Decentralized graph neural network architecture for beam forming of multi-cell multi-user MIMO (Multiple Input Multiple Output) communication system

Abstract

The invention provides a decentralized graph neural network architecture for multi-cell multi-user MIMO communication system beam forming, and the construction steps of the architecture comprise: step 1, in a method for optimizing the multi-cell multi-user MIMO communication system beam forming by using a graph neural network, constructing a local loss function of a base station, and realizing computational graph decoupling; step 2, constructing a heterogeneous graph model based on the local loss function of the base station proposed in the step 1, and realizing decoupling of a graph model level; and step 3, carrying out distributed training and reasoning on the heterogeneous graph model to obtain a decentralized graph neural network architecture oriented to multi-cell multi-user MIMO communication system beam forming.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of graph neural network architecture, in particular, a kind of decentralized graph neural network architecture for multi-cell multi-user MIMO communication system beam forming Decentralized graph neural network architecture for multi-cell multi-user MIMO (Multiple Input Multiple Output) communication system beam forming problem. BACKGROUND

[0002] The part provided in this section is merely background information related to the present disclosure, which does not necessarily have to be prior art.

[0003] In the evolution of 5G and future 6G communication technology, with the access of massive devices and the increase of base station transmit antenna number, the method of AI+communication becomes the mainstream of future communication and is included in the 6G white paper. The method using AI can realize end-to-end output, which effectively improves the system performance while avoiding the long iteration and high computational complexity of traditional optimization algorithm, meeting the requirements of low latency, high throughput and endogenous intelligence of communication system. The method based on multi-layer perception MLP, convolutional neural network CNN, reinforcement learning and other deep learning methods has been widely applied in recent years. Graph neural network (GNN) can effectively learn the topology information of network due to its unique graph building method and message passing paradigm, and has the advantages of excellent performance, strong generalization and good scalability in communication system beam forming.

[0004] However, with the increase of the number of users, the number of cells and the number of base station transmit antennas, the computational load of the central computing node of the system for solving the communication system beam forming problem is increasing. Although the use of AI method greatly reduces the computational complexity compared with the traditional iterative method, the centralized computing of central processing unit still brings high computational burden to the central computing unit. In addition, the centralized GNN training and inference method needs to collect all the information of the system users, which is not conducive to the privacy protection demand of users, so a decentralized GNN training and inference method is needed.

[0005] With the deepening of the research on GNN, the existing technologies extend the deployment of GNN to a distributed scenario (reference: M. Marwani and G. Kaddoum, "Graph Neural Networks Approach for Joint Wireless Power Control and Spectrum Allocation," in IEEE Transactions on Machine Learning in Communications and Networking, vol. 2, pp. 717-732, 2024, doi:10.1109 / TMLCN.2024.3408723. and Y. Gu, C. She, S. Bi, Z. Quan and B. Vucetic, "Graph Neural Network for Distributed Beamforming and Power Control in Massive URLLC Networks," in IEEE Transactions on Wireless Communications, vol. 23, no. 8, pp. 9099-9112, Aug. 2024, doi: 10.1109 / TWC.2024.3358903.), message passing in distributed inference is realized through air computing and other methods, and decentralization and edge computing are realized in the distributed inference link, but the parameters on each base station still need to be trained and issued through a centralized method, and complete decentralization has not been achieved. According to the known information, there are few schemes that try to train GNN in a distributed manner, and some existing technologies (reference: J. Zhao, H. Ling, C. Yang and T. Liu, "Decentralized Training of Graph Neural Networks in Mobile Systems for Power Control," 2024 IEEE Wireless Communications and Networking Conference (WCNC), Dubai, United Arab Emirates, 2024, pp. 1-6, doi: 10.1109 / WCNC57260.2024.10571216.) consider the distributed training scenario, but do not realize complete decoupling.

[0006] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a decentralized graph neural network architecture for multi-cell multi-user MIMO communication system beamforming problem, called FDGNN (Fully Distributed Graph Neural Network, FDGNN).

[0008] In order to solve the above technical problems, the present application discloses a decentralized graph neural network architecture for multi-cell multi-user MIMO communication system beamforming problem, the construction steps of the architecture include:

[0009] Step 1, in the method of using graph neural network to optimize multi-cell multi-user MIMO communication system beamforming, a base station local loss function is constructed to realize the decoupling of the computation graph;

[0010] Step 2, based on the base station local loss function proposed in step 1, a heterogeneous graph model is constructed to realize the decoupling at the graph model level;

[0011] Step 3, distributed training and inference are performed on the heterogeneous graph model to obtain a decentralized graph neural network architecture for multi-cell multi-user MIMO communication system beamforming.

[0012] Further, the construction of the base station local loss function in step 1 is to replace the signal-to-interference-and-noise ratio (SINR) with a loss function based on the signal-to-leakage-and-noise ratio (SLNR) approximation.

[0013] Further, the base station local loss function in step 1 is represented as follows:

[0014]

[0015] wherein, is the local loss function of the mth cell base station, is the graph neural network parameter weight on the mth cell base station, is the global parameter weight of the central parameter server, is a hyperparameter, is the number of users of the mth cell base station, is the number of users of the mth cell base station, is the downlink communication rate calculated based on the signal-to-leakage-and-noise ratio (SLNR) of the nth user in the mth cell base station. ​​​

[0016] Furthermore, the aforementioned first The first of the cell base stations Downlink Class Communication Rate for Individual Users Based on Signal-to-Leakage-to-Noise Ratio (SLNR) The calculation method is as follows:

[0017]

[0018] in, Let be the signal-to-noise ratio (SNR) between the m-th cell base station and the k-th user.

[0019] Furthermore, the signal-to-noise ratio (SNR) of the m-th cell base station to the k-th user... The calculation method is as follows:

[0020]

[0021] in, Item and The items are leakage within the community and leakage between communities. The power of additive white Gaussian noise, For the first Each cell base station reaches the first cell within its own cell. Downlink channel state information for each user For the first The cell base station reaches the cell's first... Beamforming vectors for each user Let be the downlink channel state information from the m-th cell base station to the l-th user in the n-th cell. For the first Number of users per cell base station This refers to the number of base stations in the cell.

[0022] Furthermore, the construction of the heterogeneous graph model described in step 2 includes:

[0023] Step 2-1: Treat all user UEs associated with the base station (BS) in each cell as nodes, with the types being serving UEs within the cell and UEs affected by interference outside the cell.

[0024] Step 2-2: Construct a bipartite heterogeneous graph, where there is one edge between two nodes of different types;

[0025] Steps 2-3: At each base station (BS), a heterogeneous graph neural network (GNN) with identical parameters and number of layers is used, and the parameters are shared between layers.

[0026] Steps 2-4 define the inputs and outputs of the heterogeneous graph model.

[0027] Further, the input for defining the heterogeneous graph model in step 2-4 is specifically as follows:

[0028] The input for defining the heterogeneous graph model is the initial value of the node feature, that is, the channel value of the base station BS to the node representing the user, and the edge has no feature.

[0029] Further, the output for defining the heterogeneous graph model in step 2-4 is specifically as follows:

[0030] The output of the heterogeneous graph model is the corresponding beamforming vector on the serving UE node in the cell.

[0031] Further, the distributed training and inference of the heterogeneous graph model in step 3 includes:

[0032] Step 3-1, initialize the model parameters of the local graph neural network GNN of each base station , set the global aggregation period T of federated learning;

[0033] Step 3-2, sample the channel information in the cell to form a local training data set;

[0034] Step 3-3, construct the heterogeneous graph model of each base station and perform forward propagation;

[0035] Step 3-4, calculate the local loss function of the base station;

[0036] Step 3-5, calculate the backward propagation gradient and update the parameters of the heterogeneous graph model;

[0037] Step 3-6, realize parameter aggregation through federated learning, and after every T local training period, perform 1 federated average to complete global update.

[0038] Further, the hyperparameters in step 1 , are artificially preset.

[0039] Beneficial effects:

[0040] 1. The application realizes the distributed decoupling of the GNN network of the multi-cell multi-user MIMO communication system, and has the ability of completely distributed training and inference.

[0041] 2. The application can effectively solve the beamforming of the multi-cell multi-user MIMO system, and solve the problem of heavy calculation load of the central calculation unit in the original centralized training and inference process.

[0042] 3、The distributed training method provided by the application can utilize the topological structure advantage of GNN on the basis of local training, and obtain global information through parameter aggregation between base stations, so that the obtained GNN is independently scalable, and has the advantages of decentralization and large-scale application. BRIEF DESCRIPTION OF DRAWINGS

[0043] The above and / or other aspects of the application will become more apparent by describing in detail the preferred embodiments thereof with reference to the attached drawings.

[0044] Figure 1 is a schematic diagram of the physical structure of the system of the application.

[0045] Figure 2 is a schematic diagram of the communication topology of the application.

[0046] Figure 3 is a schematic diagram of the coupling of the original method of centralized GNN computation graph.

[0047] Figure 4 is a schematic diagram of the decoupling of the distributed GNN computation graph of the application.

[0048] Figure 5 is a local graph structure of the GNN base station of the application.

[0049] Figure 6 is a schematic diagram of the centralized computation architecture of the original method of centralized GNN.

[0050] Figure 7 is a schematic diagram of the distributed computation architecture of the distributed GNN of the application.

[0051] Figure 8 is a schematic diagram of the distributed training parameter aggregation method of the distributed GNN of the application.

[0052] Figure 9 is a simulation result diagram of the distributed GNN of the application. DETAILED DESCRIPTION

[0053] The application designs a fully distributed GNN for beamforming in a multi-cell multi-user MIMO (Multiple Input Multiple Output) communication system to reduce the computational burden of the central computing unit in the AI-assisted solution of the communication system beamforming problem, and even to realize decentralization, realizes distributed training and distributed inference, and has the ability of completely distributed deployment.

[0054] The original GNN for the beam allocation problem usually adopts a centralized training and inference mode, the center node has a large calculation load, and the existing decentralized GNN focuses on decentralizing the inference process, but the training process is still carried out in the center node, and the complete distribution is not achieved.

[0055] The difficulty of realizing the fully distributed GNN is how to decouple the coupled computation graph, which is also the main problem solved by the application. The specific solution method is summarized as follows:

[0056] Point one: first decouple the distributed GNN computation graph by using the signal leakage and noise ratio (SLNR) approximation method.

[0057] Point two: by associating the users of each cell base station into service nodes and interference nodes for heterogeneous graph modeling, the decoupling of the graph model is realized.

[0058] Point three: by selecting a parameter server base station to aggregate parameters, the distributed training process after decoupling is realized, and the GNN model trained thereby also naturally has the ability of distributed inference. It should be noted that the parameter server only undertakes the task of parameter gradient aggregation and model update distribution, and does not undertake the high calculation load of the training task.

[0059] The technical solutions adopted by the application are described below:

[0060] In the multi-cell multi-user MIMO system downlink communication and rate maximization problem, the goal of training the neural network is to maximize the system user and rate, that is, the negative value of the user and rate of each cell is taken as the loss function, and the loss function is minimized through back propagation. However, when calculating the user rate through the signal to interference plus noise ratio (SINR), the beamforming solution process between cells will be coupled due to inter-cell interference. In order to realize distributed training and inference, it is necessary to decouple it.

[0061] a) The first point, in the loss function, the SLNR is used to approximate the SINR.

[0062] Firstly, the infeasibility of directly using the classical SINR calculation and rate as the loss function is described. In the original SINR-based classical user rate calculation formula, the SINR is in the following form:

[0063]

[0064] wherein, This refers to the channel state information from the m-th cell base station to the k-th user within its cell. This refers to the channel state information from the nth cell base station to the kth user within that cell. Let be the beamforming vector from the m-th cell base station to the k-th user within its cell. Given the power of additive white Gaussian noise, the user rate of the k-th user in the m-th cell is obtained as follows: .

[0065] As can be seen, the SINR formula includes intra-cell interference and inter-cell interference. Inter-cell interference requires the transmit power and channel information of other base stations, and calculating the transmit power of other base stations also requires the transmission information provided by those base stations to provide the interference term, thus forming an inter-cell coupling relationship. If we directly use the negative values ​​of cell users and rates at each base station, i.e. Using the local loss function for distributed training means that each base station needs to obtain the current state of other base stations in real time when performing forward and backward propagation, such as... Figure 3 As shown, computational graph coupling in this case would cause a large amount of information transmission overhead in distributed training, which is obviously unacceptable.

[0066] To decouple the computation graph, it is necessary to first process the coupling terms that require a large amount of information exchange, namely the inter-cell interference terms in SINR. .

[0067] Analysis reveals that this is a competitive, non-cooperative game. To transform this problem, this invention draws on the concept of "altruistic" cooperative game theory, shifting the focus from minimizing interference from other base stations to minimizing leakage to users served by other base stations. This transforms the receiver-centric joint optimization problem into a transmitter-centric distributed optimization problem, thus replacing the coupling terms in the loss function. At this point, the SINR form can be expressed as the SLNR form:

[0068]

[0069] in, Item and The terms represent intra-cell leakage and inter-cell leakage, achieving decoupling from the original SINR. In summary, the purpose of SLNR is to enhance the signal of the target user while reducing interference to other users. An approximate SLNR-based rate is calculated using this method, i.e. The loss function is backpropagated using the local SLNR-based sum rate of each cell, and updated via gradient descent. At this point, the local loss function can be considered as... .

[0070] Because of calculations based on SLNR is a biased approximation of the user's true rate Without the computation of the central node, the local loss function needs to be further modified to compensate for the gap caused by the approximation. By analyzing the decoupled local loss function, it can be found that the process of minimizing the local loss function based on the SLNR approximation is a "greedy" gradient descent process, which will pull the local model to a solution that is very effective for local leakage but may be very detrimental to global interference. Therefore, by adding a proximal regularization term to the local loss function to punish the deviation of the local model and the global consensus, where is a hyperparameter defined artificially, which can be obtained by issuing the parameters of each round of federated learning.

[0071] In summary, the final local loss function is:

[0072]

[0073] At this time, each base station only needs to collect the service link channel information and the interference link channel information of other users at the base station when performing forward propagation and back propagation, without the need for other base stations to transmit local model calculation intermediate results, such as Figure 4 The calculation graph in this case can realize the decoupling between base stations and provide a basis for subsequent distributed training and inference.

[0074] b) The second point, based on the loss function, the application proposes a corresponding heterogeneous graph model, which changes the original centralized GNN's graph building method of connecting different inter-cell edges, and reconstructs and splits at each base station to realize the decoupling at the graph model level. The specific graph building method of the application is shown in Figure 5 : All users (User Equipment, UE) related to the base station (Base Station, BS) in each cell are regarded as nodes, and the types are "in-cell serving UE" and "out-of-cell interference affected UE", respectively. Thus, a bipartite heterogeneous graph is modeled, and there is an edge between two nodes of different categories. The node features are initially the channel values from the BS to the node, and the edges have no features.

[0075] The graph building method of the original centralized GNN is shown in Figure 6 This graph building method integrates all cells of the system into a whole graph, and there are edges between cells. The above graph building method considers nodes affected by interference in other cells, but there is no edge connection with the graph structure of other cells, as shown in Figure 7The local GNN of each base station can perform message passing and aggregation on the heterogeneous graph based on the heterogeneous graph model, and the final output is the corresponding beamforming vector on the "in-cell UE" node. At the same time, in order to facilitate the message aggregation of distributed training in the following, the multi-layer heterogeneous GNN with the same number of parameters and layers is adopted at each base station, and the gradients are shared between layers.

[0076] The message passing process of the local GNN after the heterogeneous graph modeling is as follows: for any node v in the graph, the message update process of the k-th layer GNN is as follows:

[0077] Message generation and aggregation: first, the input node features are converted into node messages through a learnable MLP, and then according to the connection relationship between the two heterogeneous nodes, each target node from its topological neighbors aggregate information in the form of summation, averaging or attention weighted sum, where is a permutation invariant aggregation operator, is a trainable message generation function, is the edge feature;

[0078] Node update: the node combines its own last layer feature and the aggregated message to generate a new node feature through an update function: where is a parameterized nonlinear function.

[0079] After K layers of such message passing, the final feature of each "in-cell serving UE" node is sent to a fully connected layer output layer to obtain where is the output layer parameterized nonlinear function. Finally, the power is normalized to generate the corresponding beamforming vector of the node where is the maximum transmit power of the base station.

[0080] d) The third point is to design a complete distributed training and inference process for the proposed FDGNN.

[0081] The local data owned by each BS is explicitly defined as local channel samples and local GNN weight parameters. The specific process is as follows:

[0082] First, the channel information of all service links in the cell and the channel information of the interference leakage link, i.e., the channel from the BS to all UEs interfered by it, are locally sampled. After sampling, forward propagation and backward propagation are performed on the local heterogeneous GNN, the gradient is calculated after the propagation, and the local GNN model is updated in parameters. Since the GNN parameter dimensions of each cell are consistent, the parameters can be aggregated and distributed by the central parameter server. As shown in Figure 8 , the local parameters of each base station are uploaded to the parameter server of the system, and aggregation and re-distribution are performed on the parameter server:

[0083]

[0084]

[0085] Finally, the distributed parameters are used as the latest local parameters and participate in the regularization term of the loss function .

[0086] After multiple rounds of parameter updating to convergence, the local GNN model parameters of each base station are unified, and in the deployment and inference stage, the channel information of the users in the cell and the users interfered outside the cell is input as the node features into the local GNN, and the beamforming matrix transmitted by the base station can be obtained by forward message passing.

[0087] It is worth noting that the FDGNN also has scalability, i.e., the number of users in the cell and the number of users interfered outside the cell can be dynamically changed, and at the same time, if the number of base stations of the communication system is expanded, the obtained local unified GNN can be directly migrated to the newly expanded base station without retraining.

[0088] Embodiment:

[0089] The following is a specific case of the complete process of the distributed beamforming optimization system based on GNN and SLNR, which illustrates the implementation process of the above technical solution.

[0090] Consider a multi-cell multi-user MIMO wireless communication system, as shown in Figure 1 , the number of cells of the system is M, each cell is equipped with a base station with antennas, and the th cell base station serves multi-antenna users, and it is assumed that the user receiving antenna number is 1. The optimization goal of the system is to maximize the sum of user rates under the constraint of the maximum transmit power of the base station, and the sum of user rates here is calculated based on SINR:

[0091]

[0092]

[0093] For the above optimization problem, the communication system is first modeled as a network topology, as shown in Figure 2 The red nodes represent base stations, and the blue nodes represent users. The channels between base stations and nodes can be divided into service links and interference links, depending on whether the users are within the cell and the range of the base station signal. Further, the network topology is abstracted as a graph structure , where the vertex set represents communication entities, including cell base station serving user nodes and user nodes interfered by the cell base station, and the edge set describes the interconnection relationship. The node feature matrix encapsulates the physical layer channel parameters of the base station to the node communication entity node.

[0094] First part: distributed training process:

[0095] Each base station has local data and local GNN parameters. In the initialization phase, each base station (e.g., the mth base station ) randomly initializes its local GNN model parameters . Set the federal learning global aggregation period T (e.g., perform global aggregation once after completing 100 local training periods). Set the central coordination node or designated base station as the parameter server to perform model averaging operations.

[0096] Local sampling: sample the channel information within the cell to form the local training data set, including all link channel information and interference leakage link channel information within the cell.

[0097] Establish a local graph model: model the local topology as a heterogeneous graph. The types are "in-cell serving UE" and "out-of-cell interfered UE". The node features are initialized as the channel values from the BS to the node, and the edges have no features. The GNN output is the corresponding beamforming vector on the "in-cell serving UE" node. Perform message passing on the local heterogeneous graph without passing node information from other cells.

[0098] Forward message passing: the local GNN model uses a multi-layer heterogeneous message passing architecture. Let the trainable parameters of the GNN be , which is stacked through layers. This framework gradually refines the node representation, captures the multi-hop neighborhood coupling effect, encodes the interference relationship and network dependence, and is used for communication problem optimization such as beamforming.

[0099] Calculate the local loss function: introduce SLNR to decouple cross-user, solve the problem of inter-cell computational graph coupling caused by traditional SINR-based cell and rate as local loss function. The base station local loss function can be written as:

[0100]

[0101] where, Here, SLNR can reduce the interference to other users while enhancing the target user signal. Finally, the power normalization of the node output beamforming vector satisfies the maximum transmit power constraint.

[0102] Backward pass and local model parameter update: the local GNN adopts a gradient descent algorithm (such as the Adam optimizer) to minimize the loss function , update its local GNN parameters , the update method is ← , where η is the learning rate.

[0103] Global update: directly realize gradient aggregation through federated learning, and the system performs federated averaging once every T local training cycles: specify a base station (optionally the system center position base station) as the system parameter server. Each base station uploads its local GNN parameters , ,..., } to the parameter server. The parameter server base station accepts the gradient information of other nodes and performs reduction, such as , and distributes it to all base stations. Each base station replaces its local parameters with the global model parameters: ← After that, a new round of distributed local training begins. Thus, the distributed training is completed.

[0104] Second part, distributed inference:

[0105] The inference process is carried out locally at each base station. Since each base station has complete heterogeneous node channel information locally, it can directly input the trained GNN for independent inference without uploading the center node for calculation and distribution. The output on each "intra-cell serving UE" node after the forward pass of the local GNN is the beamforming vector. The GNN model used for this distributed inference has scalability and can be extended to more cells without retraining.

[0106] Simulation results:

[0107] The embodiments of the application evaluate the performance of the proposed fully distributed graph neural network framework in a multi-cell downlink system containing M=16 hexagonal cells. Each base station is configured with transmit antennas, deployed at the center of the cell, and limited to maximum transmit power constraint. User devices are associated following the geographical proximity principle, i.e. all user devices located within the cell hexagon boundary are served independently by the cell base station. The 96 single-antenna user devices are uniformly distributed in the coverage area, maintaining a minimum distance of 2m from the base station to avoid near-field effects. The cell radius R is normalized to 40m, ensuring the consistency of path loss comparisons. The composite channel model is where the path loss follows with a reference distance and an attenuation exponent and the log-normal shadowing is where , dB, the small-scale fading satisfies and the additive white Gaussian noise power at the receiver is W. The interference effect is modeled in the area satisfying i.e. the path loss attenuation relative to the cell edge does not exceed 20dB, and the super parameter is set to 0.001.

[0108] The beamforming matrix is expressed in real-imaginary part separation, with the input feature dimension of The FDGNN adopts three fully connected neural networks to realize heterogeneous message passing: the message generation network is a neural layer structure with dimensions of [32, 32, 64], and the activation function uses LeakyReLU; the node update network is a neural layer structure with dimensions of [96, 32, 32], and the activation function uses LeakyReLU; the parameters are shared between layers.

[0109] Table 1 is a system and rate comparison (bps / Hz) table, and the simulation results show that the proposed FDGNN can greatly reduce the computational load of the central node of the communication system while ensuring performance.

[0110] Table 1 System and rate comparison (bps / Hz) table

[0111]

[0112] Table 2 is a training computation comparison table, where the centralized GNN training and inference are performed on the central node server, and the proposed FDGNN training and inference are performed locally on the BS.

[0113] Table 2 Training computation comparison table

[0114]

[0115] Comprehensive simulation results Figure 9As shown in Table 1, the performance of the FDGNN implementing distributed training and inference has approached or surpassed the Weighted Minimum Mean Square Error (WMMSE) algorithm commonly used as a benchmark algorithm, and is much higher than the Zero-Forcing (ZF) algorithm and the Minimum Mean Square Error (MMSE) algorithm. Although still inferior to the CGNN of the central training and inference, as shown in Table 2, the computing load of the edge node is 5.8% of the CGNN, effectively reducing the computing load of the central node and realizing the distributed training and inference capability.

[0116] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and the computer program can run part or all of the steps in the invention content and each embodiment of the decentralized graph neural network architecture for beamforming of a multi-cell multi-user MIMO communication system when executed by the data processing unit. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), etc.

[0117] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be realized by means of a computer program and its corresponding general hardware platform. Based on this understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a computer program, i.e. a software product, which can be stored in a storage medium, including a number of instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, an MCU, or a network device, etc.) containing a data processing unit to execute the method described in each embodiment or some parts of the embodiments of the present application.

[0118] The present application provides a decentralized graph neural network architecture for beamforming of a multi-cell multi-user MIMO communication system. There are many methods and ways to implement this technical solution, and the above description is only the preferred embodiment of the present application. It should be noted that for ordinary skilled persons in the technical field, without departing from the principles of the present application, some improvements and refinements can be made, which should also be considered as the protection scope of the present application. The components not explicitly described in the embodiments can be implemented by existing technology.

Claims

1. A decentralized graph neural network architecture for beamforming in multi-cell, multi-user MIMO communication systems, characterized in that, The construction steps of the architecture include: Step 1: In the beamforming optimization method for multi-cell multi-user MIMO communication systems using graph neural networks, a base station local loss function is constructed to achieve computational graph decoupling; Step 2: Based on the base station local loss function proposed in Step 1, construct a heterogeneous graph model to achieve decoupling at the graph model level; Step 3: Perform distributed training and inference on the heterogeneous graph model to obtain a decentralized graph neural network architecture for beamforming in multi-cell multi-user MIMO communication systems.

2. The decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 1, characterized in that, The construction of the base station local loss function mentioned in step 1 is to replace the signal-to-interference-plus-noise ratio (SINR) with a loss function based on the signal-to-leakage-noise ratio (SLNR) approximation.

3. The decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 2, characterized in that, The base station local loss function mentioned in step 1 is expressed as follows: in, For the first Local loss function for each cell base station These are the graph neural network parameter weights on the m-th cell base station. It is the global parameter weight of the central parameter server. For hyperparameters, For the first Number of users per cell base station For the first The first of the cell base stations Downlink class communication rate for each user, calculated based on signal-to-leakage-to-noise ratio (SLNR).

4. The decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 3, characterized in that, The first The first of the cell base stations Downlink Class Communication Rate for Individual Users Based on Signal-to-Leakage-to-Noise Ratio (SLNR) The calculation method is as follows: in, Let be the signal-to-noise ratio (SNR) between the m-th cell base station and the k-th user.

5. A decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 4, characterized in that, The signal-to-noise ratio of the m-th cell base station to the k-th user The calculation method is as follows: in, Item and The items are leakage within the community and leakage between communities. The power of additive white Gaussian noise, For the first Each cell base station reaches the first cell within its own cell. Downlink channel state information for each user For the first The cell base station reaches the cell's first... Beamforming vectors for each user Let be the downlink channel state information from the m-th cell base station to the l-th user in the n-th cell. Let n be the number of users in the nth cell base station. This refers to the number of base stations in the cell.

6. The decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 5, characterized in that, The construction of the heterogeneous graph model described in step 2 includes: Step 2-1: Treat all user UEs associated with the base station (BS) in each cell as nodes, with the types being serving UEs within the cell and UEs affected by interference outside the cell. Step 2-2: Construct a bipartite heterogeneous graph, where there is one edge between two nodes of different types; Steps 2-3: At each base station (BS), a heterogeneous graph neural network (GNN) with identical parameters and number of layers is used, and the parameters are shared between layers. Steps 2-4 define the inputs and outputs of the heterogeneous graph model.

7. A decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 6, characterized in that, The specific method for defining the input of the heterogeneous graph model as described in steps 2-4 is as follows: The input to the heterogeneous graph model is defined as the initial values ​​of the node features, i.e., the channel values ​​from the base station (BS) to the node representing the user, and the edges have no features.

8. A decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 7, characterized in that, The specific method for defining the output of the heterogeneous graph model as described in steps 2-4 is as follows: The output of the heterogeneous graph model is defined as the beamforming vector corresponding to the serving UE node within the cell.

9. A decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 8, characterized in that, Step 3, which describes distributed training and inference of the heterogeneous graph model, includes: Step 3-1: Initialize the model parameters of the local graph neural network (GNN) for each base station. Set the global aggregation period T for federated learning; Step 3-2: Sample the channel information within the cell to form a local training dataset; Step 3-3: Construct a heterogeneous graph model for each base station and perform forward propagation; Steps 3-4: Calculate the local loss function of the base station; Steps 3-5: Calculate the backpropagation gradient and update the parameters of the heterogeneous graph model; Steps 3-6 involve parameter aggregation through federated learning. After every T local training cycles, a federated average is performed once to complete the global update.

10. A decentralized graph neural network architecture for beamforming in a multi-cell, multi-user MIMO communication system according to claim 9, characterized in that, The hyperparameters mentioned in step 1 It is preset by humans.

Images