Non-cell network power control method based on Grapher

Abstract

The invention discloses a Grapher-based no-cell network power control method, which comprises the following steps of: constructing a no-cell network adopting a time division duplex mode, defining an uplink training model to carry out channel estimation, and defining a downlink data transmission model to pre-code downlink data symbols based on an estimated channel; constructing graph structure data based on the cell-free network, the distributed access point and the user equipment; building a neural network model based on a Grapher; according to the constructed graph structure data, the graph structure information is coded to fully extract the global and local information of the network, the Grapher-based neural network is trained to obtain a trained Grapher-based neural network model, and based on the transmission power of the distributed access point and the receiving noise of the user equipment, the distributed access point is determined to be a new one. And according to a channel between the distributed access point and the user equipment, realizing cell-free network power control by using the trained Grapporter-based neural network model.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and in particular, to a cell-free network power control method based on Graphormer. Background Art

[0002] To meet the continuous growth of the number and capacity requirements of mobile user equipment (UE), mobile communication systems are being upgraded and evolved. The cell-free network has become one of the important candidate architectures for 6G networks due to its advantages such as uniform signal coverage and avoidance of handovers. The cell-free network consists of multiple access points (APs) randomly distributed in the coverage area, and each AP is connected to a central processing unit (CPU) through a fronthaul link, and the CPU coordinates the cooperation between APs. All APs share the same time-frequency resources to serve each UE. However, in this architecture, the multi-user interference in the communication network increases significantly, resulting in a decrease in the user signal-to-interference-plus-noise ratio. To solve this problem, power control has become one of the key technologies in cell-free networks.

[0003] Power control methods based on traditional optimization algorithms usually require a large number of iterations, have high computational complexity, and are difficult to be applied in real time in future large-scale networks. Therefore, how to use intelligent technologies - especially deep learning technologies - to improve the efficiency of power control algorithms has become a hot research issue. In recent years, graph neural networks have been widely used in resource scheduling and allocation in the field of wireless communication due to their powerful ability to process graph-structured data. As an innovative architecture that combines Transformer and graph neural networks, Graphormer can extract global and local information of complex graph-structured data and is suitable for modeling the interference relationship in cell-free networks. However, current research based on Graphormer mainly focuses on general graph learning tasks, and its application in the field of wireless communication is still in the initial exploration stage. Summary of the Invention

[0004] According to the above-mentioned technical problems, the present invention adopts the following technical means: A cell-free network power control method based on Graphormer, comprising the following steps:

[0005] Construct a cell-free network using the time-division duplex mode, define an uplink training model for channel estimation, and define a downlink data transmission model to precode downlink data symbols based on the estimated channel;

[0006] Based on the cell-free network, distributed access points, and user equipment, construct graph-structured data;

[0007] Build a neural network model based on Graphormer;

[0008] According to the constructed graph structure data, encode the graph structure information to fully extract the global and local information of the network, train the neural network based on Graphormer, and obtain a trained neural network model based on Graphormer.

[0009] Based on the transmit power of distributed access points, the received noise of user equipment, and the channels between distributed access points and user equipment, use the trained neural network model based on Graphormer to implement cell-free network power control.

[0010] Furthermore, the graph structure data includes topological relationships, features and variables of nodes and edges, specifically as follows:

[0011] Model the cell-free network using a heterogeneous undirected bipartite graph, where distributed access points AP and user equipment UE are regarded as two types of nodes, namely AP nodes and UE nodes respectively. Regard the channels between AP nodes and UE nodes as edges. This heterogeneous undirected bipartite graph is represented as where represents the node set, ε represents the edge set, and the node set is composed of the AP node set and the UE node set The AP node set is composed of M AP nodes, and the UE node set is composed of K UE nodes. The edge set ε is composed of M×K edges, and the elements in ε are denoted as (m, k);

[0012]

[0013] Define the feature matrix composed of all AP node feature vectors as where the node feature corresponding to the m-th AP is f m =P max ; where f1,..., f M are the M elements of the feature matrix F AP ; represents the set of all M×1-dimensional matrices over the real number field;

[0014] Define the feature matrix composed of all UE node feature vectors as where the node feature corresponding to the k-th UE is f k =σ 2 ; f1,..., f K are the K elements of the feature matrix F UE ; is the set of all K×1-dimensional matrices over the real number field;

[0015] Define the feature matrix composed of all edge feature vectors as where the edge feature corresponding to edge (m, k) is e (m,k) = β m,k ; where β m,k is the large-scale fading coefficient of the channel between the m-th AP and the k-th UE, is the set of all M×K×1-dimensional matrices over the real number field;

[0016] Define the vector composed of the variables on all AP nodes as p m represents the total transmission power of the m-th AP; p1, …, p M is the vector p AP are the M elements of the vector; is the M-dimensional real space;

[0017] Define the matrix composed of the variables on all edges as represents the power ratio allocated by the m-th AP to the k-th UE;

[0018] Therefore, the power allocated by the m-th AP to the k-th UE is expressed as:

[0019] Furthermore, the neural network model of the Graphormer includes:

[0020] A preprocessing layer for using three different MLPs to map the AP node features, UE node features, and edge features from 1×1 dimension to 1×d dimension respectively;

[0021] An update layer for receiving the preprocessed feature vectors of the AP nodes, UE nodes, and edges transmitted by the preprocessing layer, and updating the node features of the AP and UE nodes respectively using the message passing mechanism and type encoding;

[0022] A postprocessing layer for receiving the updated node features transmitted by the update layer, processing the node features and edge features, and obtaining the power allocation result.

[0023] Furthermore, the encoding of the graph structure information includes using the following formula:

[0024]

[0025] The process of updating the AP node features is that the m-th AP node aggregates the features on all adjacent edges represents taking the first-order neighbor node operator for the m-th AP node, and the aggregation operator uses the average operation. After that, the obtained 1×d-dimensional feature vector is input Take The output and the 0th layer feature vector of the mth AP obtained by the preprocessing layer are added to the introduced 1×d type encoded feature vector to obtain the 1st layer feature vector of the final mth AP

[0026] The process of updating the UE node features is as follows Aggregate the features on all the edges adjacent to the kth UE node denotes taking the first-order neighbor node operator for the kth UE node. The aggregation operator uses the average operation. After that, the obtained 1×d dimensional feature vector is input Add the output and the 0th layer feature vector of the mth AP obtained by the preprocessing layer and the introduced 1×d type encoded feature vector to obtain the 1st layer feature vector of the final kth UE

[0027] where δ AP , δ UE are different 1×d dimensional learnable vectors used to distinguish different node types and are the type encodings in the graph structure information encoding

[0028] The attention coefficient A with graph structure information encoding i,j is calculated as follows

[0029]

[0030] Among them, the first term is the traditional self-attention mechanism, which uses the 1×d dimensional query vector q i and the 1×d dimensional key vector k j to calculate the attention coefficient scalar based on the node features. The second term is the spatial encoding which is the attention coefficient scalar that can capture the spatial information between nodes. The third term is the edge feature encoding which is the attention coefficient scalar that can capture the edge information between nodes

[0031] is calculated as

[0032]

[0033] where ψ(v i , v j ) represents the shortest hop count between node i and node j; b1 and b2 are two learnable scalars

[0034] The edge feature encoding is calculated as

[0035]

[0036] Among them, (1, 2) is defined as the weighted shortest path between nodes i and j, where the edge weight is the large-scale fading coefficient. and are the feature vectors on edges 1 and 2 that make up the weighted shortest path. are two learnable 1×d-dimensional vectors.

[0037] Furthermore, the process of processing node features and edge features to obtain the power allocation result is as follows:

[0038] Input the node feature vector of the m-th AP into the MLP. Among them, the activation function of this MLP is Sigmoid. Then multiply the output of the MLP by the maximum transmit power P max to obtain the transmit power p of the m-th AP m ;

[0039] Concatenate the node feature vector of the m-th AP, the node feature vector of the k-th UE, and the edge feature vector of edge (m, k), and then input them into the MLP post to obtain p' m,k , where the activation function of the MLP is LeakyReLU.

[0040] Then perform Softmax on the M×K-dimensional matrix P' composed of p' m,k row by row to obtain the M×K-dimensional matrix The elements of this matrix are the power ratios allocated by the m-th AP to the k-th UE.

[0041] Finally, obtain the power p allocated by the m-th AP to the k-th UE m,k :

[0042] Furthermore, the neural network based on Graphormer is trained in an unsupervised learning manner, and the loss function is set by maximizing the spectrum efficiency of the cell-free network.

[0043] Furthermore, the process of using three different MLPs to map AP node features, UE node features, and edge features from 1×1 dimension to 1×d dimension respectively is as follows:

[0044]

[0045] is the feature of the m-th AP node after preprocessing, f AP,m represents the node feature of the m-th AP. represents the MLP used to process the AP node features in the preprocessing layer;

[0046] is the feature of the k-th UE node after preprocessing, f UE,k represents the node feature of the k-th UE, represents the MLP used to process the AP node features in the preprocessing layer;

[0047] is the feature of the edge (m, k) after preprocessing; MLP pre is used in the preprocessing layer for feature processing MLP;

[0048] After this operation, the AP node features, UE node features, and edge features have the same dimension and contain more abundant information.

[0049] The present invention proposes a cell-free network power control method based on Graphormer, which constructs the graph structure data of the cell-free network and designs a neural network architecture based on Graphormer to fully extract global and local information. Finally, the power control result is obtained by combining node and edge-level outputs. Compared with traditional optimization algorithms, the method proposed by the present invention reduces the computational complexity and provides a feasible method for power control in future large-scale networks.

[0050] Therefore, the present invention proposes a cell-free network power control method based on Graphormer, aiming to utilize the powerful feature learning ability of Graphormer to extract the spatial distribution and intensity of multi-user interference, and improve the downlink spectral efficiency of the cell-free network by optimizing the power control between APs and UEs. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0052] Figure 1 is the flowchart of the method of the present invention.

[0053] Figure 2 is the scenario diagram used in the embodiments of the present invention.

[0054] Figure 3 is the cell-free network graph structure data modeling diagram provided by the embodiments of the present invention.

[0055] Figure 4 is the algorithm flowchart provided by the embodiments of the present invention.

[0056] Figure 5 The neural network convergence curve graph provided by the embodiment of the present invention.

[0057] Figure 6 The cumulative distribution graph of user spectral efficiency provided by the embodiment of the present invention.

[0058] Figure 7 The box plot of the spectral efficiency of the cell-free network provided by the embodiment of the present invention. Detailed implementation manners

[0059] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data used can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0061] Figure 1 It is the flowchart of the method of the present invention.

[0062] As Figure 1 shown, the embodiment of the present invention provides a cell-free network power control method based on Graphormer, including the following steps:

[0063] A cell-free network power control method based on Graphormer, characterized by including the following steps:

[0064] S1. Construct a cell-free network adopting the time-division duplex mode, define an uplink training model for channel estimation, and define a downlink data transmission model to precode downlink data symbols based on the estimated channel;

[0065] S2: Based on the cell-free network, distributed access points, and user equipment, construct graph-structured data;

[0066] S3: Build a neural network model based on Graphormer;

[0067] S4: According to the constructed graph structure data, encode the graph structure information to fully extract the global and local information of the network, train the neural network based on Graphormer, and obtain a trained neural network model of Graphormer.

[0068] S5: Based on the transmit power of distributed access points, the received noise of user equipment, and the channel between distributed access points and user equipment, use the trained neural network model based on Graphormer to implement cell-free network power control.

[0069] Steps S1 / S2 / S3 / S4 / S5 are executed in sequence;

[0070] Figure 2 This is the scenario graph used in the embodiments of the present invention;

[0071] Furthermore, a cell-free network adopting a time-division duplex mode is constructed. Define an uplink training model for channel estimation and a downlink data transmission model to precode downlink data symbols using the estimated channel. The specific content is as follows:

[0072] S11: Consider a cell-free network composed of M distributed access points AP and K user equipments UE. Denote the set of APs, Denote the set of UEs. The APs and UEs are randomly distributed within the coverage area. The APs are connected to the CPU through a fronthaul network. It is assumed that the fronthaul network has infinite capacity and can transmit information without errors. Each AP is equipped with N antennas, and each UE is equipped with a single antenna. All APs use the same time-frequency resources to serve all UEs.

[0073] The cell-free network uses a time-division duplex protocol. Since the downlink data transmission is concerned, a coherence block is divided into two stages: uplink pilot training and downlink data transmission. First, in the uplink stage, each UE uploads a pilot sequence of length τ p , and channel estimation is performed at the AP. Then, in the downlink stage, the AP uses τ d =τ c -τ p samples to transmit UE data.

[0074] In a coherence block, the correlated Rayleigh fading channel between the m-th AP and the k-th UE is an N×1-dimensional complex vector, denoted as h m,k which follows a complex Gaussian distribution and can be expressed as:

[0075]

[0076] Among them, 0 N is an N×N all-zero matrix, is an N×N complex matrix. R m,k represents the spatial correlation matrix, which is determined by the channel gain between the m-th AP and the k-th UE and the spatial correlation between the N antennas equipped by the m-th AP. The channel gain depends on the path loss. Similar to the uncorrelated Rayleigh fading channel, the large-scale fading coefficient β mk is defined as:

[0077]

[0078] Among them, tr() is the trace operation on the matrix R m,k and N is the number of antennas equipped by the m-th AP.

[0079] Assume that the number of available orthogonal pilots τ p is less than the number of users K in the cell-free network, that is, there are τ p < K orthogonal pilot sequences assigned to K UEs. The allocation method of the pilot sequences is fixed and known, and this sequence satisfies:

[0080]

[0081] Among them, () H represents the Hermitian transpose operator, t k ∈{1,…,τ p} represents the pilot number assigned to the k-th UE. The set of UEs using the same pilot sequence as the k-th UE is defined as:

[0082]

[0083] S12, the pilot signal received by the m-th AP in an uplink stage can be expressed as:

[0084]

[0085] Among them, η k is the transmission power of the pilot sequence of the k-th UE, is the received noise of the m-th AP, and its elements follow a complex Gaussian distribution with a mean of 0 and a variance of σ 2 , that is Using the received pilot signal, the AP uses the minimum mean square error algorithm for channel estimation. The estimated channel of the k-th UE at the m-th AP can be expressed as:

[0086]

[0087] Among them, I N is an N×N dimensional identity matrix, and () -1 represents the matrix inversion operator, and () * represents the conjugate transpose operator.

[0088] S13. Denote the downlink data symbol s of the k-th UE as: k It is a complex scalar with an average power of 1, that is After the m-th AP receives the data symbol s of the k-th UE transmitted by the CPU and uses the N×1 dimensional precoding vector w k calculated locally to m,k precode s k , that is, multiply s m,k by w k , denoted as w m,k s k . The modified local partial minimum mean square error precoding algorithm is adopted, and the precoding vector w m,k is calculated according to the channel estimation at the m-th AP. For the k-th UE served by the m-th AP, the precoding vector is:

[0089]

[0090] Among them, the purpose of the first formula is to normalize it for subsequent power control. |||| is the L2 norm operator, and:

[0091]

[0092] represents the set of UEs with the largest channel gain in each pilot at the m-th AP, represents the set of users using the pilot φ ε . Z m,u represents the correlation matrix of the channel estimation error , and the expression of is: m,u Therefore, Z

[0093]

[0094] Denote the downlink transmission power allocated by the m-th AP to the k-th UE as p m,k ≥0, then the sum of the precoded signals of all UEs transmitted by the m-th AP is:

[0095]

[0096] Power control needs to meet the constraints:

[0097]

[0098] where P max is the maximum transmit power of each AP.

[0099] The signal received by the k-th UE has the following expression:

[0100]

[0101] where DS k represents the signal that the k-th user expects to receive, and MUI k represents the interference signal from other users to the k-th user. is the noise at the k-th UE, which follows a complex Gaussian distribution with a mean of 0 and a variance of σ 2 .

[0102] According to the above expression, in a coherence block, the downlink achievable spectral efficiency SE of the k-th UE k can be expressed as:

[0103]

[0104] where the instantaneous signal-to-interference-plus-noise ratio SINR k is:

[0105]

[0106] Under the above modeling considerations, the present invention designs a neural network based on Graphormer by means of power control under the constraint of limited downlink transmit power of each AP to maximize the downlink spectral efficiency of the cell-free network.

[0107] Furthermore, the graph-structured data includes topological relationships, features and variables of nodes and edges. The specific content is as follows:

[0108] Figure 3 is the modeling graph of the cell-free network graph-structured data provided by the embodiment of the present invention;

[0109] S21. Model the above cell-free network using a heterogeneous undirected bipartite graph, where APs and UEs are regarded as two types of nodes, namely AP nodes and UE nodes, and the channels between AP nodes and UE nodes are regarded as edges. This heterogeneous undirected bipartite graph can be expressed as where represents the node set, ε represents the edge set, and the node set consists of the AP node set and the set of UE nodes consists of. The set of AP nodes consists of M AP nodes, and the set of UE nodes consists of K UE nodes. The edge set ε consists of M×K edges. The elements in ε are denoted as (m, k). The above description can be expressed as:

[0110]

[0111] Define the feature matrix composed of all AP node feature vectors as where the node feature corresponding to the m-th AP is f m = P max ; where f1,…, f M are the M elements of the feature matrix F AP ; denotes the set of all M×1-dimensional matrices over the real number field;

[0112] Define the feature matrix composed of all UE node feature vectors as where the node feature corresponding to the k-th UE is f k = σ 2 ; f1,..., f K are the K elements of the feature matrix F UE , is the set of all K×1-dimensional matrices over the real number field;

[0113] Define the feature matrix composed of all edge feature vectors as where the edge feature corresponding to the edge (m, k) is e (m,k) = β m,k , where, β m,k is the large-scale fading coefficient of the channel between the m-th AP and the k-th UE, is the set of all M×K×1-dimensional matrices over the real number field;

[0114] Define the vector composed of all variables on the AP nodes as p m represents the transmission power of the m-th AP, p1,..., p M are the M elements of the vector p AP ; is the M-dimensional real space;

[0115] Define the matrix composed of all variables on the edges as whose elements are It represents the power ratio allocated by the m-th AP to the k-th UE. Therefore, the power allocated by the m-th AP to the k-th UE can be expressed as the product of the transmission power of the m-th AP and the power ratio allocated by the m-th AP to the k-th UE, that is

[0116] Figure 4 It is the algorithm flowchart provided by the embodiments of the present invention;

[0117] Furthermore, the construction of the neural network architecture based on Graphormer mainly includes a preprocessing layer, an update layer, and a postprocessing layer. According to the constructed graph data structure, three graph structure information encoding methods are designed to better extract the global and local information of the network. The neural network is trained using an unsupervised learning method, which improves the downlink spectrum efficiency of the cell-free network. The specific content is as follows:

[0118] The neural network model based on Graphormer includes:

[0119] A preprocessing layer, which is used to map the AP node features, UE node features, and edge features from 1×1 dimension to 1×d dimension respectively using three different MLPs;

[0120] An update layer, which is used to receive the preprocessed feature vectors of the AP nodes, UE nodes, and edges transmitted by the preprocessing layer, and update the node features of the AP and UE nodes respectively using the message passing mechanism and the graph structure information encoding;

[0121] A postprocessing layer, which is used to receive the updated node features transmitted by the update layer, process the node features and edge features, and obtain the power allocation result.

[0122] The construction of the preprocessing layer uses three different MLPs to map the AP node features, UE node features, and edge features from 1×1 dimension to 1×d dimension respectively:

[0123]

[0124] is the feature of the m-th AP node after preprocessing, f AP,m represents the node feature of the m-th AP, represents the MLP used to process the AP node features in the preprocessing layer;

[0125] is the feature of the k-th UE node after preprocessing, f UE,k represents the node feature of the k-th UE, represents the MLP used to process the AP node features in the preprocessing layer;

[0126] The feature of the edge (m, k) after preprocessing; MLP pre is the MLP used in the preprocessing layer to process the features;

[0127] After such operations, the AP node features, UE node features, and edge features have the same dimension and contain more abundant information, facilitating subsequent processing.

[0128] The construction of the update layer is divided into two steps. First, the AP nodes and UE nodes are used to update the node features respectively using the message passing mechanism and type encoding:

[0129]

[0130] The process of updating the AP node features is that the m-th AP node aggregates the features on all the edges adjacent to it represents taking the first-order neighbor node operator for the m-th AP node, and the aggregation operator uses the average operation. After that, the obtained 1×d-dimensional feature vector is input Take the output of and the 0-th layer feature vector of the m-th AP obtained from the preprocessing layer and the introduced 1×d type encoding feature vector are added together to obtain the final 1st layer feature vector of the m-th AP

[0131] The process of updating the UE node features is aggregating the features on all the edges adjacent to the k-th UE node, represents taking the first-order neighbor node operator for the k-th UE node, and the aggregation operator uses the average operation. After that, the obtained 1×d-dimensional feature vector is input Take the output of and the 0-th layer feature vector of the m-th AP obtained from the preprocessing layer and the introduced 1×d type encoding feature vector are added together to obtain the final 1st layer feature vector of the k-th UE

[0132] where, δ AP , δ UE are different 1×d-dimensional learnable vectors used to distinguish different node types, which are the type encodings in the graph structure information encoding. The updated AP node features and UE node features contain local information and can reflect the surrounding channel condition situation.

[0133] After that, the AP node features and UE node features are concatenated, and the edge features are retained to facilitate using Graphormer to continue updating the node features and making the node feature information more abundant. The process is as follows:

[0134]

[0135] Among them, is an (M + K)×d-dimensional real feature matrix composed of the first-layer feature vectors of M APs and the first-layer feature vectors of K UEs, is an M×K×d-dimensional real feature matrix composed of the first-layer feature vectors of M×K edges.

[0136] On this graph, the Graphormer layer is used to update the node features, and the architecture of the Graphormer layer can be expressed as:

[0137]

[0138] Among them, MHA is the multi-head attention mechanism, LN is the layer normalization operation, FFN is the feed-forward neural network, and l represents the l-th Graphormer layer. The above architecture is also the classic architecture of the Transformer layer. On this basis, the Graphormer layer introduces graph structure information encoding in the process of calculating the attention coefficients to capture the graph structure information between node i and node j and the relationship between node pairs, rather than only focusing on the semantic similarity between node pairs. Inspired by this, the present invention sets three graph structure information encoding methods according to the established graph structure data, and the above type of encoding is one of them;

[0139] The following introduces the other two encoding methods. First, use the elements of the attention coefficient matrix A ∈ (M+K)×(M+K) , that is, the attention coefficients, and their calculation method is as follows:

[0140]

[0141] Among them, the first term is the traditional self-attention mechanism, which uses a 1×d-dimensional query vector q i and a 1×d-dimensional key vector k j to calculate the attention coefficient scalar based on the node features. The second term is the spatial encoding. Specifically, the attention coefficient scalar that can capture the spatial information between nodes is calculated as:

[0142]

[0143] Among them, ψ(v i , v j ) represents the shortest hop count between node i and node j. According to the established heterogeneous undirected bipartite graph, the shortest hop count between any two nodes is at most 2 and at least 1, so only the two-hop range needs to be considered; b1 and b2 are two learnable scalars;

[0144] The third item is edge feature encoding, which is a scalar attention coefficient that can capture edge information between nodes The calculation method is as follows:

[0145]

[0146] Among them, (1, 2) is defined as the weighted shortest path between nodes i and j, where the edge weight is the large-scale fading coefficient and are the feature vectors on edge 1 and edge 2 that make up the weighted shortest path are two learnable 1×d dimensional vectors

[0147] The node features after being updated by the Graphormer layer, that is, the node features output by the update layer, can be expressed as:

[0148]

[0149] S33. Finally, in the post-processing layer, the node features and edge features are processed to obtain the power allocation result

[0150] First, the node feature vector of the m-th AP is input The activation function of this MLP is Sigmoid, and then the output of the MLP is multiplied by the maximum transmission power P of the m-th AP max to obtain the transmission power p of the m-th AP m ; then

[0151] The node feature vector of the m-th AP the node feature vector of the k-th UE and the edge feature vector of edge (m, k) are concatenated and input into the MLP post to obtain p' m,k , and the activation function of this MLP is LeakyReLU; then, the M×K dimensional matrix P' composed of p' m,k is Softmaxed row by row to obtain the M×K dimensional matrix The elements of this matrix are indicating the power ratio allocated by the m-th AP to the k-th UE; finally, the transmission power p of the m-th AP m is multiplied by the power ratio allocated by the m-th AP to the k-th UE

[0152] to obtain the final power p allocated by the m-th AP to the k-th UE m,k : That is, the power control result p m,k .

[0153] The above process can be expressed as:

[0154]

[0155] Among them, Sigmoid and LeakyReLU are the activation functions corresponding to the last layer of the MLP, Softmax is performed row by row, and the above two activation functions ensure that the transmission power p of the m-th AP m is greater than or equal to 0 and less than or equal to the maximum transmission power P max and also ensure that the sum of the transmission powers allocated by the m-th AP to all UEs is 1, that is:

[0156]

[0157] S34. After constructing the neural network architecture based on Graphormer, prepare the graph structure data required for training and testing, and use the unsupervised learning method to train the neural network model to maximize the spectrum efficiency of the cell-free network. The loss function Loss of the neural network is set to the negative of the sum of the downlink spectrum efficiencies of all UEs:

[0158]

[0159] The neural network based on Graphormer is trained by the unsupervised learning method, and the loss function is set by maximizing the spectrum efficiency of the cell-free network.

[0160] Simulation conditions

[0161] In the simulation scenario, APs and UEs are randomly distributed in a rectangular area of 960m * 960m. The local scattering model is used to model the spatial correlation between antennas at the AP, and a pilot allocation method is used to avoid UEs with close distances using the same pilot. The pilot transmission power is 200mW, and the maximum downlink transmission power of the AP is 1000mW. The number of APs is 16, the number of users is 16, each AP is equipped with 4 antennas, and each UE is equipped with a single antenna.

[0162] Simulation content and result analysis

[0163] Simulation 1: Train the neural network based on Graphormer with the training set and verify it with the test set data.

[0164] As Figure 5 shown, construct the system model shown as Figure 2 into as Figure 3For the graph-structured data shown, the graph-structured data is input into a neural network based on Graphormer, the neural network is trained using a training set, and verified using a test set. The number of cell-free networks in the training set and the test set are 2048 and 512 respectively. After 20 training cycles, the neural network shows good convergence.

[0165] Simulation 2: Compare the method proposed in the present invention with the weighted least mean square error power control algorithm, the equal power control algorithm, and the fractional power control algorithm, as Figure 6 and Figure 7 It can be seen that the proposed cell-free network power control method based on Graphormer has better performance than the equal power control algorithm and the fractional power control algorithm, and approaches the weighted least mean square error power control algorithm. This verifies the effectiveness of the method proposed in the present invention.

[0166] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A Graphormer-based cell-free network power control method, characterized in that: The steps include: Build a cell-free network using time division duplex mode, define an uplink training model for channel estimation, and define a downlink data transmission model to precode downlink data symbols based on the estimated channel; Construct graph structure data based on cell-free networks, distributed access points and user devices; Build a neural network model based on Graphormer; According to the constructed graph structure data, the graph structure information is encoded to fully extract the global and local information of the network, and the Graphormer-based neural network is trained to obtain a trained Graphormer-based neural network model. Based on the transmission power of distributed access points, the receiving noise of user equipment, and the channel between distributed access points and user equipment, a trained Graphormer-based neural network model is used to implement power control in a cell-free network.

2. A Graphormer-based cell-free network power control method according to claim 1, characterized in that: The graph structure data includes topological relationships, node and edge features and variables, as follows: The cell-free network is modeled using a heterogeneous undirected bipartite graph, where the distributed access point AP and the user equipment UE are regarded as two types of nodes, namely AP nodes and UE nodes, and the channel between the AP node and the UE node is regarded as an edge. This heterogeneous undirected bipartite graph is represented as in, represents a node set, ε represents an edge set, and the node set By AP node collection and UE node set Composition, AP node set It consists of M AP nodes and UE nodes. It consists of K UE nodes, the edge set ε consists of M×K edges, and the elements in ε are denoted as (m,k); The feature matrix composed of all AP node feature vectors is defined as The node feature corresponding to the mth AP is f m =P max ; where f1,…,f M is the feature matrix F AP M elements of; Represents the set of all M×1-dimensional matrices in the real number field; The feature matrix composed of all UE node feature vectors is defined as The node feature corresponding to the kth UE is f k =σ 2 ; f1,…,f K is the feature matrix F UE The K elements of is the set of all K×1 dimensional matrices in the real number field; The characteristic matrix composed of all edge eigenvectors is defined as The edge feature corresponding to the edge (m, k) is e (m,k) =β m,k ; Among them, β m,k is the large-scale fading coefficient of the channel between the mth AP and the kth UE, is the set of all M×K×1 matrices in the real number field; Define the vector consisting of variables on all AP nodes as p m represents the total transmit power of the mth AP; p1,…,p M is the vector p AP M elements of; is an M-dimensional real number space; Define the matrix consisting of all the variables on the edge as Indicates the power ratio allocated by the m-th AP to the k-th UE; Therefore, the power allocated by the mth AP to the kth UE is expressed as:

3. A Graphormer-based cell-free network power control method according to claim 1, characterized in that: The Graphormer-based neural network model includes: The preprocessing layer is used to map the AP node features, UE node features, and edge features from 1×1 dimensions to 1×d dimensions using three different MLPs; An updating layer, used for receiving the preprocessed feature vectors of the AP node, UE node and edge transmitted by the preprocessing layer, and updating the node features of the AP and UE nodes respectively using a message passing mechanism and type coding; The post-processing layer is used to receive the updated node features transmitted by the update layer, process the node features and edge features, and obtain the power allocation result.

4. The Graphormer-based cell-free network power control method according to claim 1, characterized in that: The encoding of the graph structure information includes using the following formula: The process of updating the AP node features is that the mth AP node aggregates the features of all its adjacent edges. It means taking the first-order neighbor node operator for the mth AP node, and the aggregation operator uses the average operation. Then, the obtained 1×d-dimensional feature vector is input into Will The output and the 0th layer feature vector of the mth AP obtained by the preprocessing layer And the introduced 1×d type encoding feature vector is added to obtain the final first-layer feature vector of the mth AP The process of updating UE node characteristics is as follows: Aggregate the features of all edges adjacent to the kth UE node. It means taking the first-order neighbor node operator for the kth UE node, and the aggregation operator uses the average operation. Then, the obtained 1×d-dimensional feature vector is input into Will The output and the 0th layer feature vector of the mth AP obtained by the preprocessing layer And the introduced 1×d type coding feature vector is added to obtain the final first layer feature vector of the kth UE where δ AP , δ UE are different 1×d-dimensional learnable vectors used to distinguish different node types and are type codes in the graph structure information encoding; Attention coefficient A with graph structure information encoding i,j The calculation method is as follows: The first one is the traditional self-attention mechanism, using a 1×d-dimensional query vector q i and a 1×d-dimensional key vector k j The attention coefficient scalar based on node features is calculated, and the second term is the spatial encoding. is the attention coefficient scalar that can capture the spatial information between nodes, and the third term is the edge feature encoding; is the attention coefficient scalar that can capture the edge information between nodes; The calculation method is: Among them, ψ(v i ,v j ) represents the shortest hop count between node i and node j; b1, b2 are two learnable scalars; Edge feature encoding, which is calculated as: Among them, (1,2) is defined as the weighted shortest path between nodes i and j, where: the edge weight is the large-scale fading coefficient, and is the feature vector on edge 1 and edge 2 that constitutes the weighted shortest path, are two learnable 1×d dimensional vectors.

5. The Graphormer-based cell-free network power control method according to claim 1, characterized in that: The process of processing node features and edge features to obtain the power allocation result is as follows: The node feature vector of the mth AP is input into the MLP, where the activation function of the MLP is Sigmoid, and then the output of the MLP is multiplied by the maximum transmission power P max Get the transmit power p of the mth AP m ; The node feature vector of the mth AP, the node feature vector of the kth UE, and the edge feature vector of edge (m, k) are concatenated and input into the MLP post Get p' m,k , where the activation function of MLP is LeakyReLU, which will be replaced by p' m,k The M*K dimensional matrix P' composed of Sofmax is performed row by row to obtain an M*K dimensional matrix The elements of this matrix are the power ratios allocated by the mth AP to the kth UE. Finally, the power p allocated by the mth AP to the kth UE is obtained. m,k :

6. The Graphormer-based cell-free network power control method according to claim 1, characterized in that: The Graphormer-based neural network is trained by unsupervised learning, and the loss function is set by maximizing the spectrum efficiency of the cell-free network.

7. The Graphormer-based cell-free network power control method according to claim 1, characterized in that: The process for using three different MLPs to map AP node features, UE node features, and edge features from 1×1 dimensions to 1×d dimensions is as follows: is the node feature of the mth AP after preprocessing, f AP,m represents the node characteristics of the mth AP, represents the MLP used to process AP node features in the preprocessing layer; is the node feature of the kth UE after preprocessing, f UE,k represents the node characteristics of the kth UE, represents the MLP used to process AP node features in the preprocessing layer; is the feature of edge (m,k) after preprocessing; MLP pre is used in the preprocessing layer to (m,k) The MLP for feature processing; ε represents the set of edges (m, k); After this operation, the dimensions of AP node features, UE node features, and edge features are the same, and the information they contain is richer.

Images