D2D node communication power distribution method based on lightweight Graph U-Net

Abstract

The invention provides a D2D node communication power distribution method based on lightweight Graph U-Net, and the method comprises the steps: firstly obtaining a physical link diagram and link information of a D2D communication network, and constructing a weighted directed graph; inputting the GU-Net into the trained GU-Net, mapping the GU-Net into a high-dimensional feature map, and performing down-sampling through a screening mechanism based on a node interference degree threshold value to reduce the calculation cost; performing graph coding to obtain a coding aggregation feature, and performing up-sampling and graph decoding to obtain a decoding aggregation feature; and finally, reasoning a power prediction value of each node based on the high-dimensional feature map and the decoding aggregation feature. According to the GU-Net, a lightweight encoder-decoder architecture is adopted, gradient propagation can be stabilized, efficient optimization is achieved, the GU-Net has the multi-scale interference modeling capacity, and local and global features can be captured at the same time. According to the method, the efficiency is remarkably improved while the performance is kept, and the method is adaptive to different network topologies and has good universality and deployability.

Description

Technical Field

[0001] This application belongs to the field of wireless communication technology, specifically relating to a power allocation method for D2D node communication based on a lightweight Graph U-Net. Background Technology

[0002] In future wireless communication systems, especially in scenarios such as the Internet of Things (IoT), vehicle-to-everything (V2X), high-speed indoor data exchange, and dense hotspots, a large number of devices will need to communicate concurrently on limited spectrum resources, forming large-scale D2D networks with a large number of nodes, dense interference links, and rapid topological changes. In such networks, inter-link interference exhibits a highly coupled structure, and the system scheduling cycle is typically in the millisecond range, making it imperative that power allocation simultaneously meet the triple requirements of high dimensionality, high interference, and high real-time performance.

[0003] Existing model-driven methods, such as the Weighted Minimum Mean Square Error (WMMSE) algorithm and the FPLinQ (Fractional Programming-based Link Quality) method, typically rely on iterative solutions for explicit optimization. These methods possess a certain degree of convergence and theoretical foundation, but they suffer from the following shortcomings: (1) High computational overhead: Each round of communication scheduling requires multiple iterations to solve, which is difficult to meet the requirements of large-scale, real-time systems; (2) Poor application capability: It is highly dependent on the global channel state, which is difficult to maintain in actual scenarios.

[0004] To improve solution efficiency, data-driven resource allocation methods based on deep learning have emerged in recent years. Among them, Graph Neural Networks (GNNs) are widely used in interference analysis and power allocation problems because they can naturally model the topology of communication networks. However, existing GNN methods generally suffer from the following limitations: (1) Mostly shallow convolutional structures: They can only capture local neighborhood interference features and are difficult to learn global interference patterns across regions and scales. (2) Limited expressive power: As the network size increases, the complexity of the graph structure increases sharply. Traditional GNNs are prone to problems such as gradient vanishing and over-smoothing at too many layers, resulting in performance degradation. (3) High inference cost: The inference cost increases directly linearly or superlinearly with the number of nodes. When the number of nodes reaches hundreds or even thousands, the inference latency increases significantly, making it unsuitable for ultra-dense D2D scenarios. (4) Lack of structural compression mechanism: Most models perform calculations on the entire graph and cannot reduce the interference evaluation burden of redundant nodes through hierarchical processing.

[0005] Therefore, existing technologies cannot provide a power allocation method that can both preserve local neighborhood interference information and reflect cross-regional interference coupling in large-scale, high-interference, and real-time-critical D2D communication networks, while also enabling structured compression and efficient inference. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides a power allocation method for D2D node communication based on a lightweight Graph U-Net. The technical problem to be solved by this application is achieved through the following technical solution: Firstly, this application provides a power allocation method for D2D node communication based on a lightweight Graph U-Net, including: S100: Obtain the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; S200, the weighted directed graph is input into the trained GU-Net, so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map. The high-dimensional feature map is downsampled using a filtering mechanism based on a node interference degree threshold, and then graph encoded to obtain encoded aggregate features. The encoded aggregate features are upsampled and graph decoded to obtain decoded aggregate features. The power prediction value of each node is inferred based on the high-dimensional feature map and the decoded aggregate features.

[0007] Optionally, S100 includes: S110, Obtain the physical link diagram and link information of the D2D communication network, wherein the link information includes channel gain, path loss and interference coefficient; S120, construct a weighted graph based on the physical link graph and the link information; wherein, in the weighted graph, the transmitting link and the receiving link constitute a node, and if a link interferes with other links, there is a directed edge pointing to the interfered link; the node features include channel gain and noise power; the edge features represent the interference gain of the node.

[0008] Optionally, the high-dimensional feature map includes features of node features mapped to the high-dimensional space and features of edge features mapped to the high-dimensional space.

[0009] Optionally, the trained GU-Net includes a feature embedding module, a hierarchical graph pooling module, a hierarchical depooling module, and a power allocation inference output module connected in sequence; the hierarchical graph pooling module includes a downsampling layer and an encoder connected in sequence; the hierarchical depooling module includes an upsampling layer and a decoder connected in sequence; the power allocation inference output module includes a multilayer perceptron; the output of the feature embedding module and the output of the decoder are connected via a residual and input to the power allocation inference output module.

[0010] Optionally, the feature embedding module is used to map the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map; The downsampling layer is used to downsample the high-dimensional feature map using a filtering mechanism based on a threshold of node interference. The encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregate features; The upsampling layer is used to upsample the encoded aggregated features; The downsampling layer is used to perform graph decoding on the downsampled features to obtain decoded aggregated features; The power allocation inference output module is used to infer the power prediction value of each node based on the high-dimensional feature map and the decoded aggregated features.

[0011] Optionally, the upsampling layer restores the node features in the encoded aggregate features to the original node space according to the index position through node indexing, and fills in the nodes that were filtered out during the downsampling process by zero padding or feature copying to obtain the upsampled features.

[0012] Optionally, the S200 includes: S210, the weighted directed graph is input into the feature embedding module in the trained GU-Net, so as to use the feature embedding module to map the weighted directed graph to a high-dimensional embedding space to obtain a high-dimensional feature map; S220, the upsampling layer in the hierarchical graph pooling module is used, and the high-dimensional feature map is downsampled through a filtering mechanism based on the node interference degree threshold. Then, the encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregated features. S230, the upsampling layer in the hierarchical depooling module is used to upsample the encoded aggregated features, and then the decoder is used to perform graph decoding on the upsampled features to obtain the decoded aggregated features; S240, the power allocation inference output module infers the power prediction value of each node based on the feature embedding map and the decoded aggregated features.

[0013] Optionally, the filtering mechanism based on the node interference level threshold is to sort the nodes by the proportion of their interference gain to the total gain from largest to smallest, filter out K nodes that are higher than the threshold, and select these K nodes as the nodes that need to be downsampled.

[0014] Optionally, the training process of the trained GU-Net is carried out by setting a preset optimization objective and completing the training through a single forward propagation; the optimization objective includes maximizing system throughput, minimizing interference, and maximizing energy efficiency.

[0015] Secondly, this application provides a power allocation device for D2D node communication based on a lightweight Graph U-Net, comprising: The acquisition module is configured to acquire the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; The prediction module is configured to input the weighted directed graph into a trained GU-Net, so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map. The high-dimensional feature map is downsampled using a filtering mechanism based on a node interference level threshold, and then graph encoded to obtain encoded aggregate features. The encoded aggregate features are upsampled and graph decoded to obtain decoded aggregate features. Based on the high-dimensional feature map and the decoded aggregate features, the power prediction value of each node is inferred.

[0016] Beneficial effects: This application provides a power allocation method for D2D node communication based on a lightweight Graph U-Net. First, the physical link graph and link information of the D2D communication network are acquired to construct a weighted directed graph. This graph is then input into a trained GU-Net, mapped to a high-dimensional feature map, and downsampled using a filtering mechanism based on a node interference threshold to reduce computational cost. Next, graph encoding is performed to obtain encoded aggregated features, followed by upsampling and graph decoding to obtain decoded aggregated features. Finally, the power prediction value of each node is inferred based on the high-dimensional feature map and the decoded aggregated features. This application significantly reduces the number of active nodes during training through the node interference threshold-based filtering mechanism, thus reducing the computational cost per iteration. The pooling mechanism in the encoder and decoder reduces the number of redundant graph convolution operations. The lightweight encoder-decoder architecture in the GU-Net of this application stabilizes gradient propagation, achieves efficient optimization, and possesses multi-scale interference modeling capabilities, simultaneously capturing local and global structural features, i.e., high-dimensional feature maps and decoded aggregated features. The proposed GU-Net can significantly improve efficiency while maintaining performance, is adaptable to different network topologies, and has versatility and deployability.

[0017] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the power allocation method for D2D node communication based on lightweight Graph U-Net provided in this application. Figure 2 This is a schematic diagram of the graph modeling of the user interference channel provided in this application; Figure 3 This is the overall framework diagram of GU-Net provided in this application; Figure 4 This is a schematic diagram of graph-based sampling and upsampling provided in this application; Figure 5 This is a schematic diagram of the graph convolutional neural network provided in this application. Detailed Implementation

[0019] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0020] like Figure 1 As shown, this application provides a power allocation method for D2D node communication based on a lightweight Graph U-Net, including: S100: Obtain the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; S200, the weighted directed graph is input into a trained GU-Net (Graph U-Net) so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map. The high-dimensional feature map is downsampled using a filtering mechanism based on a node interference threshold, and then graph encoded to obtain encoded aggregated features. The encoded aggregated features are upsampled and graph decoded to obtain decoded aggregated features. Based on the high-dimensional feature map and the decoded aggregated features, the power prediction value of each node is inferred. The high-dimensional feature map includes features mapped from node features to the high-dimensional space, and features mapped from edge features to the high-dimensional space.

[0021] The filtering mechanism based on the node interference level threshold sorts the nodes by the proportion of their interference gain to the total gain from largest to smallest, filters out K nodes that are higher than the threshold, and selects these K nodes as the nodes that need to be downsampled.

[0022] The training process of the trained GU-Net is carried out through a preset optimization objective and is completed through a single forward propagation; the optimization objective includes maximizing system throughput, minimizing interference, and maximizing energy efficiency.

[0023] The interference threshold-based screening mechanism of this application compares the degree of interference (interference gain) of a node with the percentile of the total gain, and selects K nodes with interference levels higher than the threshold to construct a compressed graph, thereby reducing the calculation of invalid interference nodes.

[0024] This application significantly reduces the number of active nodes during training by employing a selection mechanism based on a node interference threshold, thereby lowering the computational cost per iteration. The pooling mechanism in the encoder and decoder reduces the number of redundant graph convolution operations. The lightweight encoder-decoder architecture in this application's GU-Net stabilizes gradient propagation, achieving efficient optimization. Furthermore, this structure possesses multi-scale interference modeling capabilities, simultaneously capturing local and global structural features, i.e., high-dimensional feature maps and decoded aggregate features. The proposed GU-Net significantly improves efficiency while maintaining performance, is adaptable to different network topologies, and exhibits versatility and deployability.

[0025] In one specific embodiment of this application, S100 includes: S110, Obtain the physical link diagram and link information of the D2D communication network, wherein the link information includes channel gain, path loss and interference coefficient; S120, construct a weighted graph based on the physical link graph and the link information; refer to Figure 2 In the weighted graph, the transmit link and the receive link constitute a node. If a link interferes with other links, there is a directed edge pointing to the interfered link. The node features include channel gain and noise power. The edge features represent the interference gain of the node.

[0026] This application constructs a weighted graph structure based on the channel gain, path loss, and interference intensity of device-to-device links. Each node corresponds to a D2D link, each edge represents the interference relationship between links, and the edge weight is used to quantify the strength of interference between links.

[0027] refer to Figure 3 The trained GU-Net includes a feature embedding module, a hierarchical graph pooling module, a hierarchical depooling module, and a power allocation inference output module connected in sequence. The hierarchical graph pooling module includes a downsampling layer and an encoder connected in sequence. The hierarchical depooling module includes an upsampling layer and a decoder connected in sequence. The power allocation inference output module includes a multilayer perceptron. The output of the feature embedding module and the output of the decoder are connected via a residual and input to the power allocation inference output module.

[0028] In one specific embodiment of this application, The feature embedding module is used to map the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map; The feature embedding module of this application maps the original node features and edge features to a unified high-dimensional space to enhance the graph representation and provide high-dimensional perturbation and node features for subsequent pooling and depooling.

[0029] The downsampling layer is used to downsample the high-dimensional feature map using a filtering mechanism based on a threshold of node interference. This application employs a filtering mechanism based on a node interference level threshold to achieve layer-by-layer downsampling of the graph structure. By calculating the influence metrics of nodes in the interference graph (such as the total amount of received interference, the weight of incoming and outgoing edges, etc.), key nodes that play a dominant role in power allocation decisions are selected, and their corresponding local structures are preserved.

[0030] This application employs a filtering mechanism based on a threshold for the degree of node interference. It performs hierarchical downsampling of the graph based on whether the interference level of a node exceeds the threshold, thereby extracting multi-scale interference patterns. The pooling process dynamically filters the links or regions with the greatest impact on overall interference, achieving structured graph compression. After each pooling operation, this application sets up a graph convolutional neural network to aggregate, transform, and enhance the interference structure of nodes and their neighborhoods, ensuring that the pooled graph still retains key interference patterns.

[0031] The encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregate features; The encoder is used for pooling, which reduces the graph size while preserving important dependencies in the interference topology, generating a multi-scale graph structure representation, and then performing message aggregation to obtain a more expressive representation of link interference relationships.

[0032] The upsampling layer is used to upsample the encoded aggregated features; The decoder is used to perform graph decoding on the downsampled features to obtain decoded aggregate features; This application utilizes a graph-based unpooling mechanism to recover the compressed graph morphological information during downsampling. Based on the node mapping relationships recorded during the sampling phase, this module gradually restores the high-level compressed features to the original graph scale, thereby achieving the fusion of global interference information and local link features to construct a complete multi-scale link feature representation.

[0033] refer to Figure 4 The upsampling layer restores the node features in the encoded aggregate features to the original node space according to the index position through node indexing, and fills in the nodes that were filtered out during the downsampling process by zero padding or feature copying to obtain the upsampled features.

[0034] The upsampling layer and decoder in this application utilize pooling indexes to depool the downsampled graph structure, achieving node-level structure recovery and maintaining node identity consistency. This is used to recover local interference structures lost during pooling, ensuring that each link obtains the corresponding power allocation value. In the depooling stage, this application also introduces a graph feature fusion unit to recover local interference structures lost during pooling.

[0035] The power allocation inference output module is used to infer the power prediction value of each node based on the high-dimensional feature map and the decoded aggregated features.

[0036] The power prediction step described in this application employs a multilayer perceptron (MLP) or graph convolutional decoding module to map the recovered node representations to corresponding transmit power values. This application utilizes multilayer nonlinear mapping to convert the recovered node representations into link power outputs, achieving efficient end-to-end power control. This technique can be optimized based on training objectives (such as maximizing the overall system rate).

[0037] This application outputs the power allocation value for each link through a trainable nonlinear mapping unit. This module can be trained according to preset optimization objectives, including but not limited to maximizing system throughput, minimizing interference, and improving energy efficiency. Since the entire inference process is based on a single forward propagation, it eliminates the need for multiple rounds of solving traditional iterative algorithms, thus enabling power allocation inference for large-scale networks to be completed in constant time, significantly improving real-time performance and scalability.

[0038] In one specific embodiment of this application, S200 includes: S210, the weighted directed graph is input into the feature embedding module in the trained GU-Net, so as to use the feature embedding module to map the weighted directed graph to a high-dimensional embedding space to obtain a high-dimensional feature map; S220, the upsampling layer in the hierarchical graph pooling module is used, and the high-dimensional feature map is downsampled through a filtering mechanism based on the node interference degree threshold. Then, the encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregated features. S230, the upsampling layer in the hierarchical depooling module is used to upsample the encoded aggregated features, and then the decoder is used to perform graph decoding on the upsampled features to obtain the decoded aggregated features; S240, the power allocation inference output module infers the power prediction value of each node based on the feature embedding map and the decoded aggregated features.

[0039] refer to Figure 5 In this application, both the encoder and decoder employ graph convolutional neural networks. Figure 5 The input consists of edge features and node features, and the output is node features.

[0040] Secondly, this application provides a power allocation device for D2D node communication based on a lightweight Graph U-Net, comprising: The acquisition module is configured to acquire the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; The prediction module is configured to input the weighted directed graph into a trained GU-Net, so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map. The high-dimensional feature map is downsampled using a filtering mechanism based on a node interference level threshold, and then graph encoded to obtain encoded aggregate features. The encoded aggregate features are upsampled and graph decoded to obtain decoded aggregate features. Based on the high-dimensional feature map and the decoded aggregate features, the power prediction value of each node is inferred.

[0041] To verify the validity of this application, we compared PCGNN (the most advanced GNN-based method) and MLP (a deep learning network-based method) under Gaussian interference channels with single antennas of different network sizes (interference gain, SNR=10dB). The results show that compared with PCGNN, the training cost is reduced by 15%-25% without loss of performance, and the inference speed is 20%-30% faster. Compared with MLP, the maximum total transmission rate performance is improved by more than 30% in large-scale networks.

[0042] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0043] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A power allocation method for D2D node communication based on a lightweight Graph U-Net, characterized in that, include: S100: Obtain the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; S200, the weighted directed graph is input into the trained GU-Net so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map. The high-dimensional feature map is downsampled using a filtering mechanism based on the node interference degree threshold, and then graph encoding is performed to obtain encoded aggregated features. The encoded aggregated features are upsampled and graph decoded to obtain decoded aggregated features; the power prediction value of each node is inferred based on the high-dimensional feature map and the decoded aggregated features.

2. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 1, characterized in that, S100 includes: S110, Obtain the physical link diagram and link information of the D2D communication network, wherein the link information includes channel gain, path loss and interference coefficient; S120, construct a weighted graph based on the physical link graph and the link information; wherein, in the weighted graph, the transmitting link and the receiving link constitute a node, and if a link interferes with other links, there is a directed edge pointing to the interfered link; the node features include channel gain and noise power; the edge features represent the interference gain of the node.

3. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 1, characterized in that, The high-dimensional feature map includes node features mapped to the high-dimensional space and edge features mapped to the high-dimensional space.

4. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 1, characterized in that, The trained GU-Net includes a feature embedding module, a hierarchical graph pooling module, a hierarchical unpooling module, and a power allocation inference output module connected in sequence. The hierarchical graph pooling module includes a downsampling layer and an encoder connected in sequence; the hierarchical depooling module includes an upsampling layer and a decoder connected in sequence; the power allocation inference output module includes a multilayer perceptron; the output of the feature embedding module and the output of the decoder are connected via residuals and input to the power allocation inference output module.

5. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 4, characterized in that, The feature embedding module is used to map the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map; The downsampling layer is used to downsample the high-dimensional feature map using a filtering mechanism based on a threshold of node interference. The encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregate features; The upsampling layer is used to upsample the encoded aggregated features; The downsampling layer is used to perform graph decoding on the downsampled features to obtain decoded aggregated features; The power allocation inference output module is used to infer the power prediction value of each node based on the high-dimensional feature map and the decoded aggregated features.

6. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 5, characterized in that, The upsampling layer restores the node features in the encoded aggregate features to the original node space according to the index position through node indexing, and fills in the nodes that were filtered out during the downsampling process by zero padding or feature copying to obtain the upsampled features.

7. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 6, characterized in that, S200 includes: S210, the weighted directed graph is input into the feature embedding module in the trained GU-Net, so as to use the feature embedding module to map the weighted directed graph to a high-dimensional embedding space to obtain a high-dimensional feature map; S220, the upsampling layer in the hierarchical graph pooling module is used, and the high-dimensional feature map is downsampled through a filtering mechanism based on the node interference degree threshold. Then, the encoder is used to perform graph encoding on the downsampled features to obtain encoded aggregated features. S230, the upsampling layer in the hierarchical depooling module is used to upsample the encoded aggregated features, and then the decoder is used to perform graph decoding on the upsampled features to obtain the decoded aggregated features; S240, the power allocation inference output module infers the power prediction value of each node based on the feature embedding map and the decoded aggregated features.

8. The power allocation method for D2D node communication based on lightweight Graph U-Net according to any one of claims 1-7, characterized in that, The filtering mechanism based on the node interference level threshold sorts the nodes by the proportion of their interference gain to the total gain from largest to smallest, filters out K nodes that are higher than the threshold, and selects these K nodes as the nodes that need to be downsampled.

9. The power allocation method for D2D node communication based on lightweight Graph U-Net according to claim 1, characterized in that, The training process of the trained GU-Net is carried out through a preset optimization objective and is completed through a single forward propagation; the optimization objective includes maximizing system throughput, minimizing interference, and maximizing energy efficiency.

10. A power allocation device for D2D node communication based on a lightweight Graph U-Net, characterized in that, include: The acquisition module is configured to acquire the physical link diagram and link information of the D2D communication network, and construct a weighted directed graph based on the physical link diagram and link information; The prediction module is configured to input the weighted directed graph into the trained GU-Net, so that the trained GU-Net maps the weighted directed graph to a high-dimensional space to obtain a high-dimensional feature map, and then performs downsampling on the high-dimensional feature map using a filtering mechanism based on the node interference degree threshold, and then performs graph encoding to obtain encoded aggregated features. The encoded aggregated features are upsampled and graph decoded to obtain decoded aggregated features; the power prediction value of each node is inferred based on the high-dimensional feature map and the decoded aggregated features.

Images