the total transmission rate of user terminal equipment in multiple cells as the optimization goal, combined with the maximum power constraint of the cell base station, establish an optimization problem; Step 2: for the optimization problem established in step 1, design a graph model as an input of the graph neural network, wherein the graph model includes base station information of neighboring cells and base station and user information of the current cell; Step 3, establish a message passing graph neural network learning framework including edge embedding; Step 4: Taking the maximization of the sum of the system user transmission rates as the loss function, train and optimize the learning parameters of the graph neural network designed in step 3; Step 5: Distributedly deploy graph neural networks on base stations in each cell to design beamforming matrices, and implement distributed resource optimization of multi-cell wireless networks based on graph neural networks.

2. The method for optimizing distributed resources of a multi-cell wireless network based on a graph neural network according to claim 1, characterized in that: Step 1 includes: Assume that a network has L cells, each cell is equipped with a base station, and each base station serves N l (l≤L) users, the base station transmits signals to the user terminal equipment through the downlink; each base station has N t antennas, and the user terminal equipment has a single antenna. The data symbol s of all user terminal equipment is expressed as in s li The data sent by the base station of cell l to user i in the cell; the channel from the base station of cell l to user i in the cell is Represents, where i≤N l , the channel from the base station of cell j to user i outside the cell is indicates; w li represents the beam vector when the base station of cell l sends a signal to the terminal device of user i in cell l, then the signal y received by the terminal device of user i in cell l is li It can be expressed as: SINR received by the terminal device of user i in cell l li It is expressed as: in represents the additive Gaussian white noise of the ith user in the lth cell, represents the variance of the noise; The total transmission rate optimization problem of user terminal equipment is: Where st represents the set of constraints on the optimization problem. is the maximum transmission power that the base station of cell l can control, SINR li is the signal-to-interference-noise ratio received by the terminal equipment of user i in cell l, is the beamforming matrix of the base station in cell l.

3. The multi-cell wireless network distributed resource optimization method based on graph neural network according to claim 2 is characterized in that: Step 2 includes: The problem is modeled as follows using graph neural networks: s.t.(N Tx ,N Rx,l ,E l )=φ(X Tx ,X Rx,l ,Y l ) The communication topology model based on cell l is used to build a heterogeneous graph model G l = {N TX , N RX，l , E l }, where N TX = {N l |l≤L} represents the transmitting nodes of the entire multi-cell communication system, and the transmitting nodes include base stations within the cell and base stations outside the cell, N TX =[n TX，1 , ..., n TX，L ] T Represents the variable on the TX node, Represents the feature matrix of the TX node; N RX,l = {N i |i≤N l ,l≤L} represents the receiving node in cell l, i.e., the user node, N RX，l =[n RX，l，1 , ..., n RX，l，L ] T Represents the variable on the RX node, Represents the feature matrix of the RX node; E l represents the edges between the nodes of the graph represented by the current cell l, represents the variable on the edge, Represents the edge features of each edge; d TX d RX and d E They represent feature dimensions respectively, where φ is the mapping function from features to variables.

4. The method for optimizing distributed resources of a multi-cell wireless network based on a graph neural network according to claim 3 is characterized in that: Step 3 includes: Constructing a graph neural network model based on a message passing model, wherein the graph neural network model includes an initialization module, a node update module, and a normalization module; The initialization module first converts the complex features (X TX , X RX，l , Y l ) is converted to real features in represents the real part, represents the imaginary part; the initial representations of TX node features, RX node features, and edge features are transformed by a single-layer MLP with a rectified linear unit as the activation function: in is the learnable parameter of the network; The node update module consists of two parts: an aggregation layer and a combination layer. The aggregation layer collects adjacent node features and edge features, and the combination layer combines the features obtained by the aggregation layer and the node features to obtain new features. In the node update module, the update strategy for the TX node features in the t-th layer of the neural network is: in yes The mth row of yes The kth row of Indicates m The set of neighboring RX nodes of a TX node, It is the value of the aggregation layer output after aggregating adjacent RX nodes and edge features. AGG represents the aggregation function. They are two different MLPs; The update strategy for the RX node features in the t-th layer of the neural network is: in represents the set of neighboring RX nodes of the mth TX node, It is the value of the aggregation layer output after aggregating adjacent TX nodes and edge features. AGG represents the aggregation function. They are two different MLPs; The update strategy for edge features in the t-th layer of the neural network is: in yes The (m, k)th element of It is the value of the node features of the adjacent TX nodes and RX nodes output by the aggregation layer. They are three MLPs with different parameters; In the normalization module, the features of each node and edge in the graph are first represented by a single-layer MLP. Mapping to output variables The single-layer MLP uses the rectified linear unit as the activation function, specifically: is a trainable parameter; Then we need to output the variable Convert to the required complex variable Specifically: For output variables Normalized to (N TX , N RX，l , E l ) so that its power satisfies the constraint; Get the beamforming matrix W of the current cell base station l The specific manifestations are: in is the power of the signal transmitted by the base station in cell l according to the current beamforming matrix, P max It is the maximum power that each base station can use to transmit signals.

5. The method for optimizing distributed resources of a multi-cell wireless network based on a graph neural network according to claim 4, characterized in that: Step 4 includes: Generate sample data as training sets and test sets. The input of the data set is the channel information of the user terminal and each base station in the current cell, and the transmit power constraints of each base station. The output of the graph neural network is the beamforming matrix of the base station in the current cell, and the loss function is the sum of the total rates of all users. Applying the graph neural network framework to the problem of maximizing the number of users and the rate, we obtain the following problem statement: The TX node feature is set to the maximum transmit power of the base station, the RX node feature is set to the user-side noise, and the edge feature is set to the channel information between antennas; the beamforming matrix output by the network of each cell is applied to the overall optimization problem, and the beamforming matrix of each base station is output.

Images