A method for constructing a physical layer converged network

Abstract

The present invention discloses a method for constructing a physical layer fusion network, comprising the following steps: collecting physical layer feature data and performing preprocessing; constructing a node feature matrix and an edge feature matrix to generate an initial graph structure; using an adjacency matrix to generate a position encoding matrix; constructing view pairs, performing self-supervised graph pre-training, and generating an initial representation of a communication node; performing a local message passing operation to obtain a local context node representation of a communication node; performing global attention modeling; calculating the connection weights between communication nodes, and updating the graph connection relationship to generate a dynamically evolving network graph structure; performing optimal path planning and communication link selection to form a network connection plan. The present invention integrates physical layer multi-mode features with a graph neural network modeling method to construct an adaptive and highly stable topological structure with the advantages of strong generalization, fast response, and accurate perception.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication intelligent optimization, and in particular to a method for constructing a physical layer fusion network. Background Art

[0002] In wireless communication systems, the construction of network topology is a key link in ensuring communication quality and system stability. Traditional network construction methods mostly rely on static policy settings at the network layer, or perform simple link judgment based on a small number of physical layer indicators. These methods are unable to accurately characterize the actual communication capabilities between nodes in complex environments. With the diversification of communication environments and the development of device heterogeneity, it is difficult to fully reflect the link quality between nodes using only a single physical layer indicator. As a result, the constructed topology structure exhibits poor robustness and adaptability when facing dynamic factors such as channel fluctuations and node mobility.

[0003] In recent years, graph neural networks have gradually been applied to the field of network structure modeling. Some studies have attempted to introduce graph structures into the modeling process of communication networks, and use the high-order relationship between node features and graph structures to achieve more effective network construction strategies. However, most existing methods focus on static graph modeling or explicit graph learning stages, and have not fully integrated the rich physical layer feature information in wireless communications. They often ignore the structural encoding of node positions and the modeling capabilities of cross-frequency domain connection states. At the same time, there is a lack of targeted self-supervised pre-training mechanisms in the graph learning process, which leads to limited expressive power of node representations and inability to effectively adapt to the dynamic changes of the actual communication environment.

[0004] Existing physical layer network modeling technologies have limitations to varying degrees in terms of multimodal feature fusion, self-supervised learning capabilities, dynamic structure adaptability, and path planning mechanisms. There is an urgent need to propose a new approach that integrates physical layer features, multi-stage graph modeling, and adaptive optimization strategies to enhance the intelligent level of network construction and communication robustness of wireless networks in dynamic and complex environments. Summary of the Invention

[0005] One purpose of the present invention is to propose a method for constructing a physical layer fusion network. The present invention integrates the multimodal characteristics of the physical layer and the graph neural network modeling method, combines the self-supervised graph pre-training mechanism, and constructs a dynamically evolving network topology structure with adaptive capabilities and physical layer perception capabilities. It can achieve high-precision connection relationship modeling and optimal path planning in complex and changeable wireless channel environments, and has the advantages of strong generalization ability, high network stability, fast network construction response speed and high link selection accuracy.

[0006] A method for constructing a physical layer converged network according to an embodiment of the present invention includes the following steps:

[0007] Collect physical layer characteristic data of each communication node in the wireless communication system and perform preprocessing;

[0008] Based on the preprocessed physical layer feature data, the node feature matrix and edge feature matrix are constructed, and the adjacency matrix is ​​generated according to the physical connection status between the communication nodes, and the initial graph structure is generated at the same time;

[0009] Generate position encoding matrix using adjacency matrix;

[0010] Based on the initial graph structure, view pairs are constructed and self-supervised graph pre-training is performed. The initial representation of the communication nodes is generated through graph comparison learning between the main encoder and the target encoder.

[0011] Perform a local message passing operation based on the initial representation of the communication node to obtain the local context node representation of the communication node;

[0012] Concatenate the local context node representation with the corresponding position encoding matrix, perform global attention modeling, and generate an updated node representation that integrates local and global information;

[0013] Based on the updated node representation, the connection weights between communication nodes are calculated, a weighted adjacency matrix is ​​constructed, and the graph connection relationship is updated according to the weighted adjacency matrix to generate a dynamically evolving network graph structure;

[0014] Based on the network graph structure, optimal path planning and communication link selection are performed to form a network connection plan at the current moment. When it is detected that the communication node has changed its position, the physical layer characteristic data has abnormal fluctuations, or the connection weight has decreased, the above steps are re-executed to achieve adaptive construction and dynamic optimization of the network graph structure.

[0015] Optionally, the physical layer characteristic data includes channel state information value, received signal strength, signal-to-noise ratio, channel phase, frequency offset and synchronization timestamp, and the preprocessing includes outlier detection and elimination, missing value filling, phase expansion and alignment, unified dimension and normalization.

[0016] Optionally, generating an initial graph structure includes:

[0017] The number of communication nodes in the wireless communication system is set to a fixed value, each communication node has pre-processed physical layer feature data, and the feature data of all communication nodes are combined in node number order to form a node feature matrix, where each row in the node feature matrix represents a complete physical layer feature entry corresponding to a communication node;

[0018] The adjacency matrix is ​​constructed based on whether there is a physical connection relationship between communication nodes. The specific rules are:

[0019] When the channel state information value between two communication nodes is greater than or equal to the preset connection threshold, it is considered that there is a communication link between the two communication nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of 1;

[0020] If the channel state information value is less than the preset connection threshold, it is considered that there is no stable communication path between the two nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of zero;

[0021] The connection threshold value is set as a fixed constant according to the system environment;

[0022] For all communication node pairs marked as connected in the adjacency matrix, edge feature data is extracted, and the edge features of all connected node pairs are sequentially combined to generate an edge feature matrix, where each row of the edge feature matrix corresponds to an established physical connection path;

[0023] The edge feature data includes three types of indicators: channel state information difference, received signal strength difference and channel phase difference, which describe the physical connection characteristics between communication nodes;

[0024] The node feature matrix, adjacency matrix and edge feature matrix are used as inputs for constructing the graph structure. The initial graph structure is defined in the form of graph structure triples, which serve as the basic structure input for graph neural network modeling and graph position encoding calculation.

[0025] Optionally, generating a position encoding matrix includes:

[0026] Constructing a degree matrix based on the adjacency matrix to represent the connection strength of each node, where each element on the main diagonal of the degree matrix is ​​equal to the number of connections of the corresponding node, and the remaining elements are zero;

[0027] The Laplace matrix of the graph is constructed based on the element-by-element difference between the adjacency matrix and the degree matrix. The eigenvalue decomposition of the Laplace matrix is ​​performed to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The first several low-frequency eigenvectors are selected as the Laplace position code vectors of the communication nodes to capture the frequency domain structure information of the nodes.

[0028] The shortest path distance between each pair of communicating nodes is further calculated to construct a path distance matrix, where each element of the path distance matrix represents the number of hops required between two nodes. The path information of each node is then reduced in dimension using a nonlinear mapping function to form a path position encoding vector.

[0029] Based on the adjacency matrix, the transition probability between communication nodes is calculated, and a transition probability matrix is ​​constructed to represent the probability of any node jumping to its adjacent node. Based on the transition probability matrix, the frequency of each node being visited is counted through multiple rounds of random walks to generate a random walk position encoding vector.

[0030] The three types of vectors, namely the Laplace position coding vector, the path position coding vector and the random walk position coding vector, are concatenated according to the column dimension to obtain a complete position coding matrix, in which each row corresponds to the structural position vector of a communication node.

[0031] Optionally, generating an initial representation of the communication node includes:

[0032] Based on the initial graph structure, using the node feature matrix and adjacency matrix of the communication node as input, a view pair is constructed. The view pair includes an original view and a perturbed view. The original view is a complete graph structure, and the perturbed view is a transformed graph structure with perturbations applied on the complete graph structure. The perturbations include masking some feature dimensions of the communication nodes, randomly discarding some edge connections in the adjacency matrix, and performing random subgraph sampling operations on the graph structure.

[0033] The original view and the perturbed view are fed into two graph encoders with the same parameter structure but different update strategies. The first encoder acts as the master encoder, whose parameters are updated by backpropagating gradients. The second encoder acts as the target encoder, whose parameters are updated by an exponential sliding average of the master encoder parameters. The two encoders generate a structure representation vector for each communication node.

[0034] Both the main encoder and the target encoder adopt graph attention network, which realizes weighted aggregation of information between communication nodes through multi-head attention mechanism;

[0035] The parameter update rule of the target encoder is:

[0036] ;

[0037] in, represents the parameters of the target encoder, Represents the parameters of the main encoder, represents the momentum coefficient, Represents an assignment operation, represents the momentum compensation factor;

[0038] The cosine similarity between the representation vectors of the same communication node in two views is used as the positive sample pair, and the remaining representation vector pairs between different nodes are used as negative sample pairs. A temperature coefficient is introduced to adjust the gradient amplitude, and a self-supervised contrast loss function is constructed to minimize the distance between positive samples and maximize the representation discrimination between negative samples:

[0039] ;

[0040] in, represents the self-supervised contrastive loss function, represents the natural exponential function, represents the cosine similarity between vectors, Indicates the master encoder The representation vector of each node, Indicates the target encoder’s The representation vector of each node, represents the temperature coefficient, Indicates the total number of communication nodes, Indicates the target encoder’s The representation vector of each node;

[0041] After the training is completed, the output of the main encoder is used as the structural representation initialization state of all communication nodes to generate the initial representation of the communication nodes.

[0042] Optionally, the process of obtaining the local context node representation of the communication node is:

[0043] ;

[0044] in, Indicates that the communication node is The local context node representation of the layer, represents the degree matrix, represents the adjacency matrix, Indicates the The weight matrix of the layer, represents a nonlinear activation function, Indicates that the communication node is The local context node representation of the layer.

[0045] Optionally, generating an updated node representation that fuses local and global information includes:

[0046] For each communication node, obtain the local context node representation and extract the position encoding vector of the corresponding position from the position encoding matrix, which respectively represent the structural semantics and position information of the communication node;

[0047] Concatenate the local context node representation and the corresponding position encoding vector according to the feature dimension to generate a fused input vector;

[0048] The fusion input vectors of all communication nodes are constructed into a fusion input matrix in the order of node index;

[0049] Perform global modeling, and use the multi-head attention mechanism to interact with and adjust the weights of the fusion input matrices of each communication node to generate updated node representations that fuse local and global information.

[0050] Optionally, generating a dynamically evolving network graph structure includes:

[0051] Extract the embedding vector of each communication node from the updated node representation and set it as the updated node representation of the corresponding communication node. According to the established connection relationship in the initial adjacency matrix, determine the set of node pairs whose edge weights need to be updated.

[0052] For each node pair to be updated, the channel state information, received signal strength, and signal-to-noise ratio in the physical layer feature data are extracted and weighted and summed according to the set weighting coefficient to obtain the physical layer perception factor corresponding to this node pair. The physical layer perception factor reflects the stability of the physical connection and the channel quality.

[0053] The cosine similarity is used to calculate the similarity score between the updated node representations of the two communicating nodes in a node pair. The similarity score and the corresponding physical layer perception factor are input into a nonlinear gating function. The adjustment value of the connection weight is dynamically output based on the physical connection quality and the similarity score as the updated weight of this edge.

[0054] If the wireless communication system supports channel acquisition of multiple frequency bands or subcarriers, the channel state information of the node pair at multiple frequency points is extracted in the frequency domain to form a frequency domain feature vector. The frequency domain feature vector is normalized and weighted fused to obtain a frequency domain fusion feature value, which reflects the average effect of cross-frequency channel consistency.

[0055] The frequency domain fusion eigenvalues ​​and the edge weights output by the nonlinear gating function are jointly calculated to generate the final connection weights of the node pairs. The weighted adjacency matrix is ​​constructed using the final connection weights of all node pairs.

[0056] Based on the original node set and the updated weighted adjacency matrix, the graph structure is reconstructed to generate an evolved graph structure with physical layer perception and frequency response capabilities.

[0057] Optionally, forming a network connection solution at the current moment includes:

[0058] Based on the dynamically evolving network graph structure, the connection weights between communication node pairs in the weighted adjacency matrix are extracted;

[0059] For each source node and target node pair, a total path cost is set. The total path cost is equal to the sum of the connection weights of all edges on the path, which measures the communication cost of the path. The smaller the total path cost, the better the overall communication quality of the path.

[0060] The Dijkstra method is used to perform the initial shortest path search operation on the dynamically evolving network graph structure to obtain the initial shortest path between the source node and the target node;

[0061] After obtaining the initial shortest path, perform a perturbation operation on each intermediate node in the initial shortest path to construct several perturbation paths, specifically:

[0062] The initial shortest path is truncated to form a path prefix from the source node to an intermediate node, and an alternative path segment is found after the intermediate node to generate a new path without repeated cycles;

[0063] For each perturbed path, recalculate the total path cost, merge all generated new paths with the original path, and sort them in ascending order based on the total path cost;

[0064] The path with the smallest total cost is selected from the sorted results as the optimal path for the current source-destination node pair. The node sequence, edge set, and total cost in the path are recorded as the network connection plan. The network connection plan represents the optimal network connection structure under the current topological state and guides the signal routing selection and communication link establishment process.

[0065] When the network graph structure or connection weights change, the above path planning process is repeated to ensure that the network connection solution continues to maintain the global optimal state in a dynamic environment.

[0066] The beneficial effects of the present invention are:

[0067] The present invention proposes a method for constructing a physical layer fusion network, which fully integrates a variety of physical layer feature data, including channel state information, received signal strength, signal-to-noise ratio, channel phase, frequency offset and synchronization timestamp, etc. Through preprocessing methods such as normalization and phase alignment, high-quality node and edge feature representations are constructed. While maintaining the accuracy of physical perception, the stability and credibility of the modeling of the connection relationship between nodes are improved. Based on the constructed initial graph structure, combined with Laplace features, shortest path coding and random walk frequency multiple position embedding methods, the relative position and potential communication capability of the communication nodes in the structure are systematically expressed, providing structural enhancement support for subsequent graph representation learning.

[0068] In the process of graph representation learning, the present invention introduces a self-supervised graph pre-training mechanism, and adopts the main encoder and the target encoder for graph comparative learning, which effectively improves the discrimination and generalization capabilities of the communication node representation. The local message passing operation is performed through the graph neural network to obtain context information, and then combined with the position encoding vector for global attention modeling, and finally generates a node update representation that integrates local perception and global structure, enhancing the structural adaptability and information expression ability of the network in dynamic scenes. At the same time, it innovatively introduces a dynamic edge weight update mechanism based on physical layer perception factors and node semantic similarity, and combines it with cross-frequency domain feature fusion to form an evolutionary graph structure with physical response capability and frequency consistency perception capability.

[0069] In terms of path planning, the proposed method constructs a multi-path perturbation scheme on a dynamically evolving network structure and achieves optimal path selection through total cost evaluation. This allows it to quickly respond to changes in the network structure or fluctuations in connection weights and output high-quality network connection solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 This is a flow chart of a method for constructing a physical layer converged network proposed by the present invention;

[0072] Figure 2 A schematic diagram of generating initial representations of communication nodes by a self-supervised graph pre-training mechanism of a physical layer fusion network construction method proposed in the present invention;

[0073] Figure 3 This is a flowchart of the optimal path planning and network connection solution selection for the physical layer converged network construction method proposed by the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0075] refer to Figure 1-3 A method for constructing a physical layer converged network includes the following steps:

[0076] Collect physical layer characteristic data of each communication node in the wireless communication system and perform preprocessing;

[0077] Based on the preprocessed physical layer feature data, the node feature matrix and edge feature matrix are constructed, and the adjacency matrix is ​​generated according to the physical connection status between the communication nodes, and the initial graph structure is generated at the same time;

[0078] Generate position encoding matrix using adjacency matrix;

[0079] Based on the initial graph structure, view pairs are constructed and self-supervised graph pre-training is performed. The initial representation of the communication nodes is generated through graph comparison learning between the main encoder and the target encoder.

[0080] Perform a local message passing operation based on the initial representation of the communication node to obtain the local context node representation of the communication node;

[0081] Concatenate the local context node representation with the corresponding position encoding matrix, perform global attention modeling, and generate an updated node representation that integrates local and global information;

[0082] Based on the updated node representation, the connection weights between communication nodes are calculated, a weighted adjacency matrix is ​​constructed, and the graph connection relationship is updated according to the weighted adjacency matrix to generate a dynamically evolving network graph structure;

[0083] Based on the network graph structure, optimal path planning and communication link selection are performed to form a network connection plan at the current moment. When it is detected that the communication node has changed its position, the physical layer characteristic data has abnormal fluctuations, or the connection weight has decreased, the above steps are re-executed to achieve adaptive construction and dynamic optimization of the network graph structure.

[0084] The present invention breaks through the limitations of traditional network construction methods that rely on single indicators or static structures by constructing a network construction method that integrates multi-dimensional feature information of the physical layer and a graph structure modeling mechanism, and realizes adaptive topology evolution facing the actual channel environment of communication nodes. This method can perceive the multi-dimensional physical connection status between communication nodes, and through a dynamic composition and update mechanism, it can quickly respond to node status changes and channel fluctuations, effectively improving the stability, robustness and transmission efficiency of network construction, and is suitable for high-dynamic communication networks in heterogeneous edge device environments.

[0085] In this embodiment, the physical layer characteristic data includes channel state information value, received signal strength, signal-to-noise ratio, channel phase, frequency offset and synchronization timestamp, and the preprocessing includes outlier detection and elimination, missing value filling, phase expansion and alignment, unified dimension and normalization.

[0086] In the collection of physical layer feature data, the present invention introduces key communication parameters including channel state information value, received signal strength, signal-to-noise ratio, and channel phase, supplemented by timing information such as frequency offset and frame synchronization, so that the collected data dimensions are more comprehensive and the channel characterization capability is stronger. Through multi-step preprocessing operations such as anomaly detection, missing filling and unified dimension, the feature quality is improved, the interference caused by sensing errors is avoided, and the accuracy and consistency of subsequent feature expression are ensured, laying a high-quality data foundation for node modeling and connection weight evaluation.

[0087] In this embodiment, generating the initial graph structure includes:

[0088] The number of communication nodes in the wireless communication system is set to a fixed value, each communication node has pre-processed physical layer feature data, and the feature data of all communication nodes are combined in node number order to form a node feature matrix, where each row in the node feature matrix represents a complete physical layer feature entry corresponding to a communication node;

[0089] The adjacency matrix is ​​constructed based on whether there is a physical connection relationship between communication nodes. The specific rules are:

[0090] When the channel state information value between two communication nodes is greater than or equal to the preset connection threshold, it is considered that there is a communication link between the two communication nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of 1;

[0091] If the channel state information value is less than the preset connection threshold, it is considered that there is no stable communication path between the two nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of zero;

[0092] The connection threshold value is set as a fixed constant according to the system environment;

[0093] For all communication node pairs marked as connected in the adjacency matrix, edge feature data is extracted, and the edge features of all connected node pairs are sequentially combined to generate an edge feature matrix, where each row of the edge feature matrix corresponds to an established physical connection path;

[0094] The edge feature data includes three types of indicators: channel state information difference, received signal strength difference and channel phase difference, which describe the physical connection characteristics between communication nodes;

[0095] The node feature matrix, adjacency matrix and edge feature matrix are used as inputs for constructing the graph structure. The initial graph structure is defined in the form of graph structure triples, which serve as the basic structure input for graph neural network modeling and graph position encoding calculation.

[0096] In the process of constructing the initial graph structure, the present invention introduces a clear definition method for the node feature matrix and the edge feature matrix, and sets the connection threshold based on the channel state information to achieve structured expression of the communication relationship. By performing three-dimensional difference modeling on the edge features, the detailed expression ability of the physical connection attributes is enhanced. Combined with the adjacency matrix generation mechanism, the physical perception accuracy of the initial network structure is ensured, providing precise input for subsequent graph model reasoning and dynamic topology evolution, and improving the practical usability and robustness of the mapping effect.

[0097] In this embodiment, generating a position coding matrix includes:

[0098] Constructing a degree matrix based on the adjacency matrix to represent the connection strength of each node, where each element on the main diagonal of the degree matrix is ​​equal to the number of connections of the corresponding node, and the remaining elements are zero;

[0099] The Laplace matrix of the graph is constructed based on the element-by-element difference between the adjacency matrix and the degree matrix. The eigenvalue decomposition of the Laplace matrix is ​​performed to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The first several low-frequency eigenvectors are selected as the Laplace position code vectors of the communication nodes to capture the frequency domain structure information of the nodes.

[0100] The shortest path distance between each pair of communicating nodes is further calculated to construct a path distance matrix, where each element of the path distance matrix represents the number of hops required between two nodes. The path information of each node is then reduced in dimension using a nonlinear mapping function to form a path position encoding vector.

[0101] Based on the adjacency matrix, the transition probability between communication nodes is calculated, and a transition probability matrix is ​​constructed to represent the probability of any node jumping to its adjacent node. Based on the transition probability matrix, the frequency of each node being visited is counted through multiple rounds of random walks to generate a random walk position encoding vector.

[0102] The three types of vectors, namely the Laplace position coding vector, the path position coding vector and the random walk position coding vector, are concatenated according to the column dimension to obtain a complete position coding matrix, in which each row corresponds to the structural position vector of a communication node.

[0103] By fusing three types of position expressions, namely Laplace position coding, path distance coding and random walk position coding, the present invention enables each communication node to have three types of structural information, namely frequency domain structure, path reachability and local traversal probability, thereby achieving a comprehensive modeling of the structural role of the node in the network. The position coding matrix enhances the global positioning capability of the node representation and improves the discriminant ability of the subsequent fusion module to model the relationship between nodes, providing strong structural support for updating node representation and path selection.

[0104] In this embodiment, generating the initial representation of the communication node includes:

[0105] Based on the initial graph structure, using the node feature matrix and adjacency matrix of the communication node as input, a view pair is constructed. The view pair includes an original view and a perturbed view. The original view is a complete graph structure, and the perturbed view is a transformed graph structure with perturbations applied on the complete graph structure. The perturbations include masking some feature dimensions of the communication nodes, randomly discarding some edge connections in the adjacency matrix, and performing random subgraph sampling operations on the graph structure.

[0106] The original view and the perturbed view are fed into two graph encoders with the same parameter structure but different update strategies. The first encoder acts as the master encoder, whose parameters are updated by backpropagating gradients. The second encoder acts as the target encoder, whose parameters are updated by an exponential sliding average of the master encoder parameters. The two encoders generate a structure representation vector for each communication node.

[0107] Both the main encoder and the target encoder adopt graph attention network, which realizes weighted aggregation of information between communication nodes through multi-head attention mechanism;

[0108] The parameter update rule of the target encoder is:

[0109] ;

[0110] in, represents the parameters of the target encoder, Represents the parameters of the main encoder, represents the momentum coefficient, Represents an assignment operation, represents the momentum compensation factor;

[0111] The cosine similarity between the representation vectors of the same communication node in two views is used as the positive sample pair, and the remaining representation vector pairs between different nodes are used as negative sample pairs. A temperature coefficient is introduced to adjust the gradient amplitude, and a self-supervised contrast loss function is constructed to minimize the distance between positive samples and maximize the representation discrimination between negative samples:

[0112] ;

[0113] in, represents the self-supervised contrastive loss function, represents the natural exponential function, represents the cosine similarity between vectors, Indicates the master encoder The representation vector of each node, Indicates the target encoder’s The representation vector of each node, represents the temperature coefficient, Indicates the total number of communication nodes, Indicates the target encoder’s The representation vector of each node;

[0114] After the training is completed, the output of the main encoder is used as the structural representation initialization state of all communication nodes to generate the initial representation of the communication nodes.

[0115] This paper introduces a self-supervised graph pre-training mechanism, constructs structural perturbation view pairs, and generates initial node representations through graph contrast learning. It effectively captures the semantics of graph structure without the need for a large amount of labeled data. The design of collaborative updating of the main encoder and the target encoder improves the stability and generalization ability of the node representation. The multi-head attention mechanism further enhances the information aggregation effect. This method can more accurately characterize the relationship structure between communication nodes and provide a high-quality starting state for the subsequent context modeling of graph neural networks.

[0116] In this embodiment, the process of obtaining the local context node representation of the communication node is as follows:

[0117] ;

[0118] in, Indicates that the communication node is The local context node representation of the layer, represents the degree matrix, represents the adjacency matrix, Indicates the The weight matrix of the layer, represents a nonlinear activation function, Indicates that the communication node is The local context node representation of the layer.

[0119] The present invention adopts graph neural network for local message transmission, uses adjacency matrix and degree matrix to complete feature aggregation between nodes, obtains node representation with local context structure semantics, and adjusts through nonlinear activation function and weight matrix, so that the model can adaptively strengthen or suppress the information propagation path according to the adjacency relationship, retain the effective neighbor features, and filter out redundant or irrelevant information, thereby enhancing the expressive ability of the communication node representation and facilitating the recognition of local relationships under complex topological structures.

[0120] In this embodiment, generating an updated node representation that integrates local and global information includes:

[0121] For each communication node, obtain the local context node representation and extract the position encoding vector of the corresponding position from the position encoding matrix, which respectively represent the structural semantics and position information of the communication node;

[0122] Concatenate the local context node representation and the corresponding position encoding vector according to the feature dimension to generate a fused input vector;

[0123] The fusion input vectors of all communication nodes are constructed into a fusion input matrix in the order of node index;

[0124] Perform global modeling, and use the multi-head attention mechanism to interact with and adjust the weights of the fusion input matrices of each communication node to generate updated node representations that fuse local and global information.

[0125] Based on the fusion of node local representation and structural position encoding, the present invention adopts a multi-head attention mechanism for global modeling operations, fully explores the potential dependencies of communication nodes in the entire graph, enhances the input information density through feature splicing, and captures multi-scale structural semantics through input fusion, thereby improving the discriminability of updated node representation and the overall topological modeling capability. This strategy realizes local-global collaborative perception and enhances the adaptability and robustness of graph structure expression and subsequent edge weight updates.

[0126] In this embodiment, generating a dynamically evolving network graph structure includes:

[0127] Extract the embedding vector of each communication node from the updated node representation and set it as the updated node representation of the corresponding communication node. According to the established connection relationship in the initial adjacency matrix, determine the set of node pairs whose edge weights need to be updated.

[0128] For each node pair to be updated, the channel state information, received signal strength, and signal-to-noise ratio in the physical layer feature data are extracted and weighted and summed according to the set weighting coefficient to obtain the physical layer perception factor corresponding to this node pair. The physical layer perception factor reflects the stability of the physical connection and the channel quality.

[0129] The cosine similarity is used to calculate the similarity score between the updated node representations of the two communicating nodes in a node pair. The similarity score and the corresponding physical layer perception factor are input into a nonlinear gating function. The adjustment value of the connection weight is dynamically output based on the physical connection quality and the similarity score as the updated weight of this edge.

[0130] If the wireless communication system supports channel acquisition of multiple frequency bands or subcarriers, the channel state information of the node pair at multiple frequency points is extracted in the frequency domain to form a frequency domain feature vector. The frequency domain feature vector is normalized and weighted fused to obtain a frequency domain fusion feature value, which reflects the average effect of cross-frequency channel consistency.

[0131] The frequency domain fusion eigenvalues ​​and the edge weights output by the nonlinear gating function are jointly calculated to generate the final connection weights of the node pairs. The weighted adjacency matrix is ​​constructed using the final connection weights of all node pairs.

[0132] Based on the original node set and the updated weighted adjacency matrix, the graph structure is reconstructed to generate an evolved graph structure with physical layer perception and frequency response capabilities.

[0133] The present invention integrates the embedding vector similarity of communication nodes and the physical layer perception factor, designs a nonlinear gating function to dynamically adjust the connection weight, and realizes fine-grained modeling of edge relationships. At the same time, it introduces a frequency domain fusion mechanism to enhance the mapping ability to perceive the cross-frequency channel response characteristics. The evolutionary graph structure is constructed based on the updated weighted adjacency matrix, which improves the sensitivity of the graph structure to channel changes and the adaptability of the network construction strategy. This method enhances the modeling effect of the topological structure on heterogeneous communication environments.

[0134] In this embodiment, forming a network connection solution at the current moment includes:

[0135] Based on the dynamically evolving network graph structure, the connection weights between communication node pairs in the weighted adjacency matrix are extracted;

[0136] For each source node and target node pair, a total path cost is set. The total path cost is equal to the sum of the connection weights of all edges on the path, which measures the communication cost of the path. The smaller the total path cost, the better the overall communication quality of the path.

[0137] The Dijkstra method is used to perform the initial shortest path search operation on the dynamically evolving network graph structure to obtain the initial shortest path between the source node and the target node;

[0138] After obtaining the initial shortest path, perform a perturbation operation on each intermediate node in the initial shortest path to construct several perturbation paths, specifically:

[0139] The initial shortest path is truncated to form a path prefix from the source node to an intermediate node, and an alternative path segment is found after the intermediate node to generate a new path without repeated cycles;

[0140] For each perturbed path, recalculate the total path cost, merge all generated new paths with the original path, and sort them in ascending order based on the total path cost;

[0141] The path with the smallest total cost is selected from the sorted results as the optimal path for the current source-destination node pair. The node sequence, edge set, and total cost in the path are recorded as the network connection plan. The network connection plan represents the optimal network connection structure under the current topological state and guides the signal routing selection and communication link establishment process.

[0142] When the network graph structure or connection weights change, the above path planning process is repeated to ensure that the network connection solution continues to maintain the global optimal state in a dynamic environment.

[0143] The present invention constructs an optimal network construction path planning scheme driven by connection weights, dynamically screens the path with the best communication quality through the total path cost function, and combines path perturbation operations to improve path diversity and fault tolerance, ensuring the global optimality and link stability of the connection scheme in the dynamic evolution of the network. This path planning method is highly adaptable and real-time, and is particularly suitable for dynamic network construction tasks in environments with frequent device movement or drastic channel changes.

[0144] Example 1:

[0145] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a 5G micro base station communication system in a coastal city. The coverage area of ​​this communication system is densely populated, the distribution of communication nodes is complex, and it is frequently interfered with by environmental factors such as building reflections, spectrum congestion, and multipath effects. The traditional static topology network construction method has obvious defects in such an environment. The network topology responds slowly, the link quality fluctuates greatly, and misconnections and link jitters frequently occur, seriously affecting communication reliability and scheduling efficiency.

[0146] The present invention is deployed in the wireless communication cluster system in the southern part of the city, with micro base stations as communication nodes. Through the underlying driver, it extracts physical layer feature data including channel state information, received signal strength, channel phase, frequency offset and frame synchronization time, and uploads the collected data to the edge computing platform. In the data center, the node feature matrix and edge feature matrix are constructed through normalization and standardization processing, and the adjacency matrix is ​​generated according to the connection threshold to further form the graph structure input. After the graph construction is completed, self-supervised graph pre-training is performed to integrate the structural differences between nodes and guide the stable convergence of the graph neural network parameters. Subsequently, the system executes local message passing operations to obtain local context representation, and then combines the graph position encoding for global information modeling to generate the fused node representation, and finally calculates the edge weights and dynamically updates the topology structure.

[0147] After being generated, this topology structure is used in the path optimization process. Initial path planning is performed in combination with the Dijkstra method, and multi-path perturbation and global path optimization are performed. This makes the final network connection solution more adaptable and stable. During a week of actual operation, the system automatically triggers the topology reconstruction mechanism every day based on the actual link status, allowing the network to maintain a high connectivity rate and stability under different communication periods and environmental disturbances.

[0148] In order to verify the performance of the present invention in implementation, a comprehensive performance comparison was conducted with a traditional physical layer fusion network method.

[0149] Table 1. Experimental comparison between the present invention and the traditional physical layer fusion network method.

[0150]

[0151] As can be seen from Table 1, the present invention shows advantages over traditional methods in multiple key performance indicators, which fully demonstrates the effectiveness and advancement of the proposed method in actual communication network construction tasks.

[0152] In terms of topology adaptation delay, the method of the present invention shortens the time required for network structure adjustment to 168ms, compared with 342ms of the traditional method, indicating that it has faster network adaptation capabilities and is particularly suitable for dynamic communication environments with frequent node changes. In terms of path planning success rate, an indicator that directly reflects the effectiveness of network construction, the present invention reaches 90.6%, which is higher than the 82.1% of the traditional method, indicating that the proposed self-supervised representation strategy has outstanding effects in enhancing global path cognition.

[0153] In terms of edge weight update accuracy, the method of the present invention reduces the error to 0.023, which is a significant decrease compared to 0.082 of the traditional solution, further verifying its robustness in physical layer perception and connection weight modeling. At the same time, the packet loss rate per unit time in terms of communication link stability is reduced from 0.071 of the traditional method to 0.030, indicating that the constructed network has stronger anti-interference ability and transmission reliability.

[0154] In the multi-frequency domain consistency scoring dimension, the average consistency score of the proposed method after integrating frequency domain information was 0.890, higher than the 0.630 of the traditional method, reflecting a significant improvement in channel stability in a multi-band environment. The frequency of adaptive network establishment was also increased from the traditional 1.4 times / hour to 4.2 times / hour, demonstrating its more dynamic perception and real-time responsiveness.

[0155] In terms of the global path reconstruction rate, the present invention reaches 84.5%, which is higher than the 56.7% of the traditional method, further reflecting its global adaptability to the evolution of network topology. In terms of the average total path cost, it is reduced from 0.720 of the traditional method to 0.350, which means that the selected path has lower communication costs and energy consumption levels.

[0156] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for constructing a physical layer converged network, characterized in that: The steps include: Collect physical layer characteristic data of each communication node in the wireless communication system and perform preprocessing; Based on the preprocessed physical layer feature data, the node feature matrix and edge feature matrix are constructed, and the adjacency matrix is ​​generated according to the physical connection status between the communication nodes, and the initial graph structure is generated at the same time; Generate position encoding matrix using adjacency matrix; Based on the initial graph structure, view pairs are constructed and self-supervised graph pre-training is performed. The initial representation of the communication nodes is generated through graph comparison learning between the main encoder and the target encoder. Perform a local message passing operation based on the initial representation of the communication node to obtain the local context node representation of the communication node; Concatenate the local context node representation with the corresponding position encoding matrix, perform global attention modeling, and generate an updated node representation that integrates local and global information; Based on the updated node representation, the connection weights between communication nodes are calculated, a weighted adjacency matrix is ​​constructed, and the graph connection relationship is updated according to the weighted adjacency matrix to generate a dynamically evolving network graph structure; Based on the network graph structure, optimal path planning and communication link selection are performed to form a network connection plan at the current moment. When a change in the position of a communication node, abnormal fluctuations in physical layer characteristic data, or a decrease in connection weight is detected, the above steps are re-executed to achieve adaptive construction and dynamic optimization of the network graph structure. The generating of the dynamically evolving network graph structure includes: Extract the embedding vector of each communication node from the updated node representation and set it as the updated node representation of the corresponding communication node. According to the established connection relationship in the initial adjacency matrix, determine the set of node pairs whose edge weights need to be updated. For each node pair to be updated, the channel state information, received signal strength, and signal-to-noise ratio in the physical layer feature data are extracted and weighted and summed according to the set weighting coefficient to obtain the physical layer perception factor corresponding to this node pair. The physical layer perception factor reflects the stability of the physical connection and the channel quality. The cosine similarity is used to calculate the similarity score between the updated node representations of the two communicating nodes in a node pair. The similarity score and the corresponding physical layer perception factor are input into a nonlinear gating function. The adjustment value of the connection weight is dynamically output based on the physical connection quality and the similarity score as the updated weight of this edge. If the wireless communication system supports channel acquisition of multiple frequency bands or subcarriers, the channel state information of the node pair at multiple frequency points is extracted in the frequency domain to form a frequency domain feature vector. The frequency domain feature vector is normalized and weighted fused to obtain a frequency domain fusion feature value, which reflects the average effect of cross-frequency channel consistency. The frequency domain fusion eigenvalues ​​and the edge weights output by the nonlinear gating function are jointly calculated to generate the final connection weights of the node pairs. The weighted adjacency matrix is ​​constructed using the final connection weights of all node pairs. Based on the original node set and the updated weighted adjacency matrix, the graph structure is reconstructed to generate an evolved graph structure with physical layer perception and frequency response capabilities.

2. A method for constructing a physical layer converged network according to claim 1, characterized in that: The physical layer characteristic data includes channel state information value, received signal strength, signal-to-noise ratio, channel phase, frequency offset and synchronization timestamp, and the preprocessing includes outlier detection and elimination, missing value filling, phase expansion and alignment, unified dimension and normalization.

3. The method for constructing a physical layer converged network according to claim 1, wherein: Generating the initial graph structure includes: The number of communication nodes in the wireless communication system is set to a fixed value, each communication node has pre-processed physical layer feature data, and the feature data of all communication nodes are combined in node number order to form a node feature matrix, where each row in the node feature matrix represents a complete physical layer feature entry corresponding to a communication node; The adjacency matrix is ​​constructed based on whether there is a physical connection relationship between communication nodes. The specific rules are: When the channel state information value between two communication nodes is greater than or equal to the preset connection threshold, it is considered that there is a communication link between the two communication nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of 1; If the channel state information value is less than the preset connection threshold, it is considered that there is no stable communication path between the two nodes, and the corresponding position in the adjacency matrix is ​​assigned a value of zero; The connection threshold value is set as a fixed constant according to the system environment; For all communication node pairs marked as connected in the adjacency matrix, edge feature data is extracted, and the edge features of all connected node pairs are sequentially combined to generate an edge feature matrix, where each row of the edge feature matrix corresponds to an established physical connection path; The edge feature data includes three types of indicators: channel state information difference, received signal strength difference and channel phase difference, which describe the physical connection characteristics between communication nodes; The node feature matrix, adjacency matrix and edge feature matrix are used as inputs for constructing the graph structure. The initial graph structure is defined in the form of graph structure triples, which serve as the basic structure input for graph neural network modeling and graph position encoding calculation.

4. The method for constructing a physical layer converged network according to claim 1, wherein: Generating a position encoding matrix comprises: Constructing a degree matrix based on the adjacency matrix to represent the connection strength of each node, where each element on the main diagonal of the degree matrix is ​​equal to the number of connections of the corresponding node, and the remaining elements are zero; The Laplace matrix of the graph is constructed based on the element-by-element difference between the adjacency matrix and the degree matrix. The eigenvalue decomposition of the Laplace matrix is ​​performed to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The first several low-frequency eigenvectors are selected as the Laplace position code vectors of the communication nodes to capture the frequency domain structure information of the nodes. The shortest path distance between each pair of communicating nodes is further calculated to construct a path distance matrix, where each element of the path distance matrix represents the number of hops required between two nodes. The path information of each node is then reduced in dimension using a nonlinear mapping function to form a path position encoding vector. Based on the adjacency matrix, the transition probability between communication nodes is calculated, and a transition probability matrix is ​​constructed to represent the probability of any node jumping to its adjacent node. Based on the transition probability matrix, the frequency of each node being visited is counted through multiple rounds of random walks to generate a random walk position encoding vector. The three types of vectors, namely the Laplace position coding vector, the path position coding vector and the random walk position coding vector, are concatenated according to the column dimension to obtain a complete position coding matrix, in which each row corresponds to the structural position vector of a communication node.

5. The method for constructing a physical layer converged network according to claim 1, wherein: Generating the initial representation of the communication node includes: Based on the initial graph structure, using the node feature matrix and adjacency matrix of the communication node as input, a view pair is constructed. The view pair includes an original view and a perturbed view. The original view is a complete graph structure, and the perturbed view is a transformed graph structure with perturbations applied on the complete graph structure. The perturbations include masking some feature dimensions of the communication nodes, randomly discarding some edge connections in the adjacency matrix, and performing random subgraph sampling operations on the graph structure. The original view and the perturbed view are fed into two graph encoders with the same parameter structure but different update strategies. The first encoder acts as the master encoder, whose parameters are updated by backpropagating gradients. The second encoder acts as the target encoder, whose parameters are updated by an exponential sliding average of the master encoder parameters. The two encoders generate a structure representation vector for each communication node. Both the main encoder and the target encoder adopt graph attention network, which realizes weighted aggregation of information between communication nodes through multi-head attention mechanism; The parameter update rule of the target encoder is: ; in, represents the parameters of the target encoder, Represents the parameters of the main encoder, represents the momentum coefficient, Represents an assignment operation, represents the momentum compensation factor; The cosine similarity between the representation vectors of the same communication node in two views is used as the positive sample pair, and the remaining representation vector pairs between different nodes are used as negative sample pairs. A temperature coefficient is introduced to adjust the gradient amplitude, and a self-supervised contrastive loss function is constructed to minimize the distance between positive samples and maximize the representation discrimination between negative samples. After the training is completed, the output of the main encoder is used as the structural representation initialization state of all communication nodes to generate the initial representation of the communication nodes.

6. A method for constructing a physical layer converged network according to claim 1, characterized in that: Generating an updated node representation that fuses local and global information includes: For each communication node, obtain the local context node representation and extract the position encoding vector of the corresponding position from the position encoding matrix, which respectively represent the structural semantics and position information of the communication node; Concatenate the local context node representation and the corresponding position encoding vector according to the feature dimension to generate a fused input vector; The fusion input vectors of all communication nodes are constructed into a fusion input matrix in the order of node index; Perform global modeling, and use the multi-head attention mechanism to interact with and adjust the weights of the fusion input matrices of each communication node to generate updated node representations that fuse local and global information.

7. The method for constructing a physical layer converged network according to claim 1, wherein: The network connection solution formed at the current moment includes: Based on the dynamically evolving network graph structure, the connection weights between communication node pairs in the weighted adjacency matrix are extracted; For each source node and target node pair, set the total path cost; The Dijkstra method is used to perform the initial shortest path search operation on the dynamically evolving network graph structure to obtain the initial shortest path between the source node and the target node; After obtaining the initial shortest path, perform a perturbation operation on each intermediate node in the initial shortest path to construct several perturbation paths, specifically: The initial shortest path is truncated to form a path prefix from the source node to an intermediate node, and an alternative path segment is found after the intermediate node to generate a new path without repeated cycles; For each perturbed path, recalculate the total path cost, merge all generated new paths with the original path, and sort them in ascending order based on the total path cost; The path with the smallest total cost is selected from the sorted results as the optimal path for the current source-destination node pair. The node sequence, edge set, and total cost in the path are recorded as the network connection plan. The network connection plan represents the optimal network connection structure under the current topological state and guides the signal routing selection and communication link establishment process. When the network graph structure or connection weights change, the above path planning process is repeated to ensure that the network connection solution continues to maintain the global optimal state in a dynamic environment.

Images