Mode selection method and device for D2D auxiliary power internet of things

Abstract

The invention discloses a mode selection method and device for a D2D auxiliary power internet of things, and the method comprises the steps: building a graph neural network model, modeling the interference relation between D2D user pairs and cellular users in a power network and channels into a graph structure in a graph building module, coding the node features and edge features in the graph structure into feature vectors, and carrying out the coding of the feature vectors, in a graph convolution module, a graph convolution neural network is utilized to perform feature extraction on the feature vectors to obtain node hidden states including D2D user pair nodes and cellular user nodes, and in a classification decision module, the node hidden states are input into a classifier to calculate channel selection variables so as to determine a selected channel mode; and training the graph neural network model and then deploying and applying. According to the method, the pattern selection of D2D communication is optimized and the system throughput is maximized by using the graph neural network, so that the calculation complexity is further reduced, the calculation speed is improved, and the overall performance of the communication system can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a mode selection method and device for a D2D-assisted power Internet of Things. Background Art

[0002] Device-to-Device (D2D) communication technology is the fifth generation (the 5 th D2D is one of the key technologies of 5G (5th Generation) mobile communication technology. D2D refers to a communication mode in which terminal users communicate directly without the need for forwarding by the base station. D2D communication can use both authorized and unauthorized frequency band resources, so it can solve the problem of wireless resource shortage. It can improve the system frequency band utilization and increase the information transmission rate. User terminals that perform D2D communication can transmit information at low power, effectively saving energy consumption, and the transmission delay is very low when transmitting information between users. The reduction in transmission delay helps to increase the transmission rate. In addition, terminal direct communication is not controlled by the base station, so the base station load is small.

[0003] D2D communication has many advantages, but the introduction of D2D communication in cellular networks also brings many problems to the system. When D2D users use unlicensed frequency bands, it is difficult to discover devices and control user terminals, resulting in a decrease in network performance; when D2D users use licensed frequency bands, the resource utilization of the system can be improved, but at the same time, the original network will introduce new interference. This interference caused by the reuse of cellular user resources will affect the communication quality of the network and make the network environment worse. In order to improve system performance, previous studies have focused on two key technologies in D2D technology: mode selection technology and power allocation technology. Among them, mode selection technology represents the working mode selected by D2D users. Different working modes occupy different frequency bands, which solves the problem of frequency band resource allocation well; power control technology adjusts the transmission power of the user transmitter according to the communication task and user needs, thereby reducing the interference generated when reusing resources. The two technologies can be jointly modeled in an optimization problem framework, but the previous methods often have too much computational effort and are of little value in actual system applications.

[0004] D2D communication modes are divided into multiplexing mode and dedicated mode. In the multiplexing mode, D2D users and cellular users share a spectrum, which can effectively save system spectrum resources, but communication interference will occur between cellular users and D2D users; in the dedicated mode, the spectrum resources used by D2D users are orthogonal to the spectrum resources of cellular users, so there is no interference, but a large amount of spectrum resources will be occupied. Therefore, how to maximize the system throughput while balancing spectrum resources and communication interference becomes the core issue of mode selection. The rationality of mode selection determines the performance of the network. At present, the academic community is increasingly conducting research on D2D mode selection technology. Yang Wujun et al. proposed a mode selection method based on user distance. By distinguishing users according to communication intervals and setting priorities according to user channel states, the mode selection for users was effectively improved. Ma Bei et al. proposed a method based on incomplete channel state information (CSI). The service quality of different users was used as a constraint to obtain the minimum power closed solution of D2D users and cellular communication links, and the mode selection scheme was derived from this. Liu Jinxin et al. proposed a mode selection scheme based on signal to interference plus noise ratio (SINR). The downlink was defined as the interference area, and SINR was used as a constraint to quickly select the channel scheme by solving the interference suppression method. In terms of solution algorithms, this type of research mainly focuses on mathematical models based on convex optimization, and uses traditional algorithms such as implicit enumeration and Hungarian algorithm to solve. This type of algorithm is often not conducive to practical application. The reasons are mainly reflected in two aspects: First, each time a channel is allocated, an algorithm optimization needs to be performed, which leads to a very high complexity of each optimization. Second, with the increase of network scale, the complexity of optimization increases exponentially. Therefore, it is necessary to further optimize the mode selection of D2D communication to achieve performance superior to traditional algorithms. Summary of the invention

[0005] In view of the above, the purpose of the present invention is to provide a mode selection method and device for D2D-assisted power Internet of Things, which innovatively introduces a graph neural network (GNN) architecture to solve the channel allocation problem, and uses the graph neural network to optimize the mode selection of D2D communication and maximize the system throughput, further reduce the computational complexity and improve the computational speed, greatly improve the overall performance of the communication system, have high practical application value in solving the problems of spectrum resource shortage and communication interference, and provide an efficient solution for future power Internet of Things communication systems.

[0006] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows:

[0007] In a first aspect, an embodiment of the present invention provides a mode selection method for a D2D-assisted power Internet of Things, comprising the following steps:

[0008] The channel allocation label values ​​are calculated based on the D2D user pairs and cellular users in the power network and constructed as a training data set;

[0009] A graph neural network model including a graph construction module, a graph convolution module and a classification decision module is constructed. In the graph construction module, the interference relationship between the D2D user pairs and the cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the feature vectors are extracted using a graph convolution neural network to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into a classifier to calculate a channel selection variable to determine a selected channel mode.

[0010] Train the graph neural network model using the training dataset and loss function;

[0011] The trained graph neural network model is deployed to the base station to select the channel mode and send it down for communication.

[0012] Preferably, the channel allocation label values ​​are calculated based on the D2D user pairs and cellular users in the power network and constructed as a training data set, including:

[0013] Different numbers of D2D user pairs and cellular users are randomly generated, and the branch and bound method is used to calculate the channel assignment solution as the correct label value. All D2D user pairs and cellular users and their corresponding channel assignment labels are constructed as a training dataset.

[0014] Preferably, in the graph construction module, the interference relationship between the D2D user pairs in the power network and the cellular users and their channels is modeled as a graph structure, including:

[0015] Using the node modeling method, the D2D user pairs and cellular users in the power network are modeled as nodes of the graph, and the interference relationship between user channels is modeled as the edge of the graph. The final constructed graph structure is fully connected, that is, there is interference between each node. The node characteristics are the transmission power and channel gain of the D2D user pairs and cellular users, and the edge characteristics are the channel state of the interference link and the interference size.

[0016] Preferably, the node features and edge features in the graph structure are encoded into feature vectors, including:

[0017] The structure2vec graph embedding algorithm is used to calculate the feature vector of each node based on the adjacency matrix, node features, and edge features of the graph structure. The specific nonlinear iterative expression is:

[0018]

[0019] in, represents the feature vector of node v after the tth iteration, initialized to zero vector, Γ(·) represents nonlinear operation, X v represents the characteristics of node v, represents the hidden state of its input edge feature after the last iteration, represents the hidden state of the node's predecessor neighbor after the last iteration, u represents the u-th node in the set of all nodes N(v) in the graph structure, e represents the incoming edge feature, {D2D, CU} represents the node set of D2D user pairs and cellular users, and the structure2vec algorithm finally obtains the node feature vector h after iteration v .

[0020] Preferably, the step of extracting features from feature vectors using a graph convolutional neural network in a graph convolution module to obtain node hidden states including D2D user pair nodes and cellular user nodes includes:

[0021] Let the matrix H = [h v ] v∈{D2D,CU} Represents the feature matrix of all nodes after feature encoding, where h v represents the feature vector of node v, {D2D, CU} represents the node set of D2D user pairs and cellular users;

[0022] The graph convolutional neural network uses the features of aggregated neighbor nodes to update the representation of the node. The formula is expressed as:

[0023]

[0024] in, represents an adjacency matrix with self-loops, express The degree matrix, H (l) represents the node feature matrix of the lth layer of the graph convolutional neural network, H (0) Represents the input feature matrix, that is, the result of the feature encoding in the previous step, W (l) represents the weight matrix of the lth layer of the graph convolutional neural network and is continuously updated during the training process. σ represents the nonlinear operation. As the training progresses, the graph convolutional neural network will gradually learn the node hidden states h' including D2D user pairs and cellular user nodes in a semi-supervised learning manner. v .

[0025] Preferably, the step of inputting the node hidden state into a classifier to calculate a channel selection variable in the classification decision module to determine the selected channel mode comprises:

[0026] A deep neural network is used to construct an N-tuple classifier to calculate the channel selection probability. N is the total number of uplink orthogonal channels available in the cellular network. The formula is as follows:

[0027]

[0028] Among them, p(v) represents the probability that node v selects channel n, h' v represents the hidden state of the node, represents the weight vector of channel n, represents the weight vector of any channel n' including channel n, and the superscript T represents transposition;

[0029] After obtaining the probability of node v selecting all channels, the channel corresponding to the maximum probability is selected as the channel selection result of node v. The channel selection results of all nodes are combined to finally construct the channel selection variable a of the D2D user pair. k,n and the channel selection variable a of the cellular user m,n , through a k,n and a m,n It reflects the channel mode selection results of each D2D user pair and cellular user.

[0030] Preferably, the loss function adopts the cross entropy loss function The formula is:

[0031]

[0032] Where p(v) represents the probability prediction value of node v selecting channel n, q(v) represents the probability label value of node v selecting channel n, {D2D, CU} represents the node set of D2D user pairs and cellular users, and N represents the total number of uplink orthogonal channels available in the cellular network.

[0033] Preferably, the step of deploying the trained graph neural network model to a base station for channel mode selection and sending it down for communication includes:

[0034] Terminal information collection: The base station sends pilot signals to each terminal to collect channel status information of all D2D user pairs and cellular users;

[0035] Base station processing: The base station deploys a trained graph neural network model, combines the channel status and interference information of D2D user pairs and cellular users, and optimizes the mode selection of D2D user pairs and cellular users. The mode selection is dedicated mode or multiplexing mode.

[0036] Decision-making: The base station sends the optimized communication mode to each terminal;

[0037] Performing communication: Each D2D user pair and the cellular user perform communication according to the assigned mode.

[0038] In a second aspect, to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a mode selection device for D2D-assisted power Internet of Things, which is implemented by the above-mentioned mode selection method for D2D-assisted power Internet of Things, including: a data set construction module, a model construction module, a model training module and a model deployment module;

[0039] The data set construction module is used to calculate the channel allocation label value based on the D2D user pairs and cellular users in the power network to construct a training data set;

[0040] The model construction module is used to construct a graph neural network model including a graph construction module, a graph convolution module and a classification decision module. In the graph construction module, the interference relationship between the D2D user pairs and the cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the feature vectors are extracted using a graph convolution neural network to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into a classifier to calculate a channel selection variable to determine a selected channel mode.

[0041] The model training module is used to train the graph neural network model using a training data set and a loss function;

[0042] The model deployment module is used to deploy the trained graph neural network model to the base station to select the channel mode and send it down to execute communication.

[0043] In a third aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides an electronic device, comprising a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-mentioned mode selection method for D2D-assisted power Internet of Things when executing the computer program.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) Improved spectrum utilization: Through mode selection, the multiplexing mode can be selected in the appropriate scenario, which can significantly improve the spectrum utilization of the system and thus improve the overall throughput of the communication network.

[0046] (2) Reduce system interference: The D2D communication mode is optimized by using a graph neural network algorithm, which can effectively reduce the interference between D2D communication and cellular users and ensure the quality of service (QoS) of the communication system.

[0047] (3) Reduced computational complexity: Compared with traditional optimization methods (such as implicit enumeration and Hungarian algorithm), graph neural network models can perform pattern selection with lower complexity, which is especially suitable for scenarios with large network scale.

[0048] (4) Improved system throughput: By combining the optimization of dedicated mode and multiplexing mode, the overall communication rate of the system can be significantly improved, enabling the system to have better throughput performance in complex communication environments.

[0049] (5) The effect is more significant when the system scale is expanded: Due to the particularity of graph neural network training, when the system scale increases, it only corresponds to an increase in the number of graph neural network nodes. It can draw on the parameters of small-scale training to avoid an increase in complexity and reduce training costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 It is a flow chart of a mode selection method for D2D-assisted power Internet of Things provided by an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of a communication system deployment environment provided by an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of the structure of a communication system provided by an embodiment of the present invention;

[0054] Figure 4 It is a schematic diagram of the graph neural network model architecture provided by an embodiment of the present invention;

[0055] Figure 5 It is a structural schematic diagram of a mode selection device for D2D-assisted power Internet of Things provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0057] The inventive concept of the present invention is: in response to the problem of high computational complexity of mode selection of D2D communication in the prior art, an embodiment of the present invention provides a mode selection method and device for a D2D-assisted power Internet of Things, by constructing the D2D user pairs and cellular users in the communication system and the interference relationship between them as a graph structure, and optimizing the mode selection of D2D communication and maximizing the system throughput based on the graph structure using a graph neural network model, thereby further reducing the computational complexity and improving the computational speed, thereby greatly improving the overall performance of the communication system.

[0058] The communication system of the power Internet of Things involved in the present invention is as follows Figure 2 As shown in the figure, there are two types of users in the system: cellular users and D2D user pairs. D2D user pairs (D2D-T represents the transmitter and D2D-R represents the receiver) can communicate directly with each other, and there are two modes to choose from: dedicated mode and reuse mode. In dedicated mode, D2D user pairs use channel resources exclusively and do not reuse with cellular users. Therefore, the D2D communication link has no interference from cellular users and has high communication quality. However, the spectrum utilization rate is low. In reuse mode, the D2D link can reuse the uplink channel resources of the cellular network, that is, the link of one of the cellular users. At this time, the D2D link will share the channel with the cellular user, causing two-way interference. The interference is divided into two parts: the interference of D2D users on the uplink of cellular users and the interference of the uplink of cellular users on the D2D link. The size of this interference is related to the parameters such as the respective channel power gain and transmission power. Although it has a certain negative impact on the communication quality, it can improve the utilization efficiency of the spectrum. The overall goal of the system is to maximize the throughput of the overall system by selecting the appropriate D2D mode (dedicated or reused) under the premise of ensuring the communication quality of users.

[0059] Figure 1 FIG. 1 is a flow chart of a mode selection method for a D2D-assisted power Internet of Things provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a mode selection method for D2D-assisted power Internet of Things, including the following steps:

[0060] S1, calculate the channel allocation label value based on the D2D user pairs and cellular users in the power network and construct it as a training data set.

[0061] Constructing the channel optimization problem: Assume that the system has K D2D user pairs, defined as: Define the number of uplink orthogonal channels available in the cellular network as N, and the number of cellular users as M. Without loss of generality, assume that M≤N≤M+K, that is, some D2D links must choose the multiplexing mode to complete data transmission, and all cellular users can choose different orthogonal channels for transmission without mutual interference. Define indicator variable am,n represents the channel selection of cellular users, a k,n Indicates the channel selection of the D2D user pair. Both of them have values ​​of {0,1}. When the value is 1, it means that the corresponding orthogonal channel is selected. Each cellular user or D2D user pair can only select one channel for transmission.

[0062] For the mth cellular user, its uplink transmission rate is R m It is expressed as:

[0063]

[0064] Among them, p m represents the transmission power of cellular user m, h m,B represents the channel power gain between cellular user m and the base station, N0 represents the white noise power, m′ represents any cellular user except cellular user m, and p k represents the transmission power of the D2D user pair, h k,B represents the channel power gain of D2D for the base station, and B represents the bandwidth of the subchannel.

[0065] For the kth D2D user pair, its transmission rate R k It is expressed as:

[0066]

[0067] Among them, p k represents the transmission power of the kth D2D user pair, h k represents the channel power gain of the kth D2D user pair, N0 represents the white noise power, k′ represents any D2D user pair except D2D user pair k, and p m and h m,k When multiplexing channels, the corresponding transmission power and channel power gain of cellular user m, h k′,k represents the channel gain between two D2D user pairs, and B represents the bandwidth of the subchannel.

[0068] The goal of the system is to maximize the overall communication rate by adjusting the channel selection of D2D user pairs and cellular users, which can be expressed as the following channel optimization problem:

[0069]

[0070] Among them, R k represents the transmission rate of the kth D2D user pair, R m represents the uplink transmission rate of the mth cellular user, a k,n and a m,nThey represent the channel selection variable of the kth D2D user pair and the channel selection variable of the mth cellular user respectively, and n represents the index of the channel.

[0071] The constraints of the channel optimization problem are:

[0072]

[0073] SINR k ,SINR m ≥r th

[0074] Among them, SINR k and SINR m denote the signal-to-interference-noise ratio of the kth D2D user pair and the signal-to-interference-noise ratio of the mth cellular user, r th Represents a given threshold. The first two constraints indicate that each cellular user or D2D user pair can only select one channel for transmission, and the third constraint indicates that the signal-to-interference-to-noise ratio of D2D users and cellular users cannot be lower than a given value, in order to ensure the quality of service (QoS) of the communication system. In terms of optimization methods, since this optimization problem is a 0-1 integer optimization problem, the traditional optimization algorithm can be used to solve it using branch and bound method, cutting plane method, dynamic programming method, Lagrangian relaxation method and various heuristic algorithms.

[0075] In order to train the graph neural network model, it is necessary to construct a data set for the channel allocation problem in the case of coexistence of D2D user pairs and cellular users. Considering efficiency and computational complexity, the present invention adopts the branch and bound method to solve the problem. The idea of ​​using the branch and bound method to solve this problem is to allocate the channel association variable a k,n and a m,n Branches are made, such as the first D2D user reuses the channel of the first cellular user, and the second D2D user reuses the channel of the second cellular user. For each branch, check whether it meets the constraints. If it meets the constraints, the upper bound of the system's communication rate is calculated assuming that the remaining allocation is ideal. If it is lower than the optimal solution, it is pruned, otherwise the optimal solution is updated. Then backtracking is performed until all branches are traversed to obtain the optimal branch (optimal channel allocation) and the maximum rate of the system.

[0076] The advantage of the branch and bound method is that it does not require offline calculations in the early stage and can directly obtain the optimal channel allocation solution based on the model. However, the optimization problem needs to be solved once for each channel allocation, which increases the complexity of the algorithm. When the scale of the communication system increases, the complexity of the algorithm increases exponentially, so a less complex algorithm is needed.

[0077] Different numbers of D2D user pairs and cellular users are randomly generated, the channel allocation solution is calculated as the correct label value, the required training dataset is constructed, and then the base station is trained according to the graph neural network model framework in an offline state.

[0078] S2, construct a graph neural network model including a graph construction module, a graph convolution module and a classification decision module. In the graph construction module, the interference relationship between D2D user pairs and cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the graph convolution neural network is used to extract features from the feature vectors to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into the classifier to calculate the channel selection variable to determine the selected channel mode.

[0079] The application of Graph Neural Network (GNN) to communication systems has been a hot research direction in academia in recent years, providing a new solution for optimizing mode selection in large-scale communication systems. The GNN model can quickly and efficiently select modes and allocate resources by capturing the complex relationships between nodes (users and base stations) in the communication network, especially when dealing with multi-user interference and channel conditions. Compared with traditional optimization algorithms, the GNN model has the advantages of low operating complexity and fast calculation speed when facing large-scale networks, which can greatly improve the overall performance of the system. Due to the natural fit between the topological properties of the communication system and the data structure of the graph, previous studies have achieved results that are superior to traditional algorithms in terms of link planning, privacy protection, and resource allocation. The present invention will optimize the mode selection of D2D communication based on the graph neural network to achieve performance that is superior to traditional algorithms.

[0080] The graph neural network model architecture and processing process constructed by the embodiment of the present invention are as follows Figure 4 As shown, it mainly includes the following three modules.

[0081] S2.1, Graph building blocks.

[0082] The method based on graph neural network first needs to topology the communication system into a graph structure. There are two main graph modeling methods for communication systems: node modeling method and link modeling method. The node modeling method models all base stations and users of the system as communication nodes, models the user's transmission power and channel selection as node features, models the existing communication links as edges, and models the corresponding edge features as channel gain and interference, etc. This modeling method is very intuitive, that is, the topology of the communication system is directly presented in the form of a graph network; the link modeling method models the communication links and interference links of the system as nodes and edges, respectively. The graph modeled in this way is a directed graph, and the direction of the edge represents different interference links. This modeling method performs very well in a pure D2D communication system. Since D2D links and cellular users coexist in the power system involved in the present invention, the present invention selects an improved node modeling method: D2D user pairs and cellular users are modeled as nodes of the graph, and the entire graph is fully connected, that is, there are interference links between each node, which is consistent with the modeling method of rate calculation. In the graph structure modeled in the embodiment of the present invention, the node characteristics are the transmission power and channel gain of the D2D user pair or the cellular user, and the edge characteristics are the channel state of the interference link and the interference size. Figure 3 The graph structure diagram is shown, where the location of the base station is only used as a reference for the relative position of each node when modeling. In the subsequent feature extraction process, the relative position information will be extracted into the embedded features of the edge as the gain of each communication link and the attenuation value of each interference link. Cellular node x CU The node characteristics of D2D node x include its gain to the base station communication link and the transmission power allocated by the base station; D2D The node characteristics include the transmission power and channel gain of the D2D link.

[0083] The first step in training a graph neural network is feature encoding, which encodes the features of nodes and edges into feature vectors through an encoding algorithm for further calculations. In terms of the selection of the encoding architecture, the present invention adopts the structure2vec graph embedding algorithm, which calculates the p-dimensional feature vector of each node based on the adjacency matrix of the graph structure (the adjacency matrix refers to the matrix composed of the connection relationship between nodes), node features and edge features. The specific nonlinear iterative expression is:

[0084]

[0085] Wherein, the subscript v represents any node, which can be a D2D user pair node or a cellular user node; represents the feature vector of node v after the tth iteration, which is initialized to a zero vector; Γ(·) represents a nonlinear operation, and a rectified linear unit (ReLU) is selected in the embodiment; X vIndicates the characteristics of node v. If it is a D2D user to the node, it indicates the D2D input power and channel gain. If it is a cellular user, it indicates the base station power allocation and the channel gain of the cellular user. represents the hidden state of its input edge feature after the last iteration, represents the hidden state of the node's predecessor neighbor after the last iteration, u∈N(v) represents the uth node in the set of all nodes N(v) in the graph structure, e represents the input edge feature, {D2D,CU} represents the node set of D2D user pairs and cellular users, and the structure2vec algorithm finally obtains the node feature vector h after iteration v .

[0086] S2.2, graph convolution module.

[0087] After the feature encoding is completed, the node features need to be input into a specific GNN architecture for feature extraction. In the selection of the GNN architecture, the embodiment of the present invention adopts the classic and simple graph convolutional neural network (GCN) architecture, which can efficiently solve the channel allocation problem and has low computational complexity. GCN takes the input features of the cellular user node and the D2D user to the node and the input features of the edge as input, and through the propagation mechanism of the graph neural network, uses the node's own features and the interference information of the neighboring nodes to extract the high-dimensional feature representation of the node layer by layer, and finally outputs the embedded features of each node, also called the hidden state. The hidden state is an aggregation of node features, edge features, and graph features. The purpose of extracting the hidden state is to ensure that the features of each node not only contain its own transmission power and channel gain information, but also can obtain the interference link information of its neighboring nodes through the message passing mechanism to enrich its context information. The node hidden state output by the feature extraction will be used for further channel selection.

[0088] Specifically, let the matrix H = [h v ] v∈{D2D,CU} , represents the feature matrix of all nodes after feature encoding, where h v represents the feature vector of node v, and {D2D, CU} represents the node set of D2D user pairs and cellular users. The graph convolutional neural network uses the features of aggregated neighbor nodes to update the representation of the node, and the formula is expressed as:

[0089]

[0090] in, represents an adjacency matrix with self-loops, express The degree matrix, H (l) represents the node feature matrix of the lth layer of the graph convolutional neural network, H(0) Represents the input feature matrix, that is, the result of the feature encoding in the previous step, W (l) represents the weight matrix of the lth layer of the graph convolutional neural network and is continuously updated during the training process. σ represents the nonlinear operation. As the training progresses, the graph convolutional neural network will gradually learn the node hidden states h' including D2D user pairs and cellular user nodes in a semi-supervised learning manner. v .

[0091] S2.3, classification decision module.

[0092] In addition to solving the traditional optimization algorithm, the present invention proposes a solution method based on a graph neural network model. The hidden state of the output of the GCN architecture needs to be semantically transformed to obtain the respective channel allocation results. In the present invention, a classic deep neural network (DNN) is used to construct an N-ary classifier to output the result, that is, the output channel selection variable a k,n and a m,n The value of .

[0093] In the embodiment, based on the node hidden state extracted by the graph neural network model, a deep neural network is used to construct an N-tuple classifier to calculate the channel selection probability. Specifically, a SOFTMAX classifier is used, and the formula is as follows:

[0094]

[0095] Among them, p(v) represents the probability that node v selects channel n, h' v represents the hidden state of the node, represents the weight vector of channel n, represents the weight vector of any channel n' including channel n, and the superscript T represents transpose.

[0096] After obtaining the probability of node v selecting all channels, the channel corresponding to the maximum probability is selected as the channel selection result of node v. The channel selection results of all nodes are combined to finally construct the channel selection variable a of the D2D user pair. k,n and the channel selection variable a of the cellular user m,n , through a k,n and a m,n It reflects the channel mode selection results of each D2D user pair and cellular user.

[0097] S3, train the graph neural network model using the training dataset and loss function.

[0098] In the embodiment, the graph neural network model is trained using a training data set and a loss function, and the loss function adopts a cross entropy loss function. The formula is:

[0099]

[0100] Where p(v) represents the probability prediction value of node v selecting channel n, q(v) represents the probability label value of node v selecting channel n, {D2D, CU} represents the node set of D2D user pairs and cellular users, and N represents the total number of uplink orthogonal channels available in the cellular network.

[0101] S4, deploy the trained graph neural network model to the base station to select the channel mode and send it down to execute communication.

[0102] During the execution of the system, the D2D terminal and the base station need to exchange information. Specifically, the terminal needs to send its communication mode selection requirements and channel condition information, and the base station will make a global resource allocation decision based on the status feedback of all users. The specific process is as follows.

[0103] S4.1, terminal information collection: the base station sends a pilot signal to each terminal to collect the channel state information (CSI) of all users (including cellular users and D2D user pairs).

[0104] S4.2, base station processing: The base station deploys a trained graph neural network model, combines the channel status and interference information of the D2D user pairs and cellular users, and optimizes the mode selection of the D2D user pairs and cellular users. The mode selection is dedicated mode or multiplexing mode.

[0105] S4.3, decision issuance: the base station sends the optimized communication mode to each terminal.

[0106] S4.4, perform communication: each D2D user pair and the cellular user perform communication according to the assigned mode.

[0107] When the framework is running, the base station first trains the graph neural network model offline, sends a pilot signal online to obtain information about the number of users in the current cell, then loads the trained graph neural network model for fine-tuning, and sends channel selection results to each user in the cell based on the model output. The advantage of the graph neural network model is that the model operation complexity is low, and because the graph neural network model is transferable, when the network scale is relatively large, the parameters of the small-scale network can be partially migrated to the large-scale network, which greatly reduces the time required for training. When the network scale is relatively large, the advantage over the traditional optimization model is very significant.

[0108] In summary, the embodiment of the present invention provides a mode selection method for D2D-assisted power Internet of Things, which uses graph neural networks to optimize the mode selection of D2D communication and maximize the system throughput, which can improve spectrum utilization, reduce system interference, reduce computational complexity, and improve system throughput. The effect is more significant when the system scale is expanded.

[0109] Based on the same inventive concept, Figure 5 As shown, an embodiment of the present invention further provides a mode selection device 500 for D2D-assisted power Internet of Things, including: a data set construction module 510, a model construction module 520, a model training module 530 and a model deployment module 540.

[0110] The data set construction module 510 is used to calculate the channel allocation label value based on the D2D user pairs and cellular users in the power network, and construct it as a training data set.

[0111] The model construction module 520 is used to construct a graph neural network model including a graph construction module, a graph convolution module and a classification decision module. In the graph construction module, the interference relationship between D2D user pairs and cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the feature vectors are feature extracted using a graph convolution neural network to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into the classifier to calculate the channel selection variable to determine the selected channel mode.

[0112] The model training module 530 is used to train the graph neural network model using a training data set and a loss function.

[0113] The model deployment module 540 deploys the trained graph neural network model to the base station for channel mode selection and sends it down for communication execution.

[0114] Based on the same inventive concept, an embodiment of the present invention further provides an electronic device, including a memory and one or more processors, the memory is used to store a computer program, and the processor is used to implement the above-mentioned mode selection method for D2D-assisted power Internet of Things when executing the computer program.

[0115] It should be noted that the mode selection device and electronic device for D2D-assisted power Internet of Things provided in the above-mentioned embodiments belong to the same inventive concept as a mode selection method for D2D-assisted power Internet of Things. The specific implementation process is detailed in an embodiment of a mode selection method for D2D-assisted power Internet of Things, which will not be repeated here.

[0116] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A mode selection method for D2D-assisted power Internet of Things, characterized in that: The following steps are involved: The channel allocation label values ​​are calculated based on the D2D user pairs and cellular users in the power network and constructed as a training data set; A graph neural network model including a graph construction module, a graph convolution module and a classification decision module is constructed. In the graph construction module, the interference relationship between the D2D user pairs and the cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the feature vectors are extracted using a graph convolution neural network to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into a classifier to calculate a channel selection variable to determine a selected channel mode. Train the graph neural network model using the training dataset and loss function; The trained graph neural network model is deployed to the base station to select the channel mode and send it down for communication.

2. The mode selection method for D2D-assisted power Internet of Things according to claim 1 is characterized in that: The channel allocation label values ​​are calculated based on the D2D user pairs and cellular users in the power network and constructed as a training data set, including: Different numbers of D2D user pairs and cellular users are randomly generated, and the branch and bound method is used to calculate the channel assignment solution as the correct label value. All D2D user pairs and cellular users and their corresponding channel assignment labels are constructed as a training dataset.

3. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: In the graph construction module, the interference relationship between D2D user pairs and cellular users and their channels in the power network is modeled as a graph structure, including: Using the node modeling method, the D2D user pairs and cellular users in the power network are modeled as nodes of the graph, and the interference relationship between user channels is modeled as the edge of the graph. The final constructed graph structure is fully connected, that is, there is interference between each node. The node characteristics are the transmission power and channel gain of the D2D user pairs or cellular users, and the edge characteristics are the channel state of the interference link and the interference size.

4. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: Encode the node features and edge features in the graph structure into feature vectors, including: The structure2vec graph embedding algorithm is used to calculate the feature vector of each node based on the adjacency matrix, node features, and edge features of the graph structure. The specific nonlinear iterative expression is: in, represents the feature vector of node v after the tth iteration, initialized to zero vector, Γ(·) represents nonlinear operation, X v represents the characteristics of node v, represents the hidden state of its input edge feature after the last iteration, represents the hidden state of the node's predecessor neighbor after the last iteration, u represents the u-th node in the set of all nodes N(v) in the graph structure, e represents the incoming edge feature, {D2D, CU} represents the node set of D2D user pairs and cellular users, and the structure2vec algorithm finally obtains the node feature vector h after iteration v .

5. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: The feature vector is extracted by using a graph convolution neural network in the graph convolution module to obtain node hidden states including D2D user pair nodes and cellular user nodes, including: Let the matrix H = [h v ] v∈{D2D,CU} Represents the feature matrix of all nodes after feature encoding, where h v represents the feature vector of node v, {D2D, CU} represents the node set of D2D user pairs and cellular users; The graph convolutional neural network uses the features of aggregated neighbor nodes to update the representation of the node. The formula is expressed as: in, represents an adjacency matrix with self-loops, express The degree matrix, H (l) represents the node feature matrix of the lth layer of the graph convolutional neural network, H (0) Represents the input feature matrix, which is the result of the feature encoding in the previous step, W (l) represents the weight matrix of the lth layer of the graph convolutional neural network and is continuously updated during the training process. σ represents the nonlinear operation. As the training progresses, the graph convolutional neural network will gradually learn the node hidden states h including the D2D user pair nodes and the cellular user nodes. ’ v .

6. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: The step of inputting the node hidden state into the classifier in the classification decision module to calculate the channel selection variable to determine the selected channel mode comprises: A deep neural network is used to construct an N-tuple classifier to calculate the channel selection probability. N is the total number of uplink orthogonal channels available in the cellular network. The formula is as follows: Among them, p(v) represents the probability that node v selects channel n, h ’ v represents the hidden state of the node, represents the weight vector of channel n, represents any channel n including channel n ’ The weight vector of , the superscript T represents transpose; After obtaining the probability of node v selecting all channels, the channel corresponding to the maximum probability is selected as the channel selection result of node v. The channel selection results of all nodes are combined to finally construct the channel selection variable a of the D2D user pair. k,n and the channel selection variable a of the cellular user m,n , through a k,n and a m,n It reflects the channel mode selection results of each D2D user pair and cellular user.

7. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: The loss function uses the cross entropy loss function The formula is: Where p(v) represents the probability prediction value of node v selecting channel n, q(v) represents the probability label value of node v selecting channel n, {D2D, CU} represents the node set of D2D user pairs and cellular users, and N represents the total number of uplink orthogonal channels available in the cellular network.

8. The mode selection method for D2D-assisted power Internet of Things according to claim 1, characterized in that: The trained graph neural network model is deployed to the base station to select the channel mode and send it to perform communication, including: Terminal information collection: The base station sends pilot signals to each terminal to collect channel status information of all D2D user pairs and cellular users; Base station processing: The base station deploys a trained graph neural network model, combines the channel status and interference information of D2D user pairs and cellular users, and optimizes the mode selection of D2D user pairs and cellular users. The mode selection is dedicated mode or multiplexing mode. Decision-making: The base station sends the optimized communication mode to each terminal; Performing communication: Each D2D user pair and the cellular user perform communication according to the assigned mode.

9. A mode selection device for D2D-assisted power Internet of Things, implemented by using the mode selection method for D2D-assisted power Internet of Things according to any one of claims 1 to 8, characterized in that: include: Dataset building module, model building module, model training module and model deployment module; The data set construction module is used to calculate the channel allocation label value based on the D2D user pairs and cellular users in the power network to construct a training data set; The model construction module is used to construct a graph neural network model including a graph construction module, a graph convolution module and a classification decision module. In the graph construction module, the interference relationship between the D2D user pairs and the cellular users and their channels in the power network is modeled as a graph structure and the node features and edge features in the graph structure are encoded as feature vectors. In the graph convolution module, the feature vectors are extracted using a graph convolution neural network to obtain node hidden states including D2D user pair nodes and cellular user nodes. In the classification decision module, the node hidden states are input into a classifier to calculate a channel selection variable to determine a selected channel mode. The model training module is used to train the graph neural network model using a training data set and a loss function; The model deployment module is used to deploy the trained graph neural network model to the base station to select the channel mode and send it down to execute communication.

10. An electronic device comprising a memory and one or more processors, wherein the memory is used to store a computer program, characterized in that: The processor is used to implement the mode selection method for D2D-assisted power Internet of Things as described in any one of claims 1-8 when executing the computer program.

Images