Link scheduling method for heterogeneous wireless D2D networks based on graph neural network

Abstract

This invention discloses a link scheduling method for heterogeneous wireless D2D networks based on graph neural networks, comprising: constructing a heterogeneous wireless D2D network; determining a training set and a test set; modeling the heterogeneous wireless D2D network within the training set and test set as a heterogeneous graph; constructing a heterogeneous graph convolutional neural network model and initializing the neural network weights; training the heterogeneous graph convolutional neural network model using the training set and validating it using the test set to obtain a trained heterogeneous graph convolutional neural network; inputting data into the trained heterogeneous graph convolutional neural network and outputting a link scheduling strategy. This invention is used for wireless resource scheduling in D2D networks to maximize the transmission rate performance of heterogeneous D2D networks.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication networks, and in particular to a heterogeneous wireless D2D network link scheduling method based on graph neural network. Background Art

[0002] D2D (Device to Device) technology is one of the key technologies in the development of 5G wireless communication. D2D technology enables user communication devices within a certain distance to communicate directly, aiming to reduce the load of the service base station. There are three types of D2D communication methods. The first is the dedicated mode, where D2D pairs exchange data directly without involving base stations. The spectrum resources allocated in this transmission mode are dedicated to a certain pair and are not shared with other users. The second type is the shared mode, in which cellular users and D2D users, and D2D users and D2D users can share the same spectrum resources. The third type is the cellular mode, just like traditional cellular communication, all users transmit and receive data through the core network without applying direct communication between users. Devices that are close to each other use D2D technology to communicate, which can reduce latency, power consumption and cost between devices. D2D technology can also establish an emergency communication network in a short time and replace damaged infrastructure. It is an effective solution for building emergency networks. D2D technology can also be applied to cellular networks to improve spectrum utilization and network capacity. Although D2D technology has many advantages, on the other hand, it also faces many challenges. In the sharing mode, since the same resources are shared in the same area, users using D2D technology may cause serious interference, resulting in a decrease in the overall network transmission rate. Appropriate wireless link scheduling is required to balance the interference between D2D devices and maximize the network data transmission rate. Wireless link scheduling mainly refers to shutting down some devices in the communication network without affecting the overall communication, thereby achieving the purpose of reducing interference and improving network transmission efficiency. In recent years, related machine learning technologies have been used for link scheduling problems. For example, convolutional neural networks (CNNs) are used to learn the mapping of the distance between the transmitter and the receiver to the scheduling strategy in an unsupervised manner to obtain near-optimal performance.

[0003] Most of these machine learning methods are based on homogeneous wireless communication networks. Due to the different types of D2D devices in heterogeneous wireless D2D networks, the communication between devices may use different numbers of antennas. Traditional machine learning techniques cannot extract heterogeneous network features and are difficult to promote in heterogeneous networks. Therefore, it is necessary to provide an improved technical solution to address the deficiencies in the above-mentioned prior art. Summary of the invention

[0004] The purpose of the present invention is to provide a heterogeneous wireless D2D network link scheduling method based on graph neural network, which is used for wireless resource scheduling of D2D networks and maximizes the transmission rate performance of heterogeneous D2D networks.

[0005] The technical solution to achieve the above purpose is:

[0006] A heterogeneous wireless D2D network link scheduling method based on graph neural network includes:

[0007] Build heterogeneous wireless D2D networks;

[0008] Determine the training set and test set;

[0009] The heterogeneous wireless D2D networks in the training set and the test set are modeled as heterogeneous graphs;

[0010] Build a heterogeneous graph convolutional neural network model and initialize the neural network weights;

[0011] The heterogeneous graph convolutional neural network model is trained using the training set and verified using the test set to obtain a trained heterogeneous graph convolutional neural network;

[0012] Input data to the trained heterogeneous graph convolutional neural network and output the link scheduling strategy.

[0013] Preferably, the construction of a heterogeneous wireless D2D network includes:

[0014] On a side of length d area In the square area of ​​, D heterogeneous D2D pairs are generated; M = {1, 2, 3, ..., m} represents the type of D2D pair, D = {D1, D2, D3, ..., D m} represents the total number of D2D pairs; where D m represents the total number of D2D pairs of type m; represents the i-th D2D pair of type m, each D2D pair has a transmitter and a receiver Calculate the channel response of the communication link as and the channel response of the interfering link is

[0015] Preferably, the determining of the training set and the test set comprises:

[0016] Generate X and Y heterogeneous wireless D2D networks respectively, and run the FP algorithm to generate the optimal sum rate label for each heterogeneous wireless D2D network. The generated heterogeneous wireless D2D networks and the optimal sum rate labels are used as training sets and test sets. The number of samples in the training set is X, and the number of samples in the test set is Y.

[0017] Preferably, the heterogeneous wireless D2D network in the training set and the test set is modeled as a heterogeneous graph, including: considering each D2D pair in the heterogeneous wireless D2D network as a vertex and the interference link as an edge of the graph, and modeling the heterogeneous wireless D2D network as a heterogeneous graph.

[0018] Preferably, the objective function of the heterogeneous graph convolutional neural network model is:

[0019]

[0020] The constraints are:

[0021] It means that for every D2D pair of any type, there is a state vector ρ∈{0,1};

[0022] in, is the priority of the D2D pair, is the state vector; B is the communication bandwidth; It is a transmitter and receiver The channel response of the communication link; yes Transmitter With D2D Receiver The channel response between is the transmission power of the transmitter; is the environmental noise; ρ is the state vector of the D2D pair, ρ = 1 indicates that the D2D pair is activated, and ρ = 0 indicates that the D2D pair is closed; represents the state vector of the j-th D2D pair of type n, ρ im represents the state vector of the i-th D2D pair of type m; p im is the transmitter T of the i-th D2D pair of type m im The transmission power; p jn is the transmitter T of the jth D2D pair of type n jn The transmission power; Indicates the jammer T im To the interference receiver R jn The channel response between them is the channel response of the interfering link.

[0023] Preferably, the framework of the heterogeneous graph convolutional neural network model is 2 layers of heterogeneous graph convolutional layers, each of which is updated by an edge function Vertex Update Function and aggregate functions Composition, edge update function and the vertex update function is written as:

[0024]

[0025]

[0026] in, is the initialized edge feature, l is the number of layers of the graph convolutional network; is the edge feature of the graph model in the l-th layer network; It is the node feature after partial update in the l-th layer network; is node D in the l-1 layer network im Characteristics of neighbor nodes; is node D in the l-1 layer network im The node characteristics of It is the edge update function, which is used to transmit neighbor node information and edge information; It is a node partial update function, which is used to update the information passed by each type of node; the model is based on the node The type of neighbor nodes divides the graph into different network blocks, and different network blocks correspond to different network relationships, and the type of interference link is used to distinguish such relationships. Define the relationship r = (m, n) to represent the interference link from n-type D2D pairs to m-type D2D pairs; for each relationship, the model defines an edge update function parameterized by a multilayer perceptron and the vertex update function After learning the node updates for each relationship, we then aggregate the nodes to get their final updates; aggregation function:

[0027]

[0028] For Node Partial update of c I,m For Node The number of interfering links; Represented as the features of the nodes in the lth layer of the network; the heterogeneous graph is input into the heterogeneous graph convolutional neural network, and the final output layer is activated by the softmax function to generate a link scheduling strategy

[0029] Preferably, the heterogeneous graph convolutional neural network model is trained using a training set and verified using a test set to obtain a trained heterogeneous graph convolutional neural network, including:

[0030] Input the processed training set into the heterogeneous graph convolutional neural network, construct the negative expectation of the network utility function as the loss function, and select a gradient descent algorithm to optimize the weights of the neural network;

[0031] Set the number of iterations, save the training model every 5 iterations, input the test set for verification, and obtain the trained heterogeneous graph convolutional neural network after passing the test.

[0032] Preferably, the loss function is:

[0033]

[0034] in, for Hot spot representatives; for Hot spot representative. H [] indicates the expectation of calculating the formula in brackets; p jn is the transmitter T of the jth D2D pair of type n jn The transmission power is 200 Ω; a small batch gradient descent algorithm is used, each batch contains Q samples, the number of iterations is set, and the Adam optimizer is selected to optimize the parameters θ of the neural network.

[0035] The beneficial effects of the present invention are as follows: the present invention introduces the link scheduling problem into a more practical heterogeneous environment, and overcomes the serious interference problem caused by too many D2D communication links in the heterogeneous environment. The heterogeneous graph neural network is good at learning heterogeneous graph information, and is applied to heterogeneous wireless D2D networks to learn the optimal link scheduling strategy in an unsupervised learning manner; the trained neural network can make online decisions, and provide a high-performance, low-complexity and reliable link scheduling solution in real time in a heterogeneous environment, thereby maximizing the D2D network transmission rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the heterogeneous wireless D2D network link scheduling method of the present invention;

[0037] Figure 2 is a heterogeneous graph convolutional neural network graph used in the present invention;

[0038] Figure 3 is a heterogeneous D2D network diagram in the present invention;

[0039] Figure 4 It is a heterogeneous graph modeled by a heterogeneous D2D network in the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings.

[0041] See also Figure 1 The heterogeneous wireless D2D network link scheduling method of the present invention is based on a graph neural network, such as Figure 2As shown, the graph neural network is a heterogeneous graph convolutional neural network, which is responsible for learning the heterogeneous graph model to obtain the optimal link scheduling strategy. It includes the following steps:

[0042] Step 1: Build a heterogeneous wireless D2D network. area In the square area, D heterogeneous D2D pairs are generated. M = {1, 2, 3, ..., m} represents the type of D2D pair, and D = {D1, D2, D3, ..., D m} represents the total number of D2D pairs, where D m represents the total number of D2D pairs of type m, represents the i-th D2D pair of type m. Figure 3 As shown, the side length d area =400m square area, M = 12 D2D pairs are randomly generated, including 4 2×1 MISO (multiple input single output) link type D2D pairs and 8 SISO (single input single output) link type D2D pairs. The transmitters of each D2D pair are evenly distributed in the area, and the corresponding receivers are distributed in a disk centered on the transmitter, with a uniform pair distance between 2m and 50m. Large-scale fading and small-scale fading are used as channel models, where large-scale fading does not consider shadowing effects. Channel response Small-scale decline It is represented by a zero-mean complex Gaussian variable with unit variance. Large-scale fading The scaled range correlation model is used for determination. The channel bandwidth is set to 5000 MHz. The noise variance at the receiver and the transmit power of the transmitter are normalized to 1.

[0043] The heterogeneous feature is that the number of transmitter antennas of each type of D2D pair may be different, N m represents the number of D2D transmitting antennas of type m. The transmitter and receiver are represented by and Each D2D pair The distance between the transmitter and the receiver is d max and d min Calculate the channel response of the communication link and the channel response of the interfering link

[0044] Step 2: Generate training and test sets. Generate 300,000 and 1,000 heterogeneous wireless D2D networks in step 1 respectively, and run the FP (fractional programming) algorithm to generate the optimal sum rate label for each heterogeneous wireless D2D network. The generated heterogeneous wireless D2D networks and optimal sum rate labels are used as training sets, with the number of samples in the training set X = 300,000 and the number of samples in the test set Y = 1,000.

[0045] Step 3: Model the heterogeneous wireless D2D network in the training set and the test set as a heterogeneous graph. Each D2D pair in the heterogeneous wireless D2D network is regarded as a vertex, and the interference link is regarded as the edge of the graph. The heterogeneous wireless D2D network is modeled as a heterogeneous graph. Represents a collection of node types, Represents a set of edges. A heterogeneous graph is represented by two mapping functions, which map each node to the corresponding type of φ: And map each edge to the corresponding type like Figure 4 As shown, in a heterogeneous D2D network, each D2D pair is regarded as a node and each interfering link as an edge. is the i-th node of type m. For Node Neighbor nodes of type n. The properties of the nodes are obtained through the collection Indicates that the properties of the i-th D2D pair of type m are The attributes of each vertex include weight Noise variance and the direct channel response of the transceiver The attributes of the edge are obtained by the set express, Each edge (i m ,j n ) is characterized by the channel response from the jammer to the jammer receiver, i.e., is the jammer T im To the interference receiver R jn The channel response between is the jammer T jn To the interference receiver R im The channel response between .

[0046] Step 4: Build a heterogeneous graph neural network model and initialize the neural network weights. Figure 2 As shown, each graph convolution layer is updated by the edge function Vertex Update Function and aggregate functions Composition, edge update function and the vertex update function can be written as:

[0047]

[0048]

[0049] in, is the initialized edge feature, and l is the number of layers of the graph convolutional network. is the edge feature of the graph model in the l-th layer network; It is the node feature after partial update in the l-th layer network; is node D in the l-1 layer network im Characteristics of neighbor nodes; is node D in the l-1 layer network im The node characteristics of It is the edge update function, which is used to transmit neighbor node information and edge information; It is a node partial update function, which is used to update the information passed by each type of node. The type of neighbor nodes divides the graph into different network blocks. Different network blocks correspond to different network relationships, and the type of interference link is used to distinguish such relationships. Define the relationship r = (m, n) to represent the interference link from the n-type D2D pair to the m-type D2D pair. For each relationship, the model defines an edge update function parameterized by a multi-layer perceptron (MLP) and the vertex update function After learning the node updates for each relationship, we then aggregate the nodes to get the final updates. Aggregation function:

[0050]

[0051] For Node Partial update of c I,m For Node The number of interfering links; The invention uses a two-layer heterogeneous graph convolutional neural network, and the edge update function of the first layer is and the vertex update function By the size of {4N n +2N m ,32,32,8} and {8+2N m ,32,32,8} MLP parameterization. The edge update function of the second layer and the vertex update function By the size {4N n +2N m ,32,32,8} and {8+2Nm ,32,32,2} MLP parameterization. The heterogeneous D2D network in the training set is modeled as a heterogeneous graph model and input into the heterogeneous graph convolutional neural network. The final output layer is activated by the softmax (normalized exponential function) function to generate a link scheduling strategy.

[0052] The entire model rate maximization problem can be expressed as:

[0053]

[0054] The constraints are:

[0055] It means that for every D2D pair of any type, there is a state vector ρ∈{0,1}.

[0056] in, is the priority of the D2D pair, is the state vector; B is the communication bandwidth; It is a transmitter and receiver The channel response of yes Transmitter With D2D Receiver The channel response between is the transmission power of the transmitter; is the environmental noise. ρ is the state vector of the D2D pair, ρ = 1 indicates that the D2D pair is activated, and ρ = 0 indicates that the D2D pair is closed; ρ jn represents the state vector of the j-th D2D pair of type n, ρ im represents the state vector of the i-th D2D pair of type m; p im is the transmitter T of the i-th D2D pair of type m im The transmission power of jn is the transmitter T of the jth D2D pair of type n jn The transmission power; Indicates the jammer T im To the interference receiver R jn The channel response between them is the channel response of the interference link.

[0057] Step 5: Input the processed training set into the neural network to construct the negative expectation of the utility function as the loss function, and select a gradient descent algorithm to optimize the weights of the neural network. Set the number of training iterations to 200, save the training model every 5 iterations, and then input the test set to test the performance. The test result is the average of 1000 test networks, and the test result is the ratio of the sum rate generated after the neural network training to the optimal sum rate label generated by the FP algorithm. The training ends after 200 iterations.

[0058] In the training phase, the negative expectation of the utility function is selected as the loss function:

[0059]

[0060] in, for Hot spot representatives; for Hot spot representative. H [] indicates the expectation of calculating the formula in brackets. jn is the transmitter T of the jth D2D pair of type n jn The transmission power of the neural network is θ. The training set is divided into 4500 batches using the mini-batch gradient descent algorithm, each batch contains 64 samples, the number of iterations is 200, and the adam (adaptive moment estimation) optimizer is selected to optimize the neural network parameter θ.

[0061] In the testing phase, the test data set is input into the trained heterogeneous graph convolutional neural network to verify that its sum rate performance is more than 80% of the labels generated by the FP algorithm, proving the reliability of the present invention.

[0062] Figure 3 In the figure, Communication Link indicates a communication link; Interference Link indicates an interference link; Figure 4 In the diagram, Vertex and edge mean node and edge.

[0063] The above embodiments are only used to illustrate the present invention, rather than to limit the present invention. Those skilled in the relevant technical field may make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions should also belong to the scope of the present invention and should be defined by the claims.

Claims

1. A heterogeneous wireless D2D network link scheduling method based on graph neural network, characterized in that: include: Build heterogeneous wireless D2D networks; Determine the training set and test set; The heterogeneous wireless D2D networks in the training set and the test set are modeled as heterogeneous graphs; Build a heterogeneous graph convolutional neural network model and initialize the neural network weights; The heterogeneous graph convolutional neural network model is trained using the training set and verified using the test set to obtain a trained heterogeneous graph convolutional neural network; Input data to the trained heterogeneous graph convolutional neural network and output the link scheduling strategy; The construction of a heterogeneous wireless D2D network includes: On a side of length d arra In the square area of ​​, D heterogeneous D2D pairs are generated; M = {1, 2, 3, ..., m} represents the type of D2D pair, D = {D1, D2, D3, ..., D m } represents the total number of D2D pairs; where D m represents the total number of D2D pairs of type m; represents the i-th D2D pair of type m, each D2D pair has a transmitter and a receiver The determination of the training set and the test set comprises: Generate X and Y heterogeneous wireless D2D networks respectively, and run the FP algorithm to generate the optimal sum rate label for each heterogeneous wireless D2D network. The generated heterogeneous wireless D2D networks and the optimal sum rate label are used as training sets and test sets. The number of samples in the training set is X, and the number of samples in the test set is Y. Modeling the heterogeneous wireless D2D network in the training set and the test set as a heterogeneous graph, including: considering each D2D pair of the heterogeneous wireless D2D network as a vertex, and the interference link as an edge of the graph, and modeling the heterogeneous wireless D2D network as a heterogeneous graph; The objective function of the heterogeneous graph convolutional neural network model is: The constraints are: It means that for every D2D pair of any type, there is a state vector ρ∈{0,1}; in, is the priority of the D2D pair, is the state vector; B is the communication bandwidth; It is a transmitter and receiver The channel response of the communication link; yes Transmitter With D2D Receiver The channel response between is the transmission power of the transmitter; is the environmental noise; ρ is the state vector of the D2D pair, ρ = 1 indicates that the D2D pair is activated, and ρ = 0 indicates that the D2D pair is closed; represents the state vector of the j-th D2D pair of type n, ρ im represents the state vector of the i-th D2D pair of type m; p im is the transmitter T of the i-th D2D pair of type m im The transmission power of jn is the transmitter T of the jth D2D pair of type n jn The transmission power; Indicates the jammer T im To the interference receiver R jn The channel response between is the jammer T jn To the interference receiver R im The channel response between The framework of the heterogeneous graph convolutional neural network model is 2 layers of heterogeneous graph convolutional layers, each of which is updated by an edge function Vertex Update Function and aggregate functions Composition, edge update function and the vertex update function is written as: in, is the initialized edge feature, l is the number of layers of the graph convolutional network; is the edge feature of the graph model in the l-th layer network; It is the node feature after partial update in the l-th layer network; is node D in the l-1 layer network im Characteristics of neighbor nodes; is node D in the l-1 layer network im The node characteristics of It is the edge update function, which is used to transmit neighbor node information and edge information; It is a node partial update function, which is used to update the information passed by each type of node; the model is based on the node The type of neighbor nodes divides the graph into different network blocks, and different network blocks correspond to different network relationships. The type of interference link is used to distinguish such relationships. The relationship r = (m, n) is defined to represent the interference link from the n-type D2D pair to the m-type D2D pair. For each relationship, the model defines an edge update function parameterized by a multilayer perceptron. and the vertex update function After learning the node updates for each relationship, we then aggregate the nodes to get their final updates; aggregation function: For Node Partial update of c I,m For Node The number of interfering links; Represented as the features of the nodes in the lth layer of the network; the heterogeneous graph is input into the heterogeneous graph convolutional neural network, and the final output layer is activated by the softmax function to generate a link scheduling strategy 2. The heterogeneous wireless D2D network link scheduling method based on graph neural network according to claim 1 is characterized in that: The heterogeneous graph convolutional neural network model is trained using the training set and verified using the test set to obtain the trained heterogeneous graph convolutional neural network, including: Input the processed training set into the heterogeneous graph convolutional neural network, construct the negative expectation of the network utility function as the loss function, and select a gradient descent algorithm to optimize the weights of the neural network; Set the number of iterations, save the training model every 5 iterations, input the test set for verification, and obtain the trained heterogeneous graph convolutional neural network after passing the test.

3. The heterogeneous wireless D2D network link scheduling method based on graph neural network according to claim 2 is characterized in that: Loss function: in, for Hot spot representatives; for Hot spot representative; E H [] indicates the expectation of calculating the formula in brackets; is the transmitter T of the jth D2D pair of type n jn The transmission power is 200 Ω; a small batch gradient descent algorithm is used, each batch contains Q samples, the number of iterations is set, and the Adam optimizer is selected to optimize the parameters θ of the neural network.

Images