Link scheduling method for super-large-scale D2D communication system

Abstract

The invention relates to a link scheduling method for a super-large-scale D2D communication system. The method comprises the following steps: acquiring a current transceiver topological graph of a D2D communication system in real time; converting the transceiver topological graph into an interference graph, and processing the interference graph into a K-neighbor interference graph in a mode of reserving a preset nearest interference link of each node pair according to distance information between each node pair in the interference graph; inputting the K-neighbor interference graph into the trained link scheduling model for link scheduling prediction to obtain a link scheduling strategy; and scheduling the current transceiver communication link of the D2D communication system according to the link scheduling strategy, thereby improving the generalization of the link scheduling method in the large-scale D2D network.

Description

Technical Field

[0001] This application relates to the field of wireless communication technologies, and particularly to a link scheduling method for ultra-large-scale D2D communication systems. Background Art

[0002] With the continuous development of wireless communication technologies, the spectrum resources of communication systems are becoming increasingly scarce, and traditional cellular networks can no longer meet the growing communication demands. Device-to-device (D2D) communication enables direct connections between user terminals without passing through a base station, and has gradually become a key application in future integrated space-air-ground-sea networks, and is applicable to various scenarios such as the Internet of Things, smart homes, and vehicle-to-everything networks. In D2D communication, interference management is the key to improving system performance. By supporting multi-link spectrum sharing, the shortage of spectrum resources in wireless communication systems can be alleviated to a certain extent. As an important part of interference management, link scheduling not only helps to optimize resource allocation and avoid conflicts and waste of spectrum resources, but also can effectively reduce interference between devices and reasonably arrange transmission time slots and frequency bands to improve signal quality in large-scale and ultra-dense network areas. In addition, scheduling strategies can also be combined with energy consumption management to optimize transmission strategies, thereby extending the service life of devices and significantly improving the overall performance and stability of large-scale D2D networks.

[0003] However, the link scheduling problem in large-scale D2D networks faces many challenges. This problem belongs to a combinatorial optimization problem that is non-deterministic polynomial hard (NP-hard), and its constraints are non-convex, making it difficult for traditional optimization algorithms to directly solve. Some methods adopt a model-based research idea, by introducing auxiliary variables or transforming the problem into a fractional programming, which is equivalent to a convex optimization problem, and using iterative methods such as block coordinate descent and gradient descent to solve. Such methods are still regarded as one of the optimal strategies for solving the link scheduling problem at the present stage. However, due to the fact that in the actual communication environment, the channel changes violently, the algorithm delay requirement is high, and it is difficult to obtain accurate channel state information, the model-based methods not only have high computational complexity, but also rely strongly on channel state information, have insufficient robustness, and may fall into local optimal solutions. For example, the FPLinQ algorithm based on fractional programming is sensitive to the initial state, which affects whether the global optimal solution can be obtained finally.

[0004] To break through these technical bottlenecks and promote the development of ultra-large-scale D2D communication systems, a large number of machine learning-based methods have emerged in recent years. These methods utilize models such as deep neural networks and graph neural networks. By training on known network instances, they directly learn the distribution of link scheduling solutions and generate link scheduling schemes based on the trained models during the inference phase, thus usually achieving lower computational complexity compared to model-based solutions. Some methods can rely only on the location information of transceiver devices without precise channel state information, thereby possessing high robustness. However, these methods often require a large number of training samples and still have deficiencies in terms of physical interpretability, generalization, and scalability.

[0005] Therefore, the generalization of link scheduling methods in current large-scale D2D networks is poor. Summary of the Invention

[0006] Based on this, it is necessary to provide a link scheduling method for ultra-large-scale D2D communication systems that can improve the generalization of link scheduling methods in large-scale D2D networks to address the above technical problems.

[0007] A link scheduling method for an ultra-large-scale D2D communication system, the method includes:

[0008] Obtain the current transceiver topology map of the D2D communication system in real time;

[0009] Convert the transceiver topology map into an interference graph, and based on the distance information between node pairs in the interference graph, process the interference graph into a K-nearest neighbor interference graph by retaining the preset number of nearest interference links for each node pair;

[0010] Input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling prediction to obtain a link scheduling strategy;

[0011] Schedule the current transceiver communication links of the D2D communication system according to the link scheduling strategy;

[0012] Among them, the training method of the link scheduling model is:

[0013] Obtain transceiver topology map samples and sample labels, convert the transceiver topology map samples into interference graph samples, and based on the distance information between node pairs in the interference graph samples, process the interference graph samples into K-nearest neighbor interference graph samples by retaining the preset number of nearest interference links for each node pair. The sample labels are binary values indicating whether each communication link in the transceiver topology map sample is scheduled;

[0014] Train the link scheduling model based on graph-structured discrete diffusion using the K-nearest neighbor interference graph samples and sample labels to obtain a trained link scheduling model.

[0015] In one embodiment, the link scheduling model includes a Markov chain module and an anisotropic graph neural network module;

[0016] The step of training the link scheduling model based on graph-structured discrete diffusion using the K-nearest neighbor interference graph samples and sample labels to obtain a trained link scheduling model includes:

[0017] Input the sample label x0 and the K-nearest neighbor interference graph samples into the link scheduling model for iterative training until T iterations of training are completed to obtain a trained link scheduling model;

[0018] Among them, the process of the t-th iterative training is as follows:

[0019] Input the sample label x after adding noise at the (t - 1)-th time t-1 into the Markov chain module of the link scheduling model, and add noise to the sample label after adding noise at the (t - 1)-th time through the noise addition unit in the Markov chain module to obtain the sample label x after adding noise at the t-th time t ;

[0020] Input the sample label x after adding noise at the t-th time t , the sample label x0, and the K-nearest neighbor interference graph into the anisotropic graph neural network module of the link scheduling model for training.

[0021] In one embodiment, the noise addition unit is represented by the transition probability as:

[0022]

[0023] Among them, q(x t |x t-1 ) represents the transition probability from the sample label x t-1 to the sample label x t , Cat(x t , p) represents the categorical distribution on the sample label x t , and its probability is given by p. p represents the probability of the Bernoulli distribution. Q t represents the noise amplitude between the sample label x t-1 and the sample label x t , represents the one-hot code converted from x t-1 , that is, x t-1 ∈ {0, 1} N is converted to the one-hot code N represents the number of node pairs in the interference graph sample.

[0024] In one embodiment, the manner in which the trained link scheduling model performs link scheduling prediction is as follows:

[0025] Use the noise randomly generated by the Bernoulli distribution as the initial link scheduling prediction result y0 of the trained link scheduling model;

[0026] The initial link scheduling prediction result y0 and the K-nearest neighbor interference graph are input into the trained link scheduling model for iterative prediction until T times of iterative prediction are completed, and the link scheduling prediction result y of the T-th iterative prediction is output T ;

[0027] Among them, the process of the t-th iterative prediction is:

[0028] Input the link scheduling prediction result y of the (t - 1)-th time t-1 and the K-nearest neighbor interference graph into the anisotropic graph neural network module of the trained link scheduling model for prediction, and output the link scheduling estimation result of the t-th time

[0029] Input the link scheduling estimation result of the t-th time and the link scheduling prediction result y of the (t - 1)-th time t-1 into the posterior probability analysis unit of the Markov chain module for processing, and output the link scheduling prediction result y of the t-th time t .

[0030] In one embodiment, the expression of the posterior probability analysis unit is:

[0031]

[0032] Among them, represents the output of the posterior probability analysis unit, represents the link scheduling estimation result of the (t - 1)-th time, ⊙ represents the Hadamard product, Cat(y t ; p) represents the categorical distribution on the link scheduling prediction result Y t , Q t-1 represents the noise amplitude between the link scheduling prediction result y t-1 and the link scheduling prediction result y t , the superscript represents the conjugate transpose, represents the total noise amplitude added from the sample label x0 to the sample label x t ,

[0033] In one embodiment, the analysis of inputting the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling to obtain a link scheduling policy includes:

[0034] Copy the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs;

[0035] Input the m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ;

[0036] According to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results Y T , analyze the corresponding rate sum of each link scheduling prediction result;

[0037] Take the link scheduling prediction result with the largest rate sum as the link scheduling policy.

[0038] In one embodiment, the analysis of inputting the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling to obtain a link scheduling policy includes:

[0039] Detect whether the channel state information of each communication link and interference link in the D2D communication system has been obtained currently;

[0040] When the channel state information of each communication link and interference link in the D2D communication system has been obtained, copy the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs;

[0041] Input the m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ;

[0042] According to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results y T , analyze the corresponding rate sum of each link scheduling prediction result;

[0043] Take the link scheduling prediction result with the largest rate sum as the link scheduling policy;

[0044] When the channel state information of each communication link and interference link in the D2D communication system has not been obtained, input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling analysis, output a link scheduling prediction result, and take this link scheduling prediction result as the link scheduling policy.

[0045] Beneficial effects of the above link scheduling method for ultra-large-scale D2D communication systems:

[0046] 1. The D2D communication system is modeled using a K-nearest neighbor interference graph, effectively enhancing the network topology awareness and generalization ability while reducing the computational complexity;

[0047] 2. The link scheduling model is used to learn the distribution of high-quality solutions from the known network layout link scheduling schemes, and the solution space is efficiently explored through randomly initialized parallel sampling, thereby further enhancing the generalization ability of the algorithm both inside and outside the training distribution;

[0048] 3. Anisotropic graph neural network (AGNN) is introduced in the noise prediction stage to ensure excellent scalability in large-scale D2D networks. Brief Description of the Drawings

[0049] Figure 1 It is a schematic flowchart of the link scheduling method for an ultra-large-scale D2D communication system in an embodiment;

[0050] Figure 2 It is a schematic diagram of converting the transceiver topology graph into an interference graph in an embodiment;

[0051] Figure 3 It is a schematic diagram of the link scheduling model structure in an embodiment. Detailed Embodiments

[0052] In order to make the objectives, technical solutions, and advantages of this application clearer, the following further elaborates on this application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely used to explain this application and are not used to limit this application.

[0053] In one embodiment, as Figure 1 shown, a link scheduling method for an ultra-large-scale D2D communication system is provided, including the following steps:

[0054] Step S220, obtaining the current transceiver topology graph of the D2D communication system in real time.

[0055] Among them, for a single-associated single-antenna D2D communication system, it is assumed that there is a D2D communication network composed of multiple transceiver pairs, where each transmitter is associated with its corresponding receiver, and the channels between the transmitter and non-associated other receivers are used as interference channels.

[0056] Among them, the transceiver topology graph contains device location information.

[0057] In an example, a single-associated D2D network with N transceiver pairs. Each transmitter (denoted as Tx i) are all associated with the corresponding receivers (denoted as Rx i ) and both are equipped with single antennas. The communication link between them is denoted as D i . In a Gaussian interference channel, the signal y i received by the receiver Rx i is expressed as:

[0058]

[0059] where, is the channel coefficient between Tx i and Rx i , while represents the interference channel coefficient from Tx j to Rx i . represents the set of real numbers. z i represents the transmission signal sequence of Tx i , z j represents the transmission signal sequence of Tx j . represents additive white Gaussian noise, and σ 2 represents the variance of additive white Gaussian noise. For a time-varying channel, it is assumed that within each coherence block, the channel parameters remain unchanged.

[0060] In a weak interference channel, treating interference as noise (TIN) can simplify the design of encoding and decoding schemes. The signal-to-interference-plus-noise ratio (SINR) at the receiver Rx i is expressed for all i as:

[0061]

[0062] where, the variable x i ∈ {0, 1} indicates that the link D i is scheduled (x i = 1) or not scheduled (x i = 0), p i represents the maximum transmit power of Tx i , and p j represents the maximum transmit power of Tx j . The achievable data rate R i at the receiver Tx i (x) is determined by the Shannon capacity formula:

[0063] R i (x) = log(1 + SINR i )

[0064] where, SINR i represents the signal-to-interference-plus-noise ratio at the receiver Rx i .

[0065] Formulate the link scheduling problem as maximizing the weighted rate sum of N communication links within each coherent block. Let ω i represent the priority of link D i . The resulting optimization problem is expressed as:

[0066]

[0067] Step S240, convert the transceiver topology graph into an interference graph, and based on the distance information between each pair of nodes in the interference graph, process the interference graph into a K-nearest neighbor interference graph by retaining the preset number of the nearest interference links for each pair of nodes.

[0068] Among them, the topology structure of the D2D communication system is modeled in the form of an interference graph, so as to highlight the influence of the interference channel, facilitate processing, and improve the prediction effect.

[0069] Among them, the interference graph contains the distance information between node pairs.

[0070] Among them, the K-nearest neighbor interference graph modeling method can enhance the topology awareness and generalization ability of the neural network, and reduce the computational complexity.

[0071] In an example, as Figure 2 shown on the left, the transceiver topology graph is converted into an interference graph as Figure 2 shown on the right.

[0072] Among them, in the conversion process, in the original transceiver topology graph, each pair of transceivers and the desired communication link between them are represented as a node in the interference graph. For a D2D network with N transceiver pairs, its interference graph consists of vertices {D i | i = 1, …, N}, self-loops representing the desired communication links {(D i , D i ) | i = 1, …, N} and directed edges representing the interference channels {(D i , D j) It is composed of {∣i≠j, i, j = 1, …, N}. For each vertex, its self-loop and the K (K < N - 1) incoming edges with the shortest distances are retained, and the remaining edges are discarded. Therefore, the total number of edges in the top-K strongest channel graph is N·(K + 1). When the number of links is large, this method reduces the computational complexity by significantly reducing the number of edges. In addition, it enhances the differences between different graphs, thereby improving the generalization ability of the algorithm. These distance information are embedded as edge features of the graph neural network. For the sample labels used in the supervised learning stage, the binary values {0, 1} representing whether each communication link is scheduled are embedded into the graph neural network as the initialized node features to guide the training of the model, where the binary value 1 indicates being scheduled and the binary value 0 indicates not being scheduled.

[0073] Step S260, input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling prediction to obtain a link scheduling strategy.

[0074] Among them, for the K-nearest neighbor interference graph of a certain time slot in a certain scenario, one can input only one such K-nearest neighbor interference graph into the trained link scheduling model for link scheduling prediction, and the trained link scheduling model outputs a link scheduling prediction result, and use this link scheduling prediction result as the link scheduling strategy; one can also input multiple such K-nearest neighbor interference graphs into the trained link scheduling model for link scheduling prediction, and the trained link scheduling model outputs multiple link scheduling prediction results, and select the link scheduling prediction result with the maximum sum of rates from the multiple link scheduling prediction results as the link scheduling strategy.

[0075] Among them, the trained link scheduling model can be deployed on hardware platforms such as FPGA.

[0076] Among them, the trained link scheduling model outputs the probability that each node value is 1 (i.e., indicating link activation). It is converted into a binary representation {0, 1} through a threshold decoding method to obtain the final link scheduling prediction result.

[0077] Among them, the training method of the link scheduling model is as follows:

[0078] Obtain transceiver topology graph samples and sample labels, convert the transceiver topology graph samples into interference graph samples, and according to the distance information between node pairs in the interference graph samples, process the interference graph samples into K-nearest neighbor interference graph samples in the way of retaining the preset number of nearest interference links for each node pair, and the sample labels are the binary values indicating whether each communication link in the transceiver topology graph sample is scheduled; use the K-nearest neighbor interference graph samples and sample labels to train the link scheduling model based on graph-structured discrete diffusion to obtain the trained link scheduling model.

[0079] Among them, the link scheduling model based on graph-structured discrete diffusion can learn the distribution of high-quality solutions, calculate the posterior probability, and gradually denoise according to the posterior probability to solve the target result from the noise.

[0080] Among them, each sample label corresponds to a transceiver topology graph sample. Each sample label is the optimal link scheduling result of the transceiver topology graph sample obtained by processing the device location information in the transceiver topology graph sample corresponding to the sample label into a non-fully connected neighbor interference graph, normalizing the distance information of the non-fully connected neighbor interference graph, and using the FPLinQ algorithm.

[0081] It should be understood that the training of the link scheduling model is through a forward process, gradually adding noise to the sample labels of the training instances until it turns into pure noise after several steps, so that the link scheduling model learns the probability distribution of the link scheduling results in different network instances.

[0082] It should be understood that when the trained link scheduling model performs link scheduling prediction, it calculates the posterior probability from the noise distribution to the solution distribution through a reverse process, generates random noise, derives the probability that each communication link in the target network is scheduled from it, and decodes the obtained probability information into the link scheduling result.

[0083] Step S280, schedule the current transceiver communication links of the D2D communication system according to the link scheduling strategy.

[0084] In one embodiment, the link scheduling model includes a Markov chain module and an anisotropic graph neural network module.

[0085] Among them, the anisotropic graph neural network module is a module constructed with an anisotropic graph neural network (AGNN), and the anisotropic graph neural network can have strong scalability for large-scale D2D networks.

[0086] Training the link scheduling model based on graph-structured discrete diffusion using the transmitter topology graph sample and the sample label to obtain a trained link scheduling model, including:

[0087] Input the sample label x0 and the K-nearest neighbor interference graph sample into the link scheduling model for iterative training until T times of iterative training are completed to obtain a trained link scheduling model.

[0088] Among them, the process of the t-th iterative training is:

[0089] Input the sample label x after adding noise in the (t - 1)-th time t-1 into the Markov chain module of the link scheduling model, and add noise to the sample label after adding noise in the (t - 1)-th time through the noise addition unit in the Markov chain module to obtain the sample label x after adding noise in the t-th timet ; The sample label x after adding noise for the t-th time t , the sample label x0, and the K-nearest neighbor interference graph are input into the anisotropic graph neural network module of the link scheduling model for training.

[0090] It should be understood that at the first iteration training (t = 1), the sample label input into the Markov chain module of the link scheduling model should be the sample label x0 without adding noise.

[0091] It should be understood that the sample label x after adding noise for the T-th time T , and this distribution will approximate pure noise.

[0092] Among them, as Figure 3 shown, the training of the link scheduling model is the forward process of the link scheduling model, and the link scheduling prediction of the link scheduling model is the reverse process. From the perspective of probability distribution, the link scheduling model aims to estimate a probability distribution that is as close as possible to the true distribution of the target data, and both the forward process and the reverse process can be regarded as Markov chains.

[0093] Among them, as Figure 3 shown, the anisotropic graph neural network module constructed by the anisotropic graph neural network is used as a noise predictor. The noise predictor learns and trains in the forward process, learning the changes in the graph structure before and after adding noise at each step, so as to accurately predict the noise that needs to be removed in the subsequent reverse process. The noise predictor performs link scheduling prediction in the reverse process.

[0094] Among them, the anisotropic graph neural network (AGNN) with an edge gating mechanism. This anisotropic graph neural network normalizes the features of the nodes and edges in the K-nearest neighbor interference graph (i.e., the distance information in the K-nearest neighbor interference graph), which can enhance the discrimination ability between different K-nearest neighbor interference graphs and improve the prediction effect. Since AGNN can adaptively distinguish the importance of neighborhood information in different directions, this noise predictor shows excellent scalability when dealing with large-scale D2D networks.

[0095] In one embodiment, the noise addition unit is represented by a transition probability as follows:

[0096]

[0097] Among them, q(x t |x t-1 ) represents the transition probability from the sample label x t-1 to the sample label x t , and Cat(x t , p) represents at the sample label x tThe classification distribution on it, the probability of which is given by p, where p represents the probability of the Bernoulli distribution, and Q t represents the sample label x t-1 and the noise amplitude between the sample label x t represents the one-hot code of the conversion of x t-1 i.e., x t-1 ∈ {0, 1} N is converted to the one-hot code N represents the number of node pairs in the interference graph sample.

[0098] In one embodiment, the way for the trained link scheduling model to perform link scheduling prediction is as follows:

[0099] Use the noise randomly generated by the Bernoulli distribution as the initial link scheduling prediction result y0 of the trained link scheduling model; the initial link scheduling prediction result y0 and the K-nearest neighbor interference graph are input into the trained link scheduling model for iterative prediction until T times of iterative prediction are completed, and the link scheduling prediction result y of the T-th iterative prediction is output T ;

[0100] Among them, the process of the t-th iterative prediction is:

[0101] Input the link scheduling prediction result y t-1 of the (t - 1)-th time and the K-nearest neighbor interference graph into the anisotropic graph neural network module of the trained link scheduling model for prediction, and output the link scheduling estimation result of the t-th time Input the link scheduling estimation result of the t-th time and the link scheduling prediction result y t-1 of the (t - 1)-th time into the posterior probability analysis unit of the Markov chain module for processing, and output the link scheduling prediction result y t of the t-th time.

[0102] It should be understood that when the t = 1-th iterative prediction is performed, the link scheduling prediction result input into the anisotropic graph neural network module of the trained link scheduling model for prediction is the initial link scheduling prediction result y0.

[0103] Among them, since there is no known link scheduling scheme in the prediction process, the node features of the anisotropic graph neural network are initialized with the noise generated by the Bernoulli distribution. In addition, to achieve parallel sampling and subsequent denoising processing, for each D2D network layout, the edge features and node information are repeated m times and concatenated into a whole to generate the required noise for random initialization.

[0104] In one embodiment, the expression of the posterior probability analysis unit is:

[0105] ​

[0106] Among them, represents the output of the posterior probability analysis unit, represents the link scheduling estimation result of the (t - 1)-th time, ⊙ represents the Hadamard product, Cat(y t ; p) represents the categorical distribution on the link scheduling prediction result y t ; Q t-1 represents the link scheduling prediction result y t-1 and the link scheduling prediction result y t of the noise amplitude, the superscript represents the conjugate transpose, represents the sum of the noise amplitudes added from the sample label x0 to the sample label x t ;

[0107] Among them, the process of the t-th iteration prediction represented by the transition probability can be expressed as:

[0108]

[0109] Among them, p θ (y t |y t-1 ) represents the transition probability from the prediction result y t-1 to the prediction result y t ; represents the anisotropic graph neural network module in the process of the t-th iteration prediction, according to the output of y t-1 .

[0110] In one embodiment, input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling analysis to obtain a link scheduling strategy, including:

[0111] Copy the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs; input the m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ; According to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results y T , analyze the corresponding rate sum of each link scheduling prediction result; use the link scheduling prediction result with the largest rate sum as the link scheduling strategy.

[0112] It should be understood that when inputting m K-nearest neighbor interference graphs for parallel link scheduling prediction, the trained link scheduling model will randomly generate m noises as the initial link scheduling prediction results y0 corresponding to each K-nearest neighbor interference graph.

[0113] Among them, randomly initialized parallel sampling denoising helps to efficiently explore the solution space and enhance the algorithm's generalization ability both inside and outside the distribution.

[0114] Among them, since in parallel sampling denoising, a graph is randomly initialized m times, the result after prediction also needs to be split back into m sequences. Finally, using the known channel gain, calculate the sum of the rates of each sequence, and select the sequence with the largest sum of rates as the final result.

[0115] In one embodiment, input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling analysis to obtain a link scheduling strategy, including:

[0116] Detect whether the channel state information of each communication link and interference link in the D2D communication system has been obtained currently; in the case of obtaining the channel state information of each communication link and interference link in the D2D communication system, copy the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs; input the m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ; according to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results y T , analyze the sum of the rates corresponding to each link scheduling prediction result; use the link scheduling prediction result with the largest sum of rates as the link scheduling strategy; in the case of not obtaining the channel state information of each communication link and interference link in the D2D communication system, input the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling analysis, output the link scheduling prediction result, and use this link scheduling prediction result as the link scheduling strategy.

[0117] It should be understood that the channel state information of each communication link and interference link in the D2D communication system is the channel state information of all links (including communication links and interference links) in the D2D communication system.

[0118] Among them, the channel state information is used to calculate the sum of the rates and make comparisons. The input of AGNN itself is distance information. If CSI (Channel State Information, referred to as channel state information) cannot be obtained, then set m = 1 and directly output the inference result, that is, skip the parallel sampling process.

[0119] Among them, the device location information of the to-be-scheduled network is converted into a K-nearest neighbor interference graph, and the deployed and trained link scheduling model is used to infer the link scheduling probabilities through the reverse process diffusion. Finally, the probability information is decoded to generate multiple groups of link scheduling prediction results, and the optimal solution is selected from them. In the reverse process, a parallel sampling strategy is used to sample multiple groups of noises from the Bernoulli noise distribution, and a series of corresponding link scheduling prediction results are obtained through parallel calculation by the trained link scheduling model, so as to effectively explore the solution space and further improve the performance of link scheduling prediction.

[0120] The above link scheduling method for ultra-large-scale D2D communication systems includes two stages: model training and online inference. In the model training task, first, the device location information is converted into a non-fully connected K-nearest neighbor interference graph, and the link scheduling results generated by the FPLinQ algorithm are used as training sample labels and input into the link scheduling model; subsequently, noises are gradually added to these sample labels through the forward process of the link scheduling model, so that the link scheduling model learns the distribution of link scheduling results. In the online inference scheduling task, the device location information of the to-be-scheduled D2D communication network is also converted into a K-nearest neighbor interference graph, and the trained link scheduling model deployed on hardware platforms such as FPGA is used to deduce the scheduling probabilities of each link through the reverse process of the link scheduling model; in this reverse process, a parallel sampling strategy is adopted to collect multiple groups of noise samples from the Bernoulli noise distribution, and a series of corresponding link scheduling results are obtained through parallel calculation by the link scheduling model, and finally the probability information is decoded to generate multiple scheduling solutions, and the optimal solution is selected from them.

[0121] The above link scheduling method for ultra-large-scale D2D communication systems uses a K-nearest neighbor interference graph to model the D2D network, thereby effectively enhancing the network's topological awareness and generalization ability, while reducing the computational complexity; subsequently, the link scheduling model is used to learn the distribution of high-quality solutions from the known best FPLinQ solutions, and the solution space is efficiently explored in combination with randomly initialized parallel sampling to further improve the generalization performance of the algorithm inside and outside the training distribution; for the noise predictor in the link scheduling model, a strategy based on anisotropic graph neural network (AGNN) is used, making it show excellent scalability when dealing with large-scale D2D communication networks. The above link scheduling method for ultra-large-scale D2D communication systems shows excellent performance in network scenarios with different densities and scales, fully demonstrating its outstanding scalability and generalization ability.

[0122] Taking the link scheduling of a single - association single - antenna D2D communication system under the ITU - 1411 channel model as an example, this method is divided into two parts: training and inference. For the training of the model, a large number of known training instances are required. First, the device location information is processed into an un - fully - connected neighbor interference graph, and the link scheduling results calculated by the FPLinQ algorithm are used as the sample labels of the training instances. Subsequently, the link scheduling model is used to gradually add noise for learning. For the online inference scheduling task, the device location information is also processed into an un - fully - connected neighbor interference graph, and then the trained link scheduling model is deployed. For each D2D communication network to be scheduled, by means of parallel sampling and denoising, different Bernoulli noises are randomly generated first, and then the probabilities of each link being scheduled on the network are inferred in parallel from the generated noises. The probability information is decoded into a series of different link scheduling results, and finally the optimal result is selected as the final scheduling scheme. Thus, by taking advantage of the scalability of the anisotropic graph neural network and the efficient exploration ability of the link scheduling model in the solution space, the existing link scheduling algorithms can be broken through in terms of the total system rate, and it has a low computational complexity, showing excellent performance in different network scenarios. In addition, this method has strong generalization and scalability, enabling it to be trained on a small - scale network and used in a large - scale network system, thereby reducing the computational resource overhead.

[0123] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the indication of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order limit for the execution of these steps, and these steps can be executed in other orders. Moreover,

[0124] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0125] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A link scheduling method for a ultra-large-scale D2D communication system, characterized in that, The method includes: Obtaining the current transceiver topology graph of the D2D communication system in real time; Converting the transceiver topology graph into an interference graph, and processing the interference graph into a K-nearest neighbor interference graph in a manner of retaining a preset number of the nearest interference links between each pair of nodes according to the distance information between each pair of nodes in the interference graph; Inputting the K-nearest neighbor interference graph into the trained link scheduling model for link scheduling prediction to obtain a link scheduling strategy; Scheduling the current transceiver communication links of the D2D communication system according to the link scheduling strategy; Wherein, the training method of the link scheduling model is: Obtaining a transceiver topology graph sample and a sample label, converting the transceiver topology graph sample into an interference graph sample, and processing the interference graph sample into a K-nearest neighbor interference graph sample in a manner of retaining a preset number of the nearest interference links between each pair of nodes according to the distance information between each pair of nodes in the interference graph sample, and the sample label is the binary value indicating whether each communication link in the transceiver topology graph sample is scheduled; Training the link scheduling model based on graph-structured discrete diffusion by using the K-nearest neighbor interference graph sample and the sample label to obtain a trained link scheduling model.

2. The method according to claim 1, wherein The link scheduling model includes a Markov chain module and an anisotropic graph neural network module; The training the link scheduling model based on graph-structured discrete diffusion by using the K-nearest neighbor interference graph sample and the sample label to obtain a trained link scheduling model includes: Inputting the sample label x0 and the K-nearest neighbor interference graph sample into the link scheduling model for iterative training until T times of iterative training are completed to obtain a trained link scheduling model; Wherein, the process of the t-th iterative training is: The sample label \(x\) after adding noise for the \((t - 1)\)-th time t-1 is input into the Markov chain module of the link scheduling model, and noise is added to the sample label \(x\) after adding noise for the \((t - 1)\)-th time through the noise addition unit in the Markov chain module to obtain the sample label \(x\) after adding noise for the \(t\)-th time t ; The sample label x after adding noise for the t-th time t , the sample label x0 and the K-nearest neighbor interference graph are input into the anisotropic graph neural network module of the link scheduling model for training.

3. The method according to claim 2, wherein The added noise unit is represented by a transition probability as: Among them, q(x t |x t-1 ) represents the transition probability from sample label x t-1 to sample label x t , Cat(x t , p) represents the categorical distribution over sample label x t , whose probabilities are given by p, where p represents the probability of the Bernoulli distribution, Q t represents the noise amplitude between sample label x t-1 and sample label x t , represents the one - hot code converted from x t-1 , that is, x t-1 ∈ {0, 1} N is converted to the one - hot code N represents the number of node pairs in the interference graph samples.

4. The method according to claim 1, wherein The way for the trained link scheduling model to perform link scheduling prediction is: Using the noise randomly generated by the Bernoulli distribution as the initial link scheduling prediction result y0 of the trained link scheduling model; The initial link scheduling prediction result y0 and the K-nearest neighbor interference graph are input into the trained link scheduling model for iterative prediction until T times of iterative prediction are completed, and the link scheduling prediction result y of the T-th iterative prediction is output T ; Wherein, the process of the t-th iterative prediction is: Input the link scheduling prediction result y of the (t - 1)-th time t-1 and the K-nearest neighbor interference graph into the anisotropic graph neural network module of the trained link scheduling model for prediction, and output the link scheduling estimation result of the t-th time Input the link scheduling estimation result of the t-th time and the link scheduling prediction result y of the (t - 1)-th time t-1 into the posterior probability analysis unit of the Markov chain module for processing, and output the link scheduling prediction result y of the t-th time t .

5. The method according to claim 4, characterized in that The expression of the posterior probability analysis unit is: Among them, represents the output of the posterior probability analysis unit, represents the link scheduling estimation result of the (t - 1)-th time, ⊙ represents the Hadamard product, Cat(y t ; p) represents the categorical distribution on the link scheduling prediction result y t , Q t-1 represents the link scheduling prediction result y t-1 and the link scheduling prediction result y t of the noise amplitude, the superscript T represents the conjugate transpose, represents the sum of the noise amplitudes added from the sample label x0 to the sample label x t , 6. The method according to claim 1, wherein The analyzing the K-nearest neighbor interference graph by inputting the K-nearest neighbor interference graph into the trained link scheduling model to obtain a link scheduling strategy includes: Copying the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs; Input m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ; According to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results y T , analyze the sum of the rates corresponding to each link scheduling prediction result; Taking the link scheduling prediction result with the maximum sum of rates as the link scheduling strategy.

7. The method according to claim 1, characterized in that, The analyzing the K-nearest neighbor interference graph by inputting the K-nearest neighbor interference graph into the trained link scheduling model to obtain a link scheduling strategy includes: Detecting whether the channel state information of each communication link and interference link in the D2D communication system is obtained currently; In the case of obtaining the channel state information of each communication link and interference link in the D2D communication system, copying the K-nearest neighbor interference graph to obtain m K-nearest neighbor interference graphs; Input m K-nearest neighbor interference graphs into the trained link scheduling model for parallel link scheduling prediction, and output m link scheduling prediction results y T ; According to the channel state information of each communication link and interference link in the D2D communication system, and the m link scheduling prediction results y T , analyze the corresponding rate sum of each link scheduling prediction result; Taking the link scheduling prediction result with the maximum sum of rates as the link scheduling strategy; In the case where the channel state information of each communication link and interference link in the D2D communication system is not obtained, the K-nearest neighbor interference graph is input into the trained link scheduling model for link scheduling analysis, and a link scheduling prediction result is output, and this link scheduling prediction result is used as the link scheduling strategy.

Images