A machine learning-based networking method and system for internet of vehicles

Abstract

The application discloses a kind of machine learning-based vehicle networking system networking method and system, comprising: collecting vehicle position, speed, electric quantity and link delay, packet loss rate, signal-to-noise ratio;Vehicle terminal, road side unit and base station are defined as node, link index is regarded as edge attribute, and directed graph is constructed;Determine the number of reserved adjacent edges and neighborhood pruning threshold, obtain sparse adjacency matrix;Sparse adjacency matrix is input into improved SE-GCN model, and the node stability probability and access priority are output;Hierarchical quantization mapping generates access priority lookup table and transmit power lookup table;Lookup table switches link and adjusts power, completes no inference topology reconstruction;Receive link feedback, update threshold and model weight, upload cloud platform in low load.The application realizes vehicle networking link stable access, low-latency link switching and low-load networking in high-speed mobile scenario by dynamic neighborhood pruning and improved SE-GCN model.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and vehicle-to-everything (V2X) communication technology, and in particular to a machine learning-based V2X system networking method and system. Background Technology

[0002] With the development of intelligent transportation, autonomous driving, and vehicle-road cooperative technologies, the Internet of Vehicles (IoV) is gradually becoming an important foundation for vehicle information interaction, road condition perception, and traffic cooperative control. IoV systems typically achieve vehicle status uploading, road information sharing, collaborative early warning, and network resource scheduling through communication connections between vehicle terminals, roadside units, base stations, and edge servers. Because vehicles are in a high-speed moving state, the link relationships between vehicles and roadside units, base stations, and other vehicles continuously change with location, speed, and road density. IoV networking methods need to maintain stable connections in complex and dynamic environments while balancing communication latency, link reliability, and terminal power consumption.

[0003] In existing technologies, vehicle-to-everything (V2X) networking often relies on received signal strength, fixed distance thresholds, preset access priorities, or static routing rules to complete link selection and topology adjustments. Some methods also incorporate machine learning models to predict link quality or vehicle movement trends, and then perform access switching or path selection based on the prediction results. While these methods can improve V2X communication quality to some extent, fixed rules are difficult to adapt to topology changes in scenarios involving high-speed movement, dense node density, or rapid fluctuations in link quality. Furthermore, ordinary learning models often process link prediction, neighborhood selection, and networking decisions separately, resulting in long computational chains, insufficient response speeds, and problems such as frequent link interruptions, increased reconnection latency, and unstable access selection.

[0004] Furthermore, existing graph-based vehicle-to-everything (V2X) networking methods typically construct graph structures using fixed neighborhoods or uniform thresholds, failing to adequately consider the correlation between vehicle speed, node density, and link confidence. This makes it difficult to adaptively adjust the neighborhood size between sparse high-speed and dense low-speed scenarios. While some methods can output link scores or node stability results, they still require the vehicle terminal to continue performing inference calculations or policy searches, increasing the computational burden and energy consumption of onboard equipment and hindering millisecond-level link switching. Simultaneously, link selection and transmit power adjustment are often performed separately, easily leading to an imbalance between communication reliability and terminal energy consumption.

[0005] Therefore, how to provide a networking method and system for vehicle networking based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a machine learning-based networking method and system for vehicle-to-everything (V2X) systems. This invention uses an improved SE-GCN model as its core, combines vehicle status data and link quality data to construct a dynamic directed graph, and through the synergistic effect of neighborhood pruning thresholds and model-internal edge filtering operators, completes the prediction of node stability and link access priority, generating access priority lookup tables and transmit power lookup tables. This enables vehicle terminals to complete link switching and power adjustment without performing model inference, achieving adaptive reconstruction of the V2X topology. It has the advantages of high link stability, fast networking response, and low vehicle-side computational burden.

[0007] A machine learning-based vehicle networking system networking method according to an embodiment of the present invention includes: The system periodically collects vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio, and uploads them to the edge server. In the edge server, vehicle terminals, roadside units, and base stations are defined as nodes, and link latency, link packet loss rate, and received signal-to-noise ratio are defined as edge attributes to construct a dynamic directed graph. Calculate the average vehicle speed and node density, adaptively determine the number of neighboring edges to retain and the neighborhood pruning threshold, perform neighborhood pruning on the dynamic directed graph, and obtain a sparse adjacency matrix. The sparse adjacency matrix is ​​input into the improved SE-GCN model. Before each convolution, channel attention gating is used to recalibrate the node features. During convolution, the edge filtering operator that shares weights with the neighborhood pruning threshold is called to simultaneously complete feature aggregation and neighborhood compression, and output the set of node stability probability and the set of link access priority. A hierarchical quantitative mapping is performed on the node stability probability set and the link access priority set to generate access priority lookup table and transmit power lookup table, which are then encapsulated in a dual-index format and sent to the vehicle terminal. The vehicle terminal switches links according to access priority and adjusts the transmission power according to the transmission power table to complete the inference-free topology reconstruction. The edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, updates the neighborhood pruning threshold and improves the SE-GCN model weights, and uploads the updated weights to the cloud platform storage during periods of low business load.

[0008] Optionally, the periodic collection of vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio includes: The vehicle terminal reads the vehicle positioning module, vehicle speed sensing module and battery management module according to a preset sampling period to obtain the vehicle position, vehicle speed and vehicle remaining power. The vehicle terminal performs link detection with the roadside unit and base station, and records the detection message sending time, acknowledgment reception time, number of lost messages, number of sent messages and received signal strength to obtain link delay, link packet loss rate and received signal-to-noise ratio; The vehicle terminal encapsulates vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio according to vehicle terminal identifier, link object identifier, and collection timestamp, and uploads them to the edge server.

[0009] Optionally, constructing the dynamic directed graph includes: The edge server receives vehicle location, vehicle speed, vehicle remaining battery power, link latency, link packet loss rate and received signal-to-noise ratio, and aligns them according to vehicle terminal identifier, roadside unit identifier, base station identifier and collection timestamp; The vehicle terminal, roadside unit, and base station are converted into graph nodes, and the communication relationships between the vehicle terminal and the roadside unit, base station, and other vehicle terminals are converted into directed edges. Link latency, link packet loss rate, and received signal-to-noise ratio are written into the directed edge attributes, and vehicle location, vehicle speed, and vehicle remaining battery power are written into the vehicle terminal node attributes to generate a dynamic directed graph.

[0010] Optionally, the adaptive determination of the number of neighboring edges to be retained and the neighborhood pruning threshold, and the performance of neighborhood pruning on the dynamic directed graph to obtain a sparse adjacency matrix, includes: The average vehicle speed is calculated based on the vehicle speed of the vehicle terminal nodes in the dynamic directed graph, and the node density is calculated based on the number of vehicle terminal nodes and the current road coverage length. The number of neighboring edges to be retained is determined based on the average vehicle speed, node density, and a preset upper limit for neighboring edges. Specifically, the number of neighboring edges to be retained decreases when the average vehicle speed increases, and the number of neighboring edges to be retained increases when the node density increases. The link confidence of directed edges is calculated based on link delay, link packet loss rate and received signal-to-noise ratio, and the neighborhood pruning threshold is determined based on the link confidence distribution of the same node connecting neighboring edges. Delete neighboring edges whose link confidence is lower than the neighborhood pruning threshold. Sort the neighboring edges whose link confidence is not lower than the neighborhood pruning threshold in descending order of link confidence. Then, select neighboring edges sequentially from the first position of the sorted list according to the number of neighboring edges to be retained, and generate a sparse adjacency matrix.

[0011] Optionally, the output node stability probability set and the link access priority set include: An improved SE-GCN model is constructed, which includes a node-edge feature encoding layer, a channel attention gating layer, a threshold-shared edge filtering layer, a pruned coupled graph convolutional layer, and a network result output layer. The sparse adjacency matrix, vehicle position, vehicle speed, vehicle remaining battery power, link delay, link packet loss rate, and received signal-to-noise ratio are input into the node edge feature encoding layer to generate the node encoding feature set and the edge encoding feature set. The node encoded feature set is input into the channel attention gating layer, channel compression statistics are performed to generate channel statistical vectors and gating coefficients, and the node encoded feature set is recalibrated channel by channel to generate the gated node feature set. The edge encoding feature set, link confidence, and neighborhood pruning threshold are input into the threshold shared edge filtering layer. Shared weight parameters are generated based on the neighborhood pruning threshold, and edge filtering operators are generated by combining the link confidence and edge encoding feature set. The retained neighboring edges are weighted and filtered to generate the filtered edge feature set. The gating node feature set, the filtered edge feature set, and the sparse adjacency matrix are input into the pruning coupled graph convolutional layer, and neighborhood aggregation and neighborhood compression are performed simultaneously to generate the pruning coupled convolutional feature set. The pruned coupled convolution feature set is input into the network result output layer, and link priority mapping and node stability probability mapping are performed to generate link access priority set and node stability probability set. Obtain historical link switching results, historical link interruption results, and historical node connection maintenance results; construct a set of supervised labels; calculate link priority prediction error, node stability probability prediction error, and edge filtering operator output error; combine them to form a joint optimization objective; iteratively update and improve the SE-GCN model parameters according to the joint optimization objective; and end training when the preset number of iterations is reached or the change in the joint optimization objective is lower than a preset threshold.

[0012] Optionally, the step of performing hierarchical quantization mapping on the node stability probability set and the link access priority set to generate access priority lookup tables and transmit power lookup tables includes: Based on the vehicle terminal identifier and the candidate connection object identifier, the node stability probability and link access priority are matched accordingly to generate a candidate link score set; The candidate link score set is normalized, and the nodes are divided into low-stability, medium-stability and high-stability layers according to their stability probability. The links are also divided into low-priority, medium-priority and high-priority layers according to their access priority. The three types of stabilization layers are cross-mapped with the three types of priority layers to form nine types of quantization mapping units, and access ordering identifiers and transmit power levels are configured. Based on the quantization mapping unit to which the candidate link score set belongs, determine the access sorting identifier and transmit power level of the candidate connection object corresponding to each vehicle terminal; Candidate connection objects for the same vehicle terminal are sorted according to the access sorting identifier, and an access priority lookup table is generated. A lookup table for transmission power is generated based on the correspondence between vehicle terminal identifier, candidate connection object identifier, and transmission power level. Using vehicle terminal identifier and candidate connection object identifier as dual indexes, the access priority lookup table and transmit power lookup table are encapsulated into a network control table and sent to the vehicle terminal.

[0013] Optionally, the vehicle terminal switches links according to access priority and adjusts the transmission power according to transmission power by looking up a table, including: The vehicle terminal receives access priority lookup table and transmit power lookup table, and reads the double index item according to vehicle terminal identifier and candidate connection object identifier to obtain target access object identifier and target transmit power level; The vehicle terminal performs link switching based on the target access object identifier, adjusts the wireless transmission power according to the target transmission power level, and establishes a communication connection between the vehicle terminal and the target access object. The vehicle terminal updates its local adjacency table and node connection status table based on the target access object identifier, and completes the topology connection update by looking up the table based on access priority and transmit power, thus forming the network topology structure after inference-free reconstruction.

[0014] Optionally, the edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, and updates the neighborhood pruning threshold and the weights of the improved SE-GCN model, including: The edge server receives link latency, link packet loss rate, received signal-to-noise ratio and link switching results from the vehicle terminal, and generates a set of link feedback records by associating and matching them according to the vehicle terminal identifier and timestamp. Based on the link feedback record set, statistical analysis is performed on the trends of link delay, link packet loss rate, and received signal-to-noise ratio. The link confidence is then recalculated, and the link confidence distribution and neighborhood pruning threshold are updated. The link feedback record set is input into the improved SE-GCN model. The model weights are updated based on the deviation between the predicted link access priority and the actual link switching results, and the deviation between the predicted node stability probability and the actual connection maintenance results, thus generating an updated weight set. The system detects the business load status of the edge server. When the business load is lower than a preset threshold, it updates the weight set by model level, parameter type and timestamp, and uploads it to the cloud platform for storage.

[0015] According to an embodiment of the present invention, a vehicle-to-everything (V2X) system networking system based on machine learning includes the following modules: The status data acquisition module is used to acquire vehicle operating status and communication link status, and generate vehicle network status data; The dynamic graph construction module is used to build a dynamic directed graph between vehicle terminals, roadside units, and base stations based on vehicle network status data; The neighborhood pruning module is used to determine the number of neighboring edges to retain and the neighborhood pruning threshold based on the dynamic directed graph, and to generate a sparse adjacency matrix. An improved model inference module is used to input the sparse adjacency matrix into the improved SE-GCN model to generate a set of node stability probabilities and a set of link access priorities. The lookup mapping generation module is used to perform hierarchical quantization mapping on the node stability probability set and the link access priority set, and generate access priority lookup table and transmit power lookup table; The inference-free reconfiguration module is used to complete link switching, transmit power adjustment and topology reconfiguration by looking up tables based on access priority and transmit power. The feedback update module is used to update the neighborhood pruning threshold and improve the SE-GCN model weights based on the link status feedback, and upload the updated weights to the cloud platform for storage.

[0016] The beneficial effects of this invention are: This invention discloses a machine learning-based networking method and system for vehicle-to-everything (V2X) systems. By collecting vehicle location, speed, remaining battery power, link latency, packet loss rate, and received signal-to-noise ratio (SNR), and constructing a dynamic directed graph containing vehicle terminals, roadside units, and base stations on an edge server, the system allows for the representation of vehicle operating status, link quality status, and infrastructure connectivity within the same graph structure. Compared to networking methods based solely on received signal strength, fixed distance thresholds, or static access rules, this invention more comprehensively reflects the dynamic changes between vehicle nodes and communication links in high-speed mobile scenarios, providing a more stable data foundation for link selection and topology adjustments.

[0017] This invention adaptively determines the number of neighboring edges to retain and the neighborhood pruning threshold based on average vehicle speed, node density, and link confidence. It then generates a sparse adjacency matrix based on the neighborhood pruning threshold, enabling the vehicle-to-everything (V2X) network construction process to dynamically adjust the range of neighboring edges involved in the calculation according to changes in traffic flow and link quality. In this way, the invention reduces the interference of low-confidence links and invalid neighbors on network construction decisions, overcomes the difficulty of traditional fixed-neighbor or uniform-threshold methods in adapting to sparse high-speed and dense low-speed scenarios, improves the rationality and stability of V2X topology construction, and reduces the probability of frequent link interruptions and repeated handovers.

[0018] This invention inputs a sparse adjacency matrix into an improved SE-GCN model, recalibrates node features through channel attention gating, and suppresses low-confidence links using an edge filtering operator that shares weights with the neighborhood pruning threshold. This allows node feature aggregation and neighborhood compression to be completed in the same inference process, thereby improving the accuracy of node stability probability and link access priority judgment. Furthermore, this invention performs hierarchical quantization mapping on the node stability probability set and the link access priority set, generating access priority lookup tables and transmit power lookup tables. This enables vehicle terminals to complete link switching and transmit power adjustment through table lookup, eliminating the need for further model inference or policy search. This reduces onboard computing burden and terminal energy consumption, improving the networking response speed, communication reliability, and energy balance capability of the vehicle-to-everything (V2X) system in high-speed, densely populated, and link-fluctuation scenarios. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based vehicle networking system networking method proposed in this invention; Figure 2 This is a schematic diagram of the improved SE-GCN model structure of a machine learning-based vehicle networking system proposed in this invention. Figure 3 This is a schematic diagram of the network structure of a vehicle-to-everything (V2X) system based on machine learning proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 and Figure 2 A machine learning-based method for networking vehicle-to-everything (V2X) systems includes: The system periodically collects vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio, and uploads them to the edge server. In the edge server, vehicle terminals, roadside units, and base stations are defined as nodes, and link latency, link packet loss rate, and received signal-to-noise ratio are defined as edge attributes to construct a dynamic directed graph. Calculate the average vehicle speed and node density, adaptively determine the number of neighboring edges to retain and the neighborhood pruning threshold, perform neighborhood pruning on the dynamic directed graph, and obtain a sparse adjacency matrix. The sparse adjacency matrix is ​​input into the improved SE-GCN model. Before each convolution, channel attention gating is used to recalibrate the node features. During convolution, the edge filtering operator that shares weights with the neighborhood pruning threshold is called to simultaneously complete feature aggregation and neighborhood compression, and output the set of node stability probability and the set of link access priority. A hierarchical quantitative mapping is performed on the node stability probability set and the link access priority set to generate access priority lookup table and transmit power lookup table, which are then encapsulated in a dual-index format and sent to the vehicle terminal. The vehicle terminal switches links according to access priority and adjusts the transmission power according to the transmission power table to complete the inference-free topology reconstruction. The edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, updates the neighborhood pruning threshold and improves the SE-GCN model weights, and uploads the updated weights to the cloud platform storage during periods of low business load.

[0022] In this embodiment, the periodic collection of vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio includes: The vehicle terminal reads data from the onboard positioning module, vehicle speed sensing module, and battery management module according to a preset sampling period to obtain the vehicle's location, speed, and remaining battery power. The preset sampling period is determined as follows: The sampling time interval is configured based on the frequency of vehicle speed changes, the degree of link status fluctuation, and the processing capacity of the edge server. The sampling time interval is reduced when the vehicle is moving at high speed or the amplitude of link status changes increases, and the sampling time interval is increased when the vehicle is running at low speed or the link status is stable. The configuration result is written into the vehicle terminal sampling parameter table, and the vehicle terminal triggers data reading according to the corresponding time interval period in the sampling parameter table. The vehicle terminal performs link detection with roadside units and base stations, recording the transmission time of detection messages, acknowledgment time, number of lost messages, number of transmitted messages, and received signal strength. This yields link delay, link packet loss rate, and received signal-to-noise ratio (SNR). Specifically, the link delay, link packet loss rate, and SNR are obtained as follows: The vehicle terminal sends probe messages with transmission timestamps and message sequence numbers to the roadside unit and the base station, and receives acknowledgment messages returned by the roadside unit and the base station. It generates a single round-trip delay based on the difference between the acknowledgment reception time and the probe message transmission time, and averages the single round-trip delays over a continuous sampling period to obtain the link delay. It determines the number of lost messages based on the difference between the number of sent messages and the number of acknowledgment messages, and uses the ratio of the number of lost messages to the number of sent messages as the link packet loss rate. It reads the received signal strength and background noise intensity corresponding to the acknowledgment message, calculates the difference between the received signal strength and the background noise intensity, and obtains the received signal-to-noise ratio. The vehicle terminal encapsulates vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio (SNR) according to the vehicle terminal identifier, link object identifier, and collection timestamp, and uploads it to the edge server. Specifically, the encapsulation of vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received SNR according to the vehicle terminal identifier, link object identifier, and collection timestamp is as follows: Read the vehicle terminal identifier, link object identifier, and collection timestamp; establish status data field and link data field; write vehicle location, vehicle speed, and remaining vehicle battery power into the status data field; write link delay, link packet loss rate, and received signal-to-noise ratio into the link data field; generate network status data packets in the order of vehicle terminal identifier, link object identifier, collection timestamp, status data field, and link data field; and add checksum and sequence number to the network status data packets.

[0023] In this embodiment, constructing the dynamic directed graph includes: The edge server receives vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio. This data is then aligned using vehicle terminal identifier, roadside unit identifier, base station identifier, and data collection timestamp. Specifically, the alignment using these identifiers is as follows: Read data records corresponding to vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio; extract vehicle terminal identifier, roadside unit identifier, base station identifier, and collection timestamp; establish a vehicle data index set according to the vehicle terminal identifier, and establish a link object index set according to the roadside unit identifier and base station identifier; complete the association and matching of vehicle data and link data based on the connection relationship between the vehicle terminal identifier and the link object index set; divide the time window according to the collection timestamp, and group data records that are in the same time window and whose vehicle terminal identifier and link object identifier correspond to the same network status unit to form a data set that has completed identifier association and time synchronization. The vehicle terminal, roadside unit, and base station are converted into graph nodes, and the communication relationships between the vehicle terminal and the roadside unit, base station, and other vehicle terminals are converted into directed edges. Link latency, link packet loss rate, and received signal-to-noise ratio are written into the directed edge attributes, and vehicle location, vehicle speed, and remaining vehicle battery power are written into the vehicle terminal node attributes to generate a dynamic directed graph. Specifically, generating the dynamic directed graph involves: A graph node set is established based on vehicle terminal identifier, roadside unit identifier, and base station identifier. Directed connections between vehicle terminal and roadside unit, vehicle terminal and base station, and vehicle terminal are established based on link object identifier. Link delay, link packet loss rate, and received signal-to-noise ratio are written into the edge attributes of the corresponding directed connection. Vehicle location, vehicle speed, and vehicle remaining battery power are written into the node attributes of the corresponding vehicle terminal. The node attributes and edge attributes are updated according to the collection timestamp to form a dynamic directed graph that changes over time.

[0024] In this embodiment, the adaptive determination of the number of neighboring edges to be retained and the neighborhood pruning threshold, and the performance of neighborhood pruning on the dynamic directed graph to obtain a sparse adjacency matrix, includes: The average vehicle speed is calculated based on the vehicle speed of the vehicle terminal nodes in the dynamic directed graph, and the node density is calculated based on the number of vehicle terminal nodes and the current road coverage length. Specifically, the calculation of the average vehicle speed and node density involves: Read the set of vehicle terminal nodes corresponding to the current collection timestamp, extract the vehicle speed of each vehicle terminal node in the set, sum the vehicle speeds and divide by the number of vehicle terminal nodes to obtain the average vehicle speed; obtain the current road coverage length corresponding to the dynamic directed graph, count the number of vehicle terminal nodes within the current road coverage length, and divide the number of vehicle terminal nodes by the current road coverage length to obtain the node density. The number of neighboring edges to retain is determined based on the average vehicle speed, node density, and a preset upper limit for neighboring edges. Specifically, the number of retained neighboring edges decreases as the average vehicle speed increases and increases as node density increases. Configure a preset neighbor limit based on the edge server processing capacity, the allowable latency for a single network deployment decision, and the maximum number of candidate links allowed to participate in aggregation within the current road coverage length; read historical vehicle speed statistics, arrange the historical vehicle speed statistics results in ascending order, and take the speed value at the one-third mark as the boundary between the low-speed and medium-speed intervals, and the speed value at the two-thirds mark as the boundary between the medium-speed and high-speed intervals; read historical node density statistics, arrange the historical node density statistics results in ascending order, and take the node density value at the one-third mark as the boundary between the low-density and medium-density intervals, and the node density value at the three-thirds mark as the boundary between the low-density and medium-density intervals. The node density value at half the position is used as the boundary between the medium-density and high-density intervals; decreasing speed adjustment coefficients are configured for the low-speed, medium-speed, and high-speed intervals, and increasing density adjustment coefficients are configured for the low-density, medium-density, and high-density intervals, respectively; the minimum number of retained neighboring edges is determined according to the minimum connectivity requirements of the edge server; the preset upper limit of neighboring edges is multiplied by the speed adjustment coefficient to obtain the speed-corrected neighboring edge number; the speed-corrected neighboring edge number is multiplied by the density adjustment coefficient to obtain the candidate retained neighboring edge number; the candidate retained neighboring edge number is rounded down, and the rounding result is limited to between the minimum retained neighboring edge number and the preset upper limit of neighboring edges to obtain the number of retained neighboring edges; The link confidence of directed edges is calculated based on link delay, link packet loss rate, and received signal-to-noise ratio (SNR). The neighborhood pruning threshold is determined based on the distribution of link confidence among adjacent edges connected to the same node. Specifically, the calculation of the link confidence of directed edges based on link delay, link packet loss rate, and received SNR is as follows: Link delay, link packet loss rate, and received signal-to-noise ratio (SNR) under the same acquisition timestamp are normalized to obtain normalized values ​​for delay, packet loss rate, and SNR. The normalized values ​​for delay and packet loss rate are then inversely transformed to obtain reliable components for delay and packet loss. The reliable components for delay, packet loss, and SNR are then weighted and summed in the order of these components to generate a directed link confidence score. The weights are determined by the contribution ratio of the three types of link indicators to the successful handover result in historical successful handover records.

[0025] The neighborhood pruning threshold is determined based on the link confidence distribution connecting the same node to its neighbors, specifically as follows: Obtain the confidence scores of all neighboring links connected to the same vehicle terminal node. Sort all neighboring link confidence scores in ascending order of value. Calculate the difference between adjacent link confidence scores and select the neighboring link confidence score corresponding to the largest difference as the boundary position. When the number of neighboring edges after the boundary position is not less than the number of neighboring edges to be retained, determine the higher link confidence score corresponding to the boundary position as the neighborhood pruning threshold. When the number of neighboring edges after the boundary position is less than the number of neighboring edges to be retained, determine the link confidence score corresponding to the neighboring edge at the 1st position after sorting by link confidence in descending order as the neighborhood pruning threshold. Delete neighboring edges whose link confidence is lower than the neighborhood pruning threshold. Sort the neighboring edges whose link confidence is not lower than the neighborhood pruning threshold in descending order of link confidence. Then, select neighboring edges sequentially from the first position in the sorted list, based on the number of retained neighboring edges, to generate a sparse adjacency matrix. Specifically, the generation of the sparse adjacency matrix is ​​as follows: An initial zero matrix is ​​established based on the number of nodes in the dynamic directed graph. The node identifiers of vehicle terminal nodes, roadside unit nodes, and base station nodes are mapped to matrix row and column indices. The neighboring edges retained after neighborhood pruning are read, and the matrix elements are located according to the starting node index and target node index of the retained neighboring edges. The link confidence corresponding to the retained neighboring edge is written into the corresponding matrix element, and the matrix elements corresponding to the unretained neighboring edges are kept as zero, thus generating a sparse adjacency matrix.

[0026] In this embodiment, the output node stability probability set and the link access priority set include: An improved SE-GCN model is constructed, comprising a node-edge feature encoding layer, a channel attention gating layer, a threshold-shared edge filtering layer, a pruned coupled graph convolutional layer, and a network output layer. Specifically, the improved SE-GCN model is constructed as follows: Based on the graph convolution feature aggregation structure of the traditional SE-GCN model, the original fixed adjacency matrix input structure is replaced with a sparse adjacency matrix input structure obtained by dynamic directed graph pruning, forming a node edge feature encoding layer. Before the original node feature convolution structure, a processing structure is introduced to perform channel compression statistics and generate gating coefficients on the node encoded feature set, forming a channel attention gating layer. Based on the original edge weight participation graph convolution aggregation structure, a processing structure is introduced to generate shared weight parameters by neighborhood pruning threshold and generate edge filtering operators by combining link confidence, forming a threshold-shared edge filtering layer. In the original graph convolution layer, a synchronous input relationship between the gated node feature set and the filtered edge feature set is introduced, so that neighborhood aggregation and neighborhood compression are completed in the same convolution process, forming a pruning coupled graph convolution layer. Based on the original model output structure, the single node representation output is expanded into a dual-result output structure of link access priority set and node stability probability set, forming a network result output layer, resulting in the improved SE-GCN model. The sparse adjacency matrix, vehicle position, vehicle speed, remaining vehicle battery power, link delay, link packet loss rate, and received signal-to-noise ratio are input into the node edge feature encoding layer to generate a node encoded feature set and an edge encoded feature set. The node edge feature encoding layer includes: Node feature normalization unit: Unifies the coordinate scale and normalizes the values ​​of vehicle position, vehicle speed and vehicle remaining battery power to generate basic node features; Edge feature normalization unit: performs directional consistency normalization on link delay, link packet loss rate and received signal-to-noise ratio to generate basic edge features; Adjacency binding unit: Based on the non-zero matrix elements in the sparse adjacency matrix, establish the correspondence between the basic features of nodes and the basic features of edges; Encoding output unit: Performs linear mapping and nonlinear activation on the basic features of nodes and edges that have completed the corresponding relationship binding, and generates a set of node encoded features and a set of edge encoded features; In the node-edge feature encoding layer, the node feature normalization unit reads the vehicle position, vehicle speed, and remaining battery power, performs coordinate scale unification on the vehicle position, and performs numerical normalization on the vehicle speed and remaining battery power to obtain the basic node features; the edge feature normalization unit reads the link delay, link packet loss rate, and received signal-to-noise ratio, performs reverse normalization on the link delay and link packet loss rate, and performs forward normalization on the received signal-to-noise ratio to obtain the basic edge features; the adjacency binding unit reads the non-zero matrix elements in the sparse adjacency matrix and establishes a correspondence between the starting node, target node, and basic edge features corresponding to the non-zero matrix elements; the encoding output unit performs linear mapping and non-linear activation on the basic node features and basic edge features according to the correspondence, and outputs the set of node encoded features for use by the channel attention gating layer and the set of edge encoded features for use by the threshold shared edge filtering layer; The node-encoded feature set is input into the channel attention gating layer, where channel compression statistics are performed to generate channel statistical vectors and gating coefficients. The node-encoded feature set is then recalibrated channel by channel to generate a gated node feature set. The channel attention gating layer includes: Channel compression statistics unit: Aggregates and statistically analyzes the node encoding values ​​of the same feature channel in the node encoding feature set to generate a channel statistics vector; Gating coefficient generation unit: Performs linear mapping and nonlinear activation on the channel statistical vector to generate gating coefficients corresponding to the feature channels; Channel weight binding unit: Establishes a binding relationship between the gating coefficient and the corresponding feature channel in the node's encoded feature set; Node feature recalibration unit: Performs channel-wise weighting on the node encoded feature set according to the binding relationship to generate a gated node feature set; In the channel attention gating layer, the channel compression statistics unit reads the node encoding feature set and performs mean and maximum value statistics on the node encoding values ​​of each vehicle terminal node according to the feature channel dimension to obtain the channel statistics vector; the gating coefficient generation unit performs linear mapping, nonlinear activation and numerical compression on the channel statistics vector to obtain gating coefficients with limited value range; the channel weight binding unit writes the gating coefficients into the corresponding channels of the node encoding feature set according to the feature channel index; the node feature recalibration unit multiplies the node encoding value of each channel in the node encoding feature set with the corresponding gating coefficient to generate a gating node feature set for use by the pruned coupled graph convolutional layer; The edge-encoded feature set, link confidence, and neighborhood pruning threshold are input into the threshold-shared edge filtering layer. Shared weight parameters are generated based on the neighborhood pruning threshold, and an edge filtering operator is generated by combining the link confidence and edge-encoded feature set. This operator performs weighted filtering on the remaining neighboring edges to generate a filtered edge feature set. The threshold-shared edge filtering layer includes: Threshold weight generation unit: Receives the neighborhood pruning threshold, performs scale normalization and weight mapping on the neighborhood pruning threshold, and generates shared weight parameters; Confidence difference calculation unit: Receives link confidence and shared weight parameters, and calculates the confidence difference characteristics of the link confidence relative to the shared weight parameters; Edge filtering operator generation unit: fuses and maps the confidence difference features with the edge encoding feature set to generate an edge filtering operator that retains the corresponding neighboring edges; Filtered edge feature output unit: The edge filtering operator is used to perform weighted filtering on the edge encoded feature set to generate a filtered edge feature set; In the threshold-shared edge filtering layer, the threshold weight generation unit reads the neighborhood pruning threshold, normalizes the neighborhood pruning threshold to a numerical range consistent with the link confidence, and generates shared weight parameters through linear mapping; the confidence difference calculation unit reads the link confidence of each retained neighbor, calculates the difference between the link confidence and the shared weight parameters, and obtains the confidence difference feature; the edge filtering operator generation unit concatenates the confidence difference feature with the edge encoding feature of the corresponding retained neighbor, and performs linear mapping and nonlinear activation to generate the edge filtering operator; the edge filtering feature output unit weights the edge encoding feature of the retained neighbor according to the edge filtering operator, reduces the feature participation intensity of low-confidence neighbors, and generates a set of filtered edge features for use by the pruning coupled graph convolutional layer; The gated node feature set, the filtered edge feature set, and the sparse adjacency matrix are input into the pruned coupled graph convolutional layer, where neighborhood aggregation and neighborhood compression are performed simultaneously to generate the pruned coupled convolutional feature set. The pruned coupled graph convolutional layer includes: Neighborhood index reading unit: Reads non-zero matrix elements in the sparse adjacency matrix to determine the retained neighboring edges and adjacent nodes corresponding to each vehicle terminal node; Edge weight modulation unit: receives the set of filtered edge features and generates aggregate weights corresponding to the retained neighboring edges based on the set of filtered edge features; Node neighborhood aggregation unit: The gated node features of adjacent nodes are weighted and aggregated according to the aggregation weight to generate a neighborhood aggregation feature set; Convolutional compression output unit: The gating node features of the vehicle terminal node itself are fused and convolved with the neighborhood aggregated feature set to generate a pruned coupled convolutional feature set; In the pruned coupled graph convolutional layer, the neighborhood index reading unit reads the non-zero matrix elements in the sparse adjacency matrix, determines the target vehicle terminal node according to the matrix row index, determines the corresponding adjacent nodes according to the matrix column index, and matches the retained neighboring edges corresponding to the non-zero matrix elements with the filtered edge feature set; the edge weight modulation unit performs linear mapping and numerical normalization on the filtered edge feature set to generate the aggregation weights corresponding to the retained neighboring edges; the node neighborhood aggregation unit performs weighted summation on the gated node features of adjacent nodes according to the aggregation weights to obtain the neighborhood aggregation features corresponding to the target vehicle terminal node; the convolutional compression output unit concatenates and fuses the gated node features of the target vehicle terminal node itself with the neighborhood aggregation features, and performs graph convolution mapping and dimensionality compression to generate a pruned coupled convolutional feature set for use by the network result output layer; The pruned coupled convolutional feature set is input into the network result output layer, and link priority mapping and node stability probability mapping are performed to generate link access priority set and node stability probability set. The network result output layer includes: Link Feature Matching Unit: Extracts pairwise features between vehicle terminal nodes and candidate connection objects based on the pruned coupled convolution feature set, and generates candidate link matching features; Link Priority Mapping Unit: Performs linear mapping and normalized sorting on the matching features of candidate links to generate a set of link access priorities; Node stability mapping unit: Performs stability mapping on the pruned coupled convolutional features corresponding to the vehicle terminal nodes to generate a set of node stability probabilities; Result Binding Output Unit: Bind and output the link access priority set and the node stability probability set according to the vehicle terminal identifier and the candidate connection object identifier; In the network result output layer, the link feature matching unit reads the pruned coupled convolution feature set, extracts vehicle terminal node features and candidate connection object node features according to the preserved neighbor relationships in the sparse adjacency matrix, and concatenates the two node features to obtain candidate link matching features; the link priority mapping unit performs linear mapping, nonlinear activation and normalization processing on the candidate link matching features to obtain the link access priority corresponding to each candidate connection object, and forms a link access priority set according to the vehicle terminal identifier; the node stability mapping unit performs linear mapping and probability normalization processing on the pruned coupled convolution features corresponding to the vehicle terminal node to obtain the node stability probability corresponding to the vehicle terminal node, and forms a node stability probability set according to the vehicle terminal identifier; the result binding output unit associates the node stability probability under the same vehicle terminal identifier with the link access priority corresponding to the candidate connection object to generate network result data for hierarchical quantization mapping. Obtain historical link switching results, historical link interruption results, and historical node connection maintenance results; construct a supervised label set; calculate link priority prediction error, node stability probability prediction error, and edge filtering operator output error; combine these to form a joint optimization objective; iteratively update and improve the SE-GCN model parameters according to the joint optimization objective; terminate training when a preset number of iterations is reached or the change in the joint optimization objective is lower than a preset threshold. Specifically, iteratively updating and improving the SE-GCN model parameters according to the joint optimization objective involves: Historical link handover results, historical link interruption results, and historical node connection maintenance results are compiled into a supervision label set. The sparse adjacency matrix, vehicle location, vehicle speed, vehicle remaining battery power, link latency, link packet loss rate, received signal-to-noise ratio, link confidence, and neighborhood pruning threshold are input into the improved SE-GCN model to obtain the link access priority set, node stability probability set, and edge filtering operator output results during the training phase. The link priority prediction error is calculated based on the difference between the link access priority set and the historical link handover results; the node stability probability prediction error is calculated based on the difference between the node stability probability set and the historical node connection maintenance results; and the edge filtering operator output results are calculated based on the difference between the link access priority set and the historical link handover results. The difference in the results of link interruption is used to calculate the output error of the side filtering operator. The link priority prediction error, node stability probability prediction error and side filtering operator output error are weighted and summed according to a weight ratio of 4:4:2 to form a joint optimization objective. The parameter gradients of the node side feature encoding layer, channel attention gating layer, threshold sharing side filtering layer, pruned coupled graph convolutional layer and network result output layer are calculated according to the joint optimization objective, and the parameters of the SE-GCN model are updated and improved according to the parameter gradients. The preset number of iterations is set to 200 rounds and the preset threshold is set to 0.001. The training ends when the training rounds reach 200 rounds or the change of the joint optimization objective is less than 0.001 for 5 consecutive rounds.

[0027] In this embodiment, the step of performing hierarchical quantization mapping on the node stability probability set and the link access priority set to generate access priority lookup tables and transmit power lookup tables includes: Based on the vehicle terminal identifier and the candidate connection object identifier, the node stability probability and link access priority are matched to generate a candidate link score set. Specifically, the matching of node stability probability and link access priority to generate the candidate link score set is as follows: Read the vehicle terminal identifier, node stability probability, and collection timestamp from the node stability probability set; read the vehicle terminal identifier, candidate connection object identifier, link access priority, and collection timestamp from the link access priority set; establish a correspondence between the node stability probability of the same vehicle terminal at the same collection timestamp and the link access priority corresponding to each candidate connection object according to the matching condition of consistent vehicle terminal identifier and consistent collection timestamp; combine the vehicle terminal identifier, candidate connection object identifier, node stability probability, and link access priority into candidate link scoring items; aggregate the candidate link scoring items according to the vehicle terminal identifier, and remove duplicate scoring items according to the candidate connection object identifier to generate a candidate link scoring set corresponding to each vehicle terminal; The candidate link score set is normalized, and layers are divided into low-stability, medium-stability, and high-stability layers based on node stability probability, and into low-priority, medium-priority, and high-priority layers based on link access priority. Specifically: Read the candidate link score set corresponding to the same vehicle terminal, and perform minimum and maximum value statistics on the node stability probability and link access priority in the candidate link score set, respectively. Normalize the scores according to the interval from the minimum to the maximum value to obtain the normalized node stability probability and normalized link access priority. Sort the normalized node stability probabilities in ascending order of value, and take the normalized node stability probability corresponding to the last third of the sorted position as the boundary between the low-stability layer and the medium-stability layer, and take the normalized node stability probability corresponding to the last two-thirds of the sorted position as the boundary between the medium-stability layer and the high-stability layer. Sort the normalized link access priorities in ascending order of value, and take the normalized link access priority corresponding to the last third of the sorted position as the boundary between the low-priority layer and the medium-priority layer. The normalized link access priority corresponding to the latter two-thirds of the position is used as the boundary value between the medium-priority layer and the high-priority layer; candidate link scores not higher than the boundary value between the low-stability layer and the medium-stability layer are assigned to the low-stability layer; candidate link scores higher than the boundary value between the low-stability layer and the medium-stability layer but not higher than the boundary value between the medium-stability layer and the high-stability layer are assigned to the medium-stability layer; candidate link scores higher than the boundary value between the medium-stability layer and the high-stability layer are assigned to the high-stability layer; candidate link scores not higher than the boundary value between the low-priority layer and the medium-priority layer are assigned to the low-priority layer; candidate link scores higher than the boundary value between the low-priority layer and the medium-priority layer but not higher than the boundary value between the medium-priority layer and the high-priority layer are assigned to the medium-priority layer; candidate link scores higher than the boundary value between the medium-priority layer and the high-priority layer are assigned to the high-priority layer. The three types of stabilization layers are cross-mapped with the three types of priority layers to form nine types of quantization mapping units, and access ordering identifiers and transmit power levels are configured. Specifically: Using low-stability, medium-stability, and high-stability layers as stability dimensions, and low-priority, medium-priority, and high-priority layers as priority dimensions, a three-row, three-column quantization mapping table is established. Each stability layer is combined with each priority layer to form nine types of quantization mapping units: low-stability low-priority, low-stability medium-priority, low-stability high-priority, medium-stability low-priority, medium-stability medium-priority, medium-stability high-priority, high-stability low-priority, high-stability medium-priority, and high-stability high-priority. Access sorting identifiers are configured for the nine types of quantization mapping units in descending order of stability level and descending order of priority level, so that high-stability high-priority corresponds to the highest access sorting identifier, and low-stability low-priority corresponds to the lowest access sorting identifier. Transmission power levels are configured based on both stability level and priority level, where high-stability high-priority corresponds to low transmission power level, medium-stability or medium-priority corresponds to medium transmission power level, and low-stability or low-priority corresponds to high transmission power level, thus obtaining the mapping relationship between the nine types of quantization mapping units and access sorting identifiers and transmission power levels. Based on the quantization mapping unit to which the candidate link score set belongs, the access sorting identifier and transmit power level of the candidate connection object corresponding to each vehicle terminal are determined. Specifically, determining the access sorting identifier and transmit power level of the candidate connection object corresponding to each vehicle terminal is as follows: Read the vehicle terminal identifier, candidate connection object identifier, normalized node stability probability, and normalized link access priority from the candidate link score set; determine the quantization mapping unit corresponding to the candidate connection object based on the stability layer to which the normalized node stability probability belongs and the priority layer to which the normalized link access priority belongs; in the mapping relationship between the nine types of quantization mapping units and the access sorting identifier and transmit power level, read the access sorting identifier and transmit power level bound to the corresponding quantization mapping unit; combine the vehicle terminal identifier, candidate connection object identifier, access sorting identifier, and transmit power level into a candidate connection control item to obtain the access sorting identifier and transmit power level of the candidate connection object corresponding to each vehicle terminal; Candidate connection objects for the same vehicle terminal are sorted according to the access sorting identifier, and an access priority lookup table is generated. The specific steps for generating the access priority lookup table are as follows: Read the candidate connection control items corresponding to the same vehicle terminal, and extract the candidate connection object identifier and access sorting identifier; arrange the candidate connection object identifiers in descending order of access sorting identifier, and arrange them in descending order of link access priority when the access sorting identifiers are the same; write the sorting number, vehicle terminal identifier, candidate connection object identifier and access sorting identifier into the table entries, collect the table entries according to the vehicle terminal identifier, and generate the access priority lookup table corresponding to each vehicle terminal; Based on the correspondence between vehicle terminal identifier, candidate connection object identifier, and transmission power level, a transmission power lookup table is generated. Specifically, the generation of the transmission power lookup table is as follows: Read the candidate connection control items corresponding to each vehicle terminal, and extract the vehicle terminal identifier, candidate connection object identifier, and transmission power level; match the corresponding transmission power level according to the transmission power level, and write the vehicle terminal identifier, candidate connection object identifier, transmission power level, and transmission power level into the power control table item; collect the power control table items according to the vehicle terminal identifier, and generate the transmission power lookup table corresponding to each vehicle terminal; Using the vehicle terminal identifier and the candidate connection object identifier as dual indexes, the access priority lookup table and the transmit power lookup table are encapsulated into a network control table and distributed to the vehicle terminal. Specifically, the encapsulation of the access priority lookup table and the transmit power lookup table into a network control table is as follows: Read the vehicle terminal identifier, candidate connection object identifier, sorting number, and access sorting identifier from the access priority lookup table; read the vehicle terminal identifier, candidate connection object identifier, transmission power level, and transmission power level from the transmission power lookup table; establish the correspondence between the access priority lookup table and the transmission power lookup table based on the same vehicle terminal identifier and the same candidate connection object identifier; set the vehicle terminal identifier and candidate connection object identifier as dual-index fields, and set the sorting number, access sorting identifier, transmission power level, and transmission power level as control fields to generate network control table entries; collect network control table entries according to vehicle terminal identifiers, add version number, generation timestamp, and integrity verification fields, and encapsulate them to form a network control table readable by the vehicle terminal.

[0028] In this embodiment, the vehicle terminal switches links according to access priority and adjusts the transmission power according to the transmission power table, including: The vehicle terminal uses access priority lookup tables and transmit power lookup tables to read the dual index entries based on the vehicle terminal identifier and the candidate connection object identifier, obtaining the target access object identifier and the target transmit power level. Specifically, the target access object identifier and transmit power level are obtained as follows: The vehicle terminal reads the vehicle terminal identifier field from the network control table, matches its own vehicle terminal identifier with the vehicle terminal identifier in the network control table, and obtains the set of dual-index entries corresponding to the vehicle; it reads the candidate connection object record in the dual-index entry set according to the candidate connection object identifier, and selects the candidate connection object identifier with the smallest sorting number according to the access priority lookup table, thus obtaining the target access object identifier; it reads the power control table entry with the same vehicle terminal identifier and the same candidate connection object identifier as the target access object identifier from the transmit power lookup table, thus obtaining the target transmit power level; The vehicle terminal performs link switching based on the target access object identifier, adjusts the wireless transmission power according to the target transmission power level, and establishes a communication connection between the vehicle terminal and the target access object. Specifically, establishing this communication connection involves: The vehicle terminal reads the target access object identifier and the target transmission power level, and configures the wireless transmission power according to the target transmission power level; it sends an access request message to the target access object and receives an access confirmation message returned by the target access object; after the access confirmation message is verified, it writes the vehicle terminal identifier, the target access object identifier, the access timestamp, the link status flag, and the transmission power level to form a communication connection record; it updates the current service object of the vehicle terminal according to the communication connection record and establishes a communication connection relationship between the vehicle terminal and the target access object. The vehicle terminal updates its local adjacency table and node connection status table based on the target access object identifier, and completes the topology connection update by looking up the access priority and transmit power tables, forming the network topology structure after inference-free reconstruction. Specifically, the updates to the local adjacency table and node connection status table based on the target access object identifier, and the topology connection update by looking up the access priority and transmit power tables, are as follows: The vehicle terminal reads the current local adjacency table, writes the target access object identifier into the current service object field, and updates the original service object field to a candidate connection object or a disconnected object; it reads the node connection status table, marks the connection status between the vehicle terminal and the target access object as connected, and marks the connection status between the vehicle terminal and the original service object as switched; it updates the access order of candidate connection objects according to access priority, and updates the transmission power level and transmission power grade corresponding to the target access object according to transmission power; it writes the updated local adjacency table and node connection status table into the vehicle terminal's local cache, forming the network topology structure after inference-free reconstruction.

[0029] In this embodiment, the edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, and updates the neighborhood pruning threshold and the weights of the improved SE-GCN model, including: The edge server receives link latency, link packet loss rate, received signal-to-noise ratio, and link switching results from vehicle terminals. It then generates a set of link feedback records by associating and matching these records with the vehicle terminal identifier and timestamp. Specifically, the generation of this link feedback record set by associating and matching with the vehicle terminal identifier and timestamp is as follows: The edge server reads the vehicle terminal identifier, feedback timestamp, link latency, link packet loss rate, received signal-to-noise ratio, and link switching result from the vehicle terminal feedback messages; it aggregates feedback messages from the same vehicle terminal according to the vehicle terminal identifier, matches them according to the time difference between the feedback timestamp and the timestamp generated by the network control table, and determines feedback messages with a time difference less than the matching time window as valid feedback messages; it writes the link latency, link packet loss rate, received signal-to-noise ratio, and link switching result from the valid feedback messages into the same link feedback record; and it arranges the link feedback records according to the vehicle terminal identifier and feedback timestamp to generate a set of link feedback records. Based on the link feedback record set, statistical analysis is performed on the trends of link delay, link packet loss rate, and received signal-to-noise ratio. The link confidence score is then recalculated, and the link confidence score distribution and neighborhood pruning threshold are updated. Specifically, the recalculation of link confidence score and the updating of the link confidence score distribution and neighborhood pruning threshold are as follows: The edge server reads the link latency, link packet loss rate, and received signal-to-noise ratio (SNR) trends from the link feedback record set. It performs inverse normalization on the link latency and link packet loss rate, and forward normalization on the received SNR, obtaining updated reliability components for latency, packet loss, and SNR. Based on the contribution ratio of the three link metrics to the successful handover results in historical successful handover records, the reliability components for latency, packet loss, and SNR are weighted and fused, and the link confidence is recalculated. The link confidence of adjacent links connected to the same vehicle terminal is aggregated according to the vehicle terminal identifier and arranged in ascending order of value, forming an updated link confidence distribution. The difference between adjacent link confidences is calculated, and the adjacent link confidence corresponding to the largest difference is used as the candidate boundary position. The updated neighborhood pruning threshold is determined by combining this with the number of retained adjacent edges. The link feedback record set is input into the improved SE-GCN model. The model weights are updated based on the deviations between the predicted link access priority and the actual link handover results, and between the predicted node stability probability and the actual connection maintenance results, generating an updated weight set. Specifically, the updated weight set is generated by updating the model weights based on the deviations between the predicted link access priority and the actual link handover results, and between the predicted node stability probability and the actual connection maintenance results. The edge server reads the actual link switching results and actual connection hold results from the link feedback record set. Successfully switched candidate connections are marked as positive samples, and unselected or failed candidate connections are marked as negative samples, forming link access priority supervision labels. Vehicle terminals whose connection hold time reaches the stable time threshold are marked as stable samples, and vehicle terminals with interrupted connections or hold times below the stable time threshold are marked as unstable samples, forming node stability probability supervision labels. The link access priority prediction results are compared item by item with the link access priority supervision labels, and the difference between underestimation of positive sample predictions and overestimation of negative sample predictions is statistically analyzed to obtain the link priority feedback error. The node stability probability prediction results are compared item by item with the node stability probability supervision labels, and the difference between underestimation of positive sample predictions and overestimation of negative sample predictions is statistically analyzed to obtain the link priority feedback error. The difference between underestimation of stable samples and overestimation of unstable samples is used to obtain the node stability probability feedback error. The link priority feedback error and the node stability probability feedback error are weighted and summed according to the first error weight and the second error weight to obtain the feedback update target. The first error weight and the second error weight are determined based on the historical link handover success rate and the historical connection retention success rate. Based on the feedback update target, the influence direction and magnitude of the parameters of each layer in the improved SE-GCN model on the error change are calculated. The error is reduced according to the influence direction and the update step size is determined according to the influence magnitude. The model weights of the node edge feature encoding layer, the channel attention gating layer, the threshold shared edge filtering layer, the pruned coupled graph convolutional layer, and the network result output layer are updated to generate an updated weight set. The system monitors the service load status of the edge server. When the service load falls below a preset threshold, it updates the weight set by model level, parameter type, and timestamp, and uploads it to the cloud platform storage. Specifically, this process involves: The edge server reads the current processor utilization, memory utilization, number of queued tasks, and number of network control requests to calculate the business load status. Based on the historical load records of the edge server during normal network periods, it configures processor utilization below 40%, memory utilization below 50%, and the number of queued tasks below 20 as the criteria for determining that the business load is below a preset threshold. When the criteria are met, it reads and updates the model level identifier, parameter type identifier, weight value, and weight update timestamp from the weight set, and applies these parameters according to the node edge feature encoding layer, channel attention gating layer, threshold sharing edge filtering layer, and so on. The weights of the pruned coupled graph convolutional layer and the network output layer are grouped according to their hierarchical order. The weights belonging to the convolution parameters, gating parameters, edge filtering parameters, mapping parameters, and bias parameters within the same model level are written into the corresponding parameter type fields. The weight update timestamp, model version number, vehicle segment identifier, and integrity verification field are written into the encapsulation header to generate a weight update data packet. The edge server sends the weight update data packet through an encrypted communication channel with the cloud platform. The cloud platform establishes a storage index according to the model level, parameter type, and weight update timestamp, and writes the weight update data packet into the model weight storage area.

[0030] refer to Figure 3 A vehicle-to-everything (V2X) networking system based on machine learning includes the following modules: The status data acquisition module is used to acquire vehicle operating status and communication link status, and generate vehicle network status data; The dynamic graph construction module is used to build a dynamic directed graph between vehicle terminals, roadside units, and base stations based on vehicle network status data; The neighborhood pruning module is used to determine the number of neighboring edges to retain and the neighborhood pruning threshold based on the dynamic directed graph, and to generate a sparse adjacency matrix. An improved model inference module is used to input the sparse adjacency matrix into the improved SE-GCN model to generate a set of node stability probabilities and a set of link access priorities. The lookup mapping generation module is used to perform hierarchical quantization mapping on the node stability probability set and the link access priority set, and generate access priority lookup table and transmit power lookup table; The inference-free reconfiguration module is used to complete link switching, transmit power adjustment and topology reconfiguration by looking up tables based on access priority and transmit power. The feedback update module is used to update the neighborhood pruning threshold and improve the SE-GCN model weights based on the link status feedback, and upload the updated weights to the cloud platform for storage.

[0031] Example 1: In a continuous networking test cycle of a vehicle-to-everything (V2X) system on an urban expressway, the test section was 8.4 km long, with six lanes in both directions. Twelve roadside units and three base stations were deployed along the route. A total of 96 vehicles participated in the test, including 72 passenger cars, 10 buses, and 14 logistics vehicles. The test period covered morning and off-peak hours, with vehicle speeds ranging from 32 km / h to 96 km / h and remaining battery power ranging from 41% to 92%. The highest local vehicle density reached 39 vehicles / km. Vehicle terminals read data from the onboard positioning module, speed sensor module, and battery management module at a 200 ms sampling period, and simultaneously sent detection messages to the roadside units and base stations. Within a certain sampling window, vehicle V023 was located at a road mileage of 3.26 km, a speed of 86 km / h, and had 63% remaining battery power. It sent 20 probe messages to roadside unit R05, acknowledging reception of 19, with an average round-trip time of 31 ms, a packet loss rate of 5%, a received signal strength of -68 dBm, and a background noise level of -94 dBm, resulting in a received signal-to-noise ratio (SNR) of 26 dB. The same vehicle sent 20 probe messages to roadside unit R06, acknowledging reception of 17, with an average round-trip time of 76 ms, a packet loss rate of 15%, and a received SNR of 15 dB. The vehicle terminal encapsulated the vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received SNR into a network status data packet, categorized by vehicle terminal identifier, link object identifier, and collection timestamp. Each data packet was 188 bytes long. The edge server received 478 valid network status data packets within 1 second, of which 7 were invalid or failed to be validated, resulting in an effectiveness rate of 98.6%. The edge server aligns the data according to vehicle terminal identifiers, roadside unit identifiers, base station identifiers, and data collection timestamps, converting 96 vehicle terminals, 12 roadside units, and 3 base stations into graph nodes. It then converts the communication relationships between vehicle terminals and roadside units, base stations, and adjacent vehicle terminals into directed edges, initially forming 1432 directed edges. The edge server writes link latency, link packet loss rate, and received signal-to-noise ratio into the directed edge attributes, and vehicle location, vehicle speed, and remaining battery power into the vehicle terminal node attributes, forming a dynamic directed graph for the current data collection window.

[0032] When the dynamic directed graph enters the neighborhood pruning stage, the edge server counts 96 current vehicle terminal nodes, with a road coverage length of 8.4km, resulting in a node density of 11.43 vehicles / km. During the morning rush hour congestion, 43 vehicle terminals exist within a 3.1km to 4.2km section, with a local node density of 39.09 vehicles / km. The system reads the vehicle speeds of the current vehicle terminal nodes, calculating a global average vehicle speed of 68.5km / h and an average speed of 88.7km / h for the local high-speed section. Based on the edge server's processing capacity, the allowable latency for a single network decision, and the maximum number of candidate links allowed to participate in aggregation within the current road coverage length, the preset upper limit for neighboring edges is configured to 10, and the minimum number of retained neighboring edges is configured to 3. The system reads historical vehicle speed statistics, using the speed value of 46 km / h at one-third of the location as the boundary between the low-speed and medium-speed ranges, and the speed value of 78 km / h at two-thirds of the location as the boundary between the medium-speed and high-speed ranges. It also reads historical node density statistics, using the density value of 16 vehicles / km at one-third of the location as the boundary between the low-density and medium-density ranges, and the density value of 32 vehicles / km at two-thirds of the location as the boundary between the medium-density and high-density ranges. For high-speed, low-density segments, the speed adjustment coefficient is 0.55, the density adjustment coefficient is 0.90, and the number of candidate neighboring edges is 10 × 0.55 × 0.90, rounded to 5. For medium-speed, high-density segments, the speed adjustment coefficient is 0.75, the density adjustment coefficient is 1.15, rounded to 8. When calculating link confidence for directed edges, the system reverse-normalizes link delay and link packet loss rate, forward-normalizes received signal-to-noise ratio, and performs weighted fusion based on the contribution percentage determined by historical successful link handover records. Among the candidate links for vehicle V023, the link confidence score for R05 is 0.86, for base station B02 it is 0.79, for adjacent vehicle V031 it is 0.66, and for R06 it is 0.43. The system sorts the link confidence scores of the same vehicle terminal's adjacent links in ascending order and calculates the difference in confidence scores between adjacent links. The largest difference occurs between 0.43 and 0.66, so 0.66 is determined as the candidate pruning threshold. After neighborhood pruning, the initial 1432 directed edges are reduced to 512, the sparse adjacency matrix size is 111×111, and the proportion of non-zero elements is 4.15%, reducing the matrix computation by approximately 64.2% compared to before pruning.

[0033] In the inference stage of the improved SE-GCN model, the sparse adjacency matrix, vehicle position, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio (SNR) are input to the node edge feature encoding layer. The node feature normalization unit normalizes the vehicle position to the 0-1 interval and normalizes the speed and remaining battery power to generate 32-dimensional basic node features; the edge feature normalization unit normalizes the link latency, packet loss rate, and received SNR to generate 24-dimensional basic edge features; the encoding output unit outputs a 128-dimensional node encoded feature set and a 64-dimensional edge encoded feature set after linear mapping and nonlinear activation. After the node encoded feature set enters the channel attention gating layer, the system performs mean and maximum value statistics for each feature channel to obtain a channel statistical vector. In high-speed vehicle nodes, the average gating coefficient for the speed change channel is 0.74, the average gating coefficient for the remaining battery power channel is 0.48, and the average gating coefficient for the link fluctuation-related channel is 0.81, indicating that the model pays more attention to mobility and link stability in high-speed handover scenarios. After the edge encoding feature set, link confidence, and neighborhood pruning threshold are fed into the threshold-shared edge filtering layer, the system generates shared weight parameters based on the neighborhood pruning threshold. The output values ​​of the candidate edge filtering operator near the threshold are concentrated between 0.18 and 0.34, while the output values ​​of the high-confidence link filtering operator are concentrated between 0.71 and 0.92. The pruning coupled graph convolutional layer takes the gated node feature set, the filtered edge feature set, and the sparse adjacency matrix as input to complete neighborhood aggregation and neighborhood compression, outputting a 128-dimensional pruning coupled convolutional feature set. The network result output layer shows a node stability probability of 0.91 for vehicle V023, a link access priority of 0.88 for R05, a link access priority of 0.76 for B02, and a link access priority of 0.63 for V031. The system normalizes node stability probability and link access priority, then performs hierarchical quantization. V023 and R05 are assigned to the high-stability, high-priority quantization mapping unit, configured with access sorting identifier 1 and low transmit power level; V023 and B02 are assigned to the high-stability, medium-priority quantization mapping unit, configured with access sorting identifier 2 and medium transmit power level; V023 and V031 are assigned to the medium-stability, medium-priority quantization mapping unit, configured with access sorting identifier 3 and medium transmit power level. The lookup mapping generation module uses the vehicle terminal identifier and candidate connection object identifier as dual indexes to generate access priority lookup tables and transmit power lookup tables. After reading the dual index entries, the vehicle terminal determines the target access object as R05, adjusts the transmit power from 23dBm to 19dBm, completes link switching and inference-free topology reconstruction, and the entire vehicle-side lookup and execution process takes 2.4ms.

[0034] In a continuous 3-hour test, the method of this invention was compared with a fixed received signal strength threshold networking method and a conventional GCN networking method. The average link reconnection latency was 94.6 ms for the fixed threshold method, 57.8 ms for the conventional GCN method, and 29.4 ms for the method of this invention. The number of link interruptions per thousand vehicle-kilometers was 41.2 for the fixed threshold method, 25.7 for the conventional GCN method, and 11.8 for the method of this invention. The number of repeated handovers per vehicle per minute was 1.26 for the fixed threshold method, 0.73 for the conventional GCN method, and 0.34 for the method of this invention. Regarding link quality, the average packet loss rate was 4.9% for the fixed threshold method, 2.8% for the conventional GCN method, and 1.5% for the method of this invention. The average received signal-to-noise ratio was 18.2 dB for the fixed threshold method, 21.7 dB for the conventional GCN method, and 24.9 dB for the method of this invention. Regarding vehicle-side load, the typical GCN method takes approximately 16.2ms for a single network decision. This invention reduces this time to 2.4ms through table lookup execution. The average power consumption related to vehicle terminal networking is reduced from 1.82W to 1.51W, a reduction of approximately 17.0%. In high-speed segments, the link access success rate for vehicles exceeding 85km / h is increased from 77.6% with a fixed threshold to 93.1%. In node-dense segments, the average connection hold time is increased from 13.2s to 23.5s. After testing, the edge server updates the link confidence distribution and neighborhood pruning threshold based on vehicle terminal feedback records. When processor utilization is below 40%, memory utilization is below 50%, and the number of queued tasks is below 20, the updated weight set is encapsulated by model level, parameter type, and timestamp and uploaded to the cloud platform for storage. As can be seen from Example 1, the present invention demonstrates a clear data change process in data acquisition, dynamic graph modeling, neighborhood pruning, improved SE-GCN inference, lookup-based topology reconstruction, and feedback updates. This effectively addresses the issues of frequent link interruptions, high handover latency, and heavy inference burden on the vehicle side in high-speed mobile scenarios, verifying the engineering feasibility of this method in vehicle-to-everything (V2X) system networking scenarios. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A networking method for a vehicle-to-everything (V2X) system based on machine learning, characterized in that, include: The system periodically collects vehicle location, vehicle speed, remaining battery power, link latency, link packet loss rate, and received signal-to-noise ratio, and uploads them to the edge server. In the edge server, vehicle terminals, roadside units, and base stations are defined as nodes, and link latency, link packet loss rate, and received signal-to-noise ratio are defined as edge attributes to construct a dynamic directed graph. Calculate the average vehicle speed and node density, adaptively determine the number of neighboring edges to retain and the neighborhood pruning threshold, perform neighborhood pruning on the dynamic directed graph, and obtain a sparse adjacency matrix. The sparse adjacency matrix is ​​input into the improved SE-GCN model. Before each convolution, channel attention gating is used to recalibrate the node features. During convolution, the edge filtering operator that shares weights with the neighborhood pruning threshold is called to simultaneously complete feature aggregation and neighborhood compression, and output the set of node stability probability and the set of link access priority. A hierarchical quantitative mapping is performed on the node stability probability set and the link access priority set to generate access priority lookup table and transmit power lookup table, which are then encapsulated in a dual-index format and sent to the vehicle terminal. The vehicle terminal switches links according to access priority and adjusts the transmission power according to the transmission power table to complete the inference-free topology reconstruction. The edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, updates the neighborhood pruning threshold and improves the SE-GCN model weights, and uploads the updated weights to the cloud platform storage during periods of low business load.

2. The vehicle networking system networking method based on machine learning according to claim 1, characterized in that, The periodic collection of vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio includes: The vehicle terminal reads the vehicle positioning module, vehicle speed sensing module and battery management module according to a preset sampling period to obtain the vehicle position, vehicle speed and vehicle remaining power. The vehicle terminal performs link detection with the roadside unit and base station, and records the detection message sending time, acknowledgment reception time, number of lost messages, number of sent messages and received signal strength to obtain link delay, link packet loss rate and received signal-to-noise ratio; The vehicle terminal encapsulates vehicle location, vehicle speed, remaining vehicle battery power, link latency, link packet loss rate, and received signal-to-noise ratio according to vehicle terminal identifier, link object identifier, and collection timestamp, and uploads them to the edge server.

3. The vehicle networking system networking method based on machine learning according to claim 1, characterized in that, The construction of the dynamic directed graph includes: The edge server receives vehicle location, vehicle speed, vehicle remaining battery power, link latency, link packet loss rate and received signal-to-noise ratio, and aligns them according to vehicle terminal identifier, roadside unit identifier, base station identifier and collection timestamp; The vehicle terminal, roadside unit, and base station are converted into graph nodes, and the communication relationships between the vehicle terminal and the roadside unit, base station, and other vehicle terminals are converted into directed edges. Link latency, link packet loss rate, and received signal-to-noise ratio are written into the directed edge attributes, and vehicle location, vehicle speed, and vehicle remaining battery power are written into the vehicle terminal node attributes to generate a dynamic directed graph.

4. The vehicle networking system networking method based on machine learning according to claim 1, characterized in that, The adaptive determination of the number of neighboring edges to retain and the neighborhood pruning threshold, followed by neighborhood pruning of the dynamic directed graph to obtain a sparse adjacency matrix, includes: The average vehicle speed is calculated based on the vehicle speed of the vehicle terminal nodes in the dynamic directed graph, and the node density is calculated based on the number of vehicle terminal nodes and the current road coverage length. The number of neighboring edges to be retained is determined based on the average vehicle speed, node density, and a preset upper limit for neighboring edges. Specifically, the number of neighboring edges to be retained decreases when the average vehicle speed increases, and the number of neighboring edges to be retained increases when the node density increases. The link confidence of directed edges is calculated based on link delay, link packet loss rate and received signal-to-noise ratio, and the neighborhood pruning threshold is determined based on the link confidence distribution of the same node connecting neighboring edges. Delete neighboring edges whose link confidence is lower than the neighborhood pruning threshold. Sort the neighboring edges whose link confidence is not lower than the neighborhood pruning threshold in descending order of link confidence. Then, select neighboring edges sequentially from the first position of the sorted list according to the number of neighboring edges to be retained, and generate a sparse adjacency matrix.

5. The vehicle networking system networking method based on machine learning according to claim 1, characterized in that, The output node stability probability set and the link access priority set include: An improved SE-GCN model is constructed, which includes a node-edge feature encoding layer, a channel attention gating layer, a threshold-shared edge filtering layer, a pruned coupled graph convolutional layer, and a network result output layer. The sparse adjacency matrix, vehicle position, vehicle speed, vehicle remaining battery power, link delay, link packet loss rate, and received signal-to-noise ratio are input into the node edge feature encoding layer to generate the node encoding feature set and the edge encoding feature set. The node encoded feature set is input into the channel attention gating layer, channel compression statistics are performed to generate channel statistical vectors and gating coefficients, and the node encoded feature set is recalibrated channel by channel to generate the gated node feature set. The edge encoding feature set, link confidence, and neighborhood pruning threshold are input into the threshold shared edge filtering layer. Shared weight parameters are generated based on the neighborhood pruning threshold, and edge filtering operators are generated by combining the link confidence and edge encoding feature set. The retained neighboring edges are weighted and filtered to generate the filtered edge feature set. The gating node feature set, the filtered edge feature set, and the sparse adjacency matrix are input into the pruning coupled graph convolutional layer, and neighborhood aggregation and neighborhood compression are performed simultaneously to generate the pruning coupled convolutional feature set. The pruned coupled convolution feature set is input into the network result output layer, and link priority mapping and node stability probability mapping are performed to generate link access priority set and node stability probability set. Obtain historical link switching results, historical link interruption results, and historical node connection maintenance results; construct a set of supervised labels; calculate link priority prediction error, node stability probability prediction error, and edge filtering operator output error; combine them to form a joint optimization objective; iteratively update and improve the SE-GCN model parameters according to the joint optimization objective; and end training when the preset number of iterations is reached or the change in the joint optimization objective is lower than a preset threshold.

6. The vehicle-to-everything (V2X) system networking method based on machine learning according to claim 1, characterized in that, The step of performing hierarchical quantization mapping on the node stability probability set and the link access priority set to generate access priority lookup tables and transmit power lookup tables includes: Based on the vehicle terminal identifier and the candidate connection object identifier, the node stability probability and link access priority are matched accordingly to generate a candidate link score set; The candidate link score set is normalized, and the nodes are divided into low-stability, medium-stability and high-stability layers according to their stability probability. The links are also divided into low-priority, medium-priority and high-priority layers according to their access priority. The three types of stabilization layers are cross-mapped with the three types of priority layers to form nine types of quantization mapping units, and access ordering identifiers and transmit power levels are configured. Based on the quantization mapping unit to which the candidate link score set belongs, determine the access sorting identifier and transmit power level of the candidate connection object corresponding to each vehicle terminal; Candidate connection objects for the same vehicle terminal are sorted according to the access sorting identifier, and an access priority lookup table is generated. A lookup table for transmission power is generated based on the correspondence between vehicle terminal identifier, candidate connection object identifier, and transmission power level. Using vehicle terminal identifier and candidate connection object identifier as dual indexes, the access priority lookup table and transmit power lookup table are encapsulated into a network control table and sent to the vehicle terminal.

7. The vehicle-to-everything (V2X) system networking method based on machine learning according to claim 1, characterized in that, The vehicle terminal switches links according to access priority and adjusts the transmission power according to the transmission power table, including: The vehicle terminal receives access priority lookup table and transmit power lookup table, and reads the double index item according to vehicle terminal identifier and candidate connection object identifier to obtain target access object identifier and target transmit power level; The vehicle terminal performs link switching based on the target access object identifier, adjusts the wireless transmission power according to the target transmission power level, and establishes a communication connection between the vehicle terminal and the target access object. The vehicle terminal updates its local adjacency table and node connection status table based on the target access object identifier, and completes the topology connection update by looking up the table based on access priority and transmit power, thus forming the network topology structure after inference-free reconstruction.

8. The vehicle networking system networking method based on machine learning according to claim 1, characterized in that, The edge server receives the latest feedback on link latency, link packet loss rate, and received signal-to-noise ratio, and updates the neighborhood pruning threshold and the weights of the improved SE-GCN model, including: The edge server receives link latency, link packet loss rate, received signal-to-noise ratio and link switching results from the vehicle terminal, and generates a set of link feedback records by associating and matching them according to the vehicle terminal identifier and timestamp. Based on the link feedback record set, statistical analysis is performed on the trends of link delay, link packet loss rate, and received signal-to-noise ratio. The link confidence is then recalculated, and the link confidence distribution and neighborhood pruning threshold are updated. The link feedback record set is input into the improved SE-GCN model. The model weights are updated based on the deviation between the predicted link access priority and the actual link switching results, and the deviation between the predicted node stability probability and the actual connection maintenance results, thus generating an updated weight set. The system detects the business load status of the edge server. When the business load is lower than a preset threshold, it updates the weight set by model level, parameter type and timestamp, and uploads it to the cloud platform for storage.

9. A vehicle-to-everything (V2X) system networking system based on machine learning, comprising the V2X system networking method based on machine learning as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The status data acquisition module is used to acquire vehicle operating status and communication link status, and generate vehicle network status data; The dynamic graph construction module is used to build a dynamic directed graph between vehicle terminals, roadside units, and base stations based on vehicle network status data; The neighborhood pruning module is used to determine the number of neighboring edges to retain and the neighborhood pruning threshold based on the dynamic directed graph, and to generate a sparse adjacency matrix. An improved model inference module is used to input the sparse adjacency matrix into the improved SE-GCN model to generate a set of node stability probabilities and a set of link access priorities. The lookup mapping generation module is used to perform hierarchical quantization mapping on the node stability probability set and the link access priority set, and generate access priority lookup table and transmit power lookup table; The inference-free reconfiguration module is used to complete link switching, transmit power adjustment and topology reconfiguration by looking up tables based on access priority and transmit power. The feedback update module is used to update the neighborhood pruning threshold and improve the SE-GCN model weights based on the link status feedback, and upload the updated weights to the cloud platform for storage.

Images