Air-to-ground network traffic offloading decision method and system based on graph neural network

Abstract

The present invention discloses an air-space-ground-ground network traffic unloading decision method and system based on a graph neural network, which relates to the technical field of communication network optimization, and includes: receiving node-related data in an air-space-ground integrated network, dividing the node-related data into time windows, and performing feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes; inputting the node feature data into a pre-established graph neural network model GNN, encoding and aggregating the features of the nodes and edges, generating a representation vector for each node, training the GNN using the representation vector of each node, and obtaining a trained GNN; receiving the network status and node features in the air-space-ground integrated network, inputting them into the trained GNN, and outputting a performance prediction result, based on the performance prediction result, combined with multi-objective optimization technology, thereby achieving optimal traffic unloading of the entire network.

Description

Technical Field

[0001] The present invention relates to the technical field of communication network optimization, and in particular to an air-to-ground network traffic unloading decision method and system based on a graph neural network. Background Art

[0002] With the rapid development of the Space-Ground Integrated Network (SAGIN), traditional network management and data traffic distribution strategies face many challenges. SAGIN is a highly complex network that includes a multi-layer network structure on the ground, in the air, and in space, involving numerous nodes and a wide range of connection types. This network is characterized by complex topology, fast dynamic changes, and a volatile communication environment. Therefore, optimizing the traffic offloading solution of this network becomes the key to improving network performance and efficiency through network management.

[0003] Traditional network traffic management methods usually rely on static routing and bandwidth allocation strategies, which are inefficient in the dynamically changing SAGIN environment. In addition, these methods often fail to reflect changes in network status in real time, and it is difficult to predict and adapt to real-time changes in network load, latency, and energy consumption. This limits the performance of the network and may lead to uneven resource utilization, increased latency, and energy waste.

[0004] In order to solve these problems, in recent years, the research on network management and optimization using machine learning technology has gradually increased. Graph neural networks have been widely studied due to their advantages in processing complex network data. They can effectively process graph structured data and learn the deep characteristics of the network by capturing the dependencies between nodes, providing a new technical approach for network performance prediction and traffic offloading management. At the same time, in the final traffic offloading task, how to establish a multi-objective optimization model and select the best offloading path is also a major difficulty in production practice. Summary of the invention

[0005] In order to solve the deficiencies mentioned in the above-mentioned background technology, the purpose of the present invention is to provide an air-to-ground network traffic unloading decision-making method and system based on graph neural network, which can accurately predict network performance indicators by utilizing the powerful learning ability of graph neural network, and dynamically calculate and implement the optimal traffic unloading strategy by combining multi-objective optimization technology.

[0006] In the first aspect, the purpose of the present invention can be achieved by the following technical solution: a method for air-ground network traffic offloading decision-making based on a graph neural network, the method comprising the following steps:

[0007] Receive node-related data in the air-ground integrated network, divide the node-related data into time windows, and perform feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes;

[0008] The node feature data is input into the pre-established graph neural network model GNN, and the representation vector of each node is generated by encoding and aggregating the features of the nodes and edges. The GNN is trained using the representation vector of each node to obtain the trained GNN.

[0009] Receive the network status and node characteristics in the air-space-ground integrated network, input the network status and node characteristics in the air-space-ground integrated network into the trained GNN, output the performance prediction result, and combine the multi-objective optimization technology based on the performance prediction result to achieve the optimal offloading of the overall network traffic.

[0010] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: acquiring node-related data in the air-space-ground integrated network as follows: automatically collecting key data through sensors deployed on satellites Sat, unmanned aerial vehicles UAVs, and ground base stations BS; transmitting to a data processing center in real time; collecting key data in the air-space-ground integrated network, and collecting real-time location information of each node for the dynamic topological structure in the air-space-ground integrated network SAGIN. , the communication status between nodes , as well as the transmission capacity and communication quality of the nodes ;

[0011] The data vector will also include the transmit power ， Signal loss , and the distance between nodes .

[0012] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: node-related data in the air-ground integrated network is standardized or normalized, and the standardized processing takes into account dynamic changes in the network:

[0013]

[0014] in, is the original key data value of the mth indicator between nodes i and j at time t, Yes The normalized value; considering the interaction between nodes i and j, and They represent the mean and standard deviation of the index m among all nodes respectively;

[0015] Time window division process: According to the communication characteristics of the integrated air-space-ground network and the data collection frequency, the time series data is divided into several time windows. Each time window includes data within a fixed time length, reflecting the changes in network topology caused by the movement of nodes in space; among them, the time window length setting is adjusted according to the network dynamics and the communication characteristics of the nodes.

[0016] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established graph neural network model GNN includes a node and edge , where each node represents a network device and the edge represents the communication connection between nodes;

[0017] Each node has a feature vector express;

[0018] The graph convolution layer is used in the model construction to process the node features and the adjacency matrix A to learn the implicit representation of the node. The graph convolution layer updates the node status by utilizing the topological structure between nodes. The mathematical form of graph convolution is:

[0019]

[0020] in, is the node feature matrix of the lth layer, is the node feature matrix of the next layer l+1, is the adjacency matrix A plus the identity matrix , yes The degree matrix of is the weight of the lth layer, and its dimension is determined by the size of the input and output features. is the nonlinear activation function ReLU.

[0021] In combination with the first aspect, in some implementations of the first aspect, the method further includes: in the process of training the GNN using the representation vector of each node, using the mean square error MSE to construct a loss function Loss:

[0022]

[0023] in, is the predicted value of the model, y i is the true value, N is the number of samples;

[0024] Introducing L2 regularization technology and adding a new loss function Loss after the penalty term for weights * :

[0025]

[0026] in, is the regularization parameter, is the weight of the lth layer.

[0027] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: inputting the network status and node features in the air-ground integrated network into the trained GNN, and outputting a performance prediction result including a plurality of key performance indicators, wherein the plurality of key performance indicators are as follows:

[0028] Latency: The latency of each node in the network is calculated as the sum of transmission and processing delays:

[0029]

[0030] in, is the time it takes for node i data to be transmitted in the network, is the time that the data of node i is processed on the node;

[0031] Normalized load: The normalized load of a node is expressed as the ratio of the current load to the maximum processing capacity of the node:

[0032]

[0033] Among them, the left side of the equation is the normalized load of node i, and the Load on the right side of the equation i Load is the actual load of node i at present, max(capacity) i is the channel capacity of node i;

[0034] Energy Consumption: The energy consumption of each node is composed of its data transmission and processing energy consumption:

[0035]

[0036] in, and is the adjustment coefficient;

[0037] Bandwidth Utilization: The bandwidth utilization of node i in the network is calculated as the bandwidth used, B used，i With the total bandwidth B total，i proportion;

[0038]

[0039] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the GNN model defines an output vector y i as follows:

[0040]

[0041] Among them, y i is the true result of the network performance of node i, is the prediction result of the network performance of node i output by the GNN model; L i is the delay of node i in the network, Load i is the normalized load of node i, E i is the energy consumption of node i, U i is the bandwidth utilization of node i; f represents the graph neural network function, represents the eigenvector of node i, A is the adjacency matrix, are the learned model parameters.

[0042] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: a process of achieving optimal offloading of overall network traffic based on the performance prediction result and combining a multi-objective optimization technology:

[0043] Establish decision variables and define a set of decision variables , where i and j represent the data source node and target offloading node respectively. The value is 0 or 1, where 1 means unloading data from node i to node j, and 0 means not unloading data from node i to node j;

[0044] If it is decided to unload data from node i to node j at time t, then ; If there is no uninstall operation, then ;

[0045] Define the following decision variables to optimize the data traffic offloading path:

[0046] Node uninstallation decision variable x i : Indicates whether to select node i as the destination node for data unloading; Indicates that node i is selected as the unloading destination node, Indicates not being selected;

[0047] Traffic Allocation Decision Variables : represents the data flow from node i to node j;

[0048] Path selection decision variables : Indicates whether data is unloaded from node i via a specific path j;

[0049] Time Window Decision Variables : The data unloading operation of the specified node i is performed within a specific time window;

[0050] Build a channel capacity model:

[0051] Data transfer volume Directly affected by the channel capacity, calculated as follows:

[0052]

[0053] in is the real-time data transmission rate calculated from the channel capacity, is the transmission time; channel capacity Directly determined by the channel model:

[0054]

[0055] in is the spectrum bandwidth of node i communication, is the transmit power of node i, is the channel gain, It's interference. is the noise power spectral density.

[0056] In combination with the first aspect, in some implementations of the first aspect, the method further includes: an objective function of the multi-objective optimization technology is as follows:

[0057] Minimize service delays :

[0058]

[0059] in, represents the data transmission volume of node i, is the bandwidth of node i, is the processing delay of node i, It’s processing power;

[0060] The calculation of is obtained by the channel capacity formula:

[0061]

[0062] in is the signal power sent from node i to node j at time t, is the channel gain of the link between node i and node j at time t, is the co-frequency interference power received by node i when communicating with node j at time t, is the noise power spectral density;

[0063] Minimize energy consumption :

[0064] in, and denote the energy efficiency coefficients of data transmission and processing of node i, respectively;

[0065] Maximize load balancing :

[0066] in, is the average load of all nodes;

[0067] Maximize bandwidth utilization :

[0068] in, is the actual bandwidth usage of node i;

[0069] Minimize packet loss :

[0070] in, and Respectively represent the loss and total number of packets of node i;

[0071] Maximize network throughput :

[0072] in, is the transmission time of node i;

[0073] The final optimization formula integrates the above multi-objective functions and is formalized as follows:

[0074]

[0075] Among them: Minimize means that the optimization goal is to minimize the value of the above function Z, is a decision variable, indicating whether to offload data traffic from node i to node j;

[0076] The amount of data transferred from node i to node j under the offloading decision;

[0077] is the bandwidth of node i;

[0078] is the unit data transmission energy consumption of node i;

[0079] is the load of node i, is the average load of all nodes;

[0080] is the amount of traffic offloaded from node i to node j;

[0081] and In decision making The number of packet losses and total number of packets from node i to j;

[0082] is the actual bandwidth usage of node i;

[0083] is the service quality score from node i to j;

[0084] Traffic offloading decision Length of time;

[0085] arrive Respectively represent the weight of each objective function;

[0086] The constraints include: power consumption constraint, signal-to-noise ratio constraint, service delay constraint, load balancing constraint, flow conservation constraint and node selection constraint;

[0087] Finally, the genetic algorithm is applied to solve the multi-objective optimization formula of the air-ground integrated network traffic offloading decision method.

[0088] In the second aspect, in order to achieve the above-mentioned purpose, the present invention discloses an air-ground network traffic unloading decision system based on a graph neural network, comprising:

[0089] A data processing module is used to receive node-related data in the air-ground integrated network, divide the node-related data into time windows, and perform feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes;

[0090] The model optimization module is used to input the node feature data into the pre-established graph neural network model GNN, generate the representation vector of each node by encoding and aggregating the features of the nodes and edges, and train the GNN using the representation vector of each node to obtain the trained GNN;

[0091] The offloading decision module is used to receive the network status and node characteristics in the air-space-ground integrated network, input the network status and node characteristics in the air-space-ground integrated network into the trained GNN, and output the performance prediction results. Based on the performance prediction results, combined with multi-objective optimization technology, the optimal offloading of the overall network traffic can be achieved.

[0092] Beneficial effects of the present invention:

[0093] The present invention:

[0094] 1. Heterogeneous network structure modeling:

[0095] Use the graph neural network (GNN) model to capture the heterogeneous structure in the integrated space-ground network and handle the complex relationships between different types of nodes including satellites, drones, high-altitude platforms, and ground base stations;

[0096] 2. Dynamic prediction mechanism based on graph neural network: This invention uses graph neural network (GNN) to process and analyze the complexity and high dynamics of the integrated space-ground network. Through in-depth learning of network topology, GNN can capture changes in network status in real time and predict key performance indicators such as latency, load, and bandwidth utilization, thereby providing a scientific basis for decision-making. The innovation of this prediction mechanism lies in its ability to maintain efficient and accurate prediction performance in a highly dynamic and changing network environment.

[0097] 3. Multi-objective optimization decision algorithm: The present invention constructs a multi-objective optimization framework to handle multiple competing performance indicators, such as minimizing service delay, minimizing energy consumption, maximizing load balancing, and maximizing bandwidth utilization. Using heuristic algorithms such as evolutionary algorithms to solve the multi-objective optimization problem can effectively find the best offloading strategy that balances various performance indicators. The innovation of this method lies in the ability to integrate multiple network performance requirements and optimize multiple objectives at the same time, providing a comprehensive and balanced network management strategy.

[0098] 4. Adaptive traffic offloading strategy: Through the performance indicators predicted by the GNN model and the multi-objective optimization results, the present invention can dynamically adjust the traffic offloading decision to adapt to the real-time changes in network conditions. The innovation of this strategy lies in its high adaptability and flexibility, which can automatically optimize the decision according to the real-time status of the network and the prediction results, thereby maximizing the efficiency and benefits of network resources.

[0099] The present invention can dynamically optimize data traffic offloading decisions in a complex air-ground integrated network, improve network efficiency, reduce latency and energy consumption, and provide powerful technical support for network operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0101] Figure 1 It is a schematic diagram of the overall structure of the present invention;

[0102] Figure 2It is a decision execution flow chart of the multi-objective optimization algorithm for air-ground integrated network traffic unloading of the present invention;

[0103] Figure 3 It is a decision scenario diagram of air-ground integrated network traffic offloading based on graph neural network of the present invention;

[0104] Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0105] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0106] Embodiment 1:

[0107] like Figure 1 As shown, a method for air-ground network traffic offloading decision-making based on graph neural network includes the following steps:

[0108] S101: receiving node-related data in the air-ground integrated network, dividing the node-related data into time windows, and performing feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes;

[0109] Node-related data in the integrated air-ground-space network:

[0110] First, key data is automatically collected through sensors deployed on satellites (Sats), unmanned aerial vehicles (UAVs), and ground base stations (BS); these sensors can capture a variety of data types from signal quality to environmental parameters and transmit them to the data processing center in real time; automatically collect key data in the air-ground integrated network, especially for the dynamic topology in SAGIN, collect the real-time location information of each node , the communication status between nodes , as well as the transmission capacity and communication quality of the nodes ;

[0111] The data vector will also include the transmit power , signal loss , and the distance between nodes , which is particularly critical for the subsequent channel and transmission rate model calculation;

[0112] The collected data is transmitted in real time to the high-performance server of the ground base station or uploaded to the cloud storage system for further data preprocessing and analysis. The collected data is cleaned, including missing value filling and outlier removal. In view of the high dynamic characteristics of SAGIN, special attention is paid to breakpoints and data gaps in time series data. The collected data is standardized or normalized to ensure the consistency and comparability of the data before model training. The standardization process takes into account the dynamic changes in the network:

[0113]

[0114] in, is the original key data value of the mth indicator between nodes i and j at time t, Yes The normalized value; considering the interaction between nodes i and j, and They represent the mean and standard deviation of the index m among all nodes respectively;

[0115] Time window division: According to the communication characteristics of the integrated space-ground network and the frequency of data collection, the time series data is divided into several time windows. Each time window includes data within a fixed time length, reflecting the changes in network topology caused by the movement of nodes in space. The time window length setting needs to be adjusted according to the actual network dynamics and the communication characteristics of the nodes to ensure the timeliness and relevance of the data set in model training and analysis.

[0116] S102: Input the node feature data into a pre-established graph neural network model GNN, generate a representation vector for each node by encoding and aggregating the features of the nodes and edges, train the GNN using the representation vector of each node, and obtain a trained GNN;

[0117] GNN model training includes the following sub-steps:

[0118] Optimize the graph neural network (GNN) model to adapt to the dynamic and heterogeneous structure of the Space-Ground Integrated Network (SAGIN). The GNN model consists of nodes v∈V and edges e∈E, where each node represents a network element (such as a satellite, drone, or ground base station) and the edges represent the communication connections between nodes. These structures map the physical structure and communication paths of the network.

[0119] Each node is represented by a feature vector x v Representation, including pre-processed data such as location, connection status, node type, etc.;

[0120] The graph convolution layer is used in the model construction to process the node features and the adjacency matrix A to learn the implicit representation of the node. The graph convolution layer updates the node status by utilizing the topological structure between nodes, so that the model can capture and learn the complex patterns of dynamic changes in the network; the mathematical form of graph convolution is:

[0121]

[0122] in, is the node feature matrix of the lth layer, is the node feature matrix of the next layer l+1, is the adjacency matrix A plus the identity matrix , yes The degree matrix of is the weight of the lth layer, and its dimension is determined by the size of the input and output features. is a nonlinear activation function ReLU. In each layer of graph convolution, the node will aggregate information from its neighbors and update its state, thereby propagating and synthesizing information in the network; the optimization process includes adjusting the weights of each layer , to minimize the prediction error;

[0123] GNN model training and regularization, using the collected historical data to train the GNN model, the goal is to minimize the difference between the predicted performance indicators (such as latency, load, and energy consumption) and the actual value; first, the data is divided into 70% training set, 15% validation set, and 15% test set. This division ensures that the model can perform reasonable performance evaluation on the data in this scenario;

[0124] During the training process, the weights of the graph convolution layer are continuously adjusted through back propagation and gradient descent methods. ; Loss function is used in the training process, such as mean square error (MSE):

[0125]

[0126] in, is the predicted value of the model, y i is the true value, N is the number of samples;

[0127] The new loss function Loss after introducing regularization technology and adding penalty terms to weights * :

[0128]

[0129] in, is the regularization parameter, is the weight of the lth layer.

[0130] Hyperparameter optimization and model validation: Use the validation set to adjust the model's hyperparameters, such as learning rate, number and size of hidden layers, regularization strength, etc., to find the optimal parameter settings; use the early stopping strategy to prevent overfitting, that is, if the loss on the validation set does not drop significantly within several consecutive epochs, stop training to avoid overfitting the training data during training;

[0131] S103: Receive the network status and node characteristics in the air-space-ground integrated network, input the network status and node characteristics in the air-space-ground integrated network into the trained GNN, output the performance prediction result, and combine the multi-objective optimization technology based on the performance prediction result to achieve the optimal offloading of the overall network traffic.

[0132] Performance prediction aims to perform effective multi-objective optimization in the air-ground integrated network scenario to select the best traffic distribution and offloading strategy. This step requires using the trained GNN model to predict multiple key performance indicators in the network, which includes the following sub-steps:

[0133] Definition of key performance indicators;

[0134] Define key performance indicators of the network, including latency, load, energy consumption, and bandwidth utilization. These indicators provide predictions for current and future network states and support the decision-making process.

[0135] Latency: The latency of each node in the network is calculated as the sum of transmission and processing delays:

[0136]

[0137] in, is the time it takes for node i data to be transmitted in the network, is the time that the data of node i is processed on the node;

[0138] Normalized load: The normalized load of a node is expressed as the ratio of the current load to the maximum processing capacity of the node:

[0139]

[0140] Among them, the left side of the equation is the normalized load of node i, and the Load on the right side of the equation i Load is the actual load of node i at present, max(capacity) i is the channel capacity of node i;

[0141] Energy Consumption: The energy consumption of each node is composed of its data transmission and processing energy consumption:

[0142]

[0143] in, and is the adjustment coefficient;

[0144] Bandwidth Utilization: The bandwidth utilization of node i in the network is calculated as the bandwidth used, B used，i With the total bandwidth B total，i proportion;

[0145]

[0146] Use the trained graph neural network model to predict the performance indicators defined above based on the input network status and node features; the model uses the node feature vector x i For each node i, we output the vector y i as follows:

[0147]

[0148] Among them, y i is the true result of the network performance of node i, is the prediction result of the network performance of node i output by the GNN model; L i is the delay of node i in the network, Load i is the normalized load of node i, E i is the energy consumption of node i, U i is the bandwidth utilization of node i; f represents the graph neural network function, represents the eigenvector of node i, A is the adjacency matrix, are the learned model parameters.

[0149] The process of realizing the optimal traffic offloading strategy using multi-objective optimization technology:

[0150] Node uninstallation decision variable x i : Indicates whether to select node i as the destination node for data unloading; Indicates that node i is selected as the unloading destination node, Indicates not being selected;

[0151] Traffic Allocation Decision Variables : represents the data flow from node i to node j;

[0152] Path selection decision variables : Indicates whether data is unloaded from node i via a specific path j;

[0153] Time Window Decision Variables : The data unloading operation of the specified node i is performed within a specific time window;

[0154] Build a channel capacity model:

[0155] Data transfer volume Directly affected by the channel capacity, calculated as follows:

[0156]

[0157] in is the spectrum bandwidth of node i communication, is the real-time data transmission rate calculated from the channel capacity, is the transmission time; channel capacity Directly determined by the channel model:

[0158]

[0159] in is the transmit power of node i, is the channel gain, It's interference. is the noise power spectral density.

[0160] The objective function of the multi-objective optimization technique is as follows:

[0161] Minimize service delays :

[0162]

[0163] in, represents the data transmission volume of node i, is the bandwidth of node i, is the processing delay of node i, It’s processing power;

[0164] The calculation of is obtained by the channel capacity formula:

[0165]

[0166] in is the signal power sent from node i to node j at time t, is the channel gain of the link between node i and node j at time t, is the co-frequency interference power received by node i when communicating with node j at time t, is the noise power spectral density;

[0167] Minimize energy consumption :

[0168] in, and denote the energy efficiency coefficients of data transmission and processing of node i, respectively;

[0169] Maximize load balancing :

[0170] in, is the average load of all nodes;

[0171] Maximize bandwidth utilization :

[0172] in, is the actual bandwidth usage of node i;

[0173] Minimize packet loss :

[0174] in, and Respectively represent the loss and total number of packets of node i;

[0175] Maximize network throughput :

[0176] in, is the transmission time of node i;

[0177] The final optimization formula integrates the above multi-objective functions and is formalized as follows:

[0178]

[0179] Among them: Minimize means that the optimization goal is to minimize the value of the above function Z, is a decision variable, indicating whether to offload data traffic from node i to node j;

[0180] The amount of data transferred from node i to node j under the offloading decision;

[0181] is the bandwidth of node i;

[0182] is the unit data transmission energy consumption of node i;

[0183] is the load of node i, is the average load of all nodes;

[0184] is the amount of traffic offloaded from node i to node j;

[0185] and In decision making The number of packet losses and total number of packets from node i to j;

[0186] is the actual bandwidth usage of node i;

[0187] is the service quality score from node i to j;

[0188] Traffic offloading decision Length of time;

[0189] arrive Respectively represent the weight of each objective function;

[0190] The constraints include: power consumption constraint, signal-to-noise ratio constraint, service delay constraint, load balancing constraint, flow conservation constraint and node selection constraint;

[0191] Finally, the genetic algorithm is applied to solve the multi-objective optimization formula of the air-ground integrated network traffic offloading decision method.

[0192] Power consumption constraint, in order to ensure that the power consumption of the network is within an acceptable range, the power consumption of each node must not exceed the maximum power P max , the formula is:

[0193]

[0194] in, represents the power output of node i at time t;

[0195] The signal-to-noise ratio constraint ensures that the signal-to-noise ratio (SNR) of the communication link from each node to other nodes at least meets the minimum standard sinr min , the formula is:

[0196]

[0197] in, Represents the signal-to-noise ratio of the corresponding node, and the expression is:

[0198]

[0199] Service delay constraint, service delay is one of the key indicators of network performance, and it must be ensured that the maximum delay L max , the formula is:

[0200]

[0201] Among them, L i is the service delay of node i, L max is the maximum delay allowed by the network;

[0202] Load balancing constraints: To ensure stable operation of the network and optimize resource utilization, the load of each node should be close to the average load of the network. , the difference does not exceed a certain set threshold , the formula is:

[0203]

[0204] in, is the average load of all nodes, is the maximum allowable load deviation;

[0205] The flow conservation constraint ensures that the total flow sent from node i is equal to the sum of its allocated flows:

[0206]

[0207] where F i is the total traffic that node i needs to offload;

[0208] Node selection constraints ensure that traffic is offloaded only between selected nodes;

[0209]

[0210] Where M is a sufficiently large number to ensure hour .

[0211] Summary: The multi-objective optimization objective function and its constraints are expressed as follows:

[0212]

[0213] The genetic algorithm is applied to solve the multi-objective optimization formula of the air-ground integrated network traffic offloading decision method. The genetic algorithm includes the following algorithm steps:

[0214] Initialization: Randomly generate a certain number of initial solution sets. Each solution represents a possible traffic offloading strategy, that is, the direction and amount of data flow between nodes;

[0215] Evaluation: Calculate the fitness of each solution, that is, evaluate the performance indicators of each decision based on the current state of the network, including latency, energy consumption, bandwidth utilization, etc.;

[0216] Selection: Select the better solutions to enter the next generation based on fitness. Use methods such as tournament selection and roulette selection to determine which solutions will be retained or eliminated;

[0217] Crossover and mutation: Generate new solutions through crossover and mutation operations. The crossover operation allows two solutions to merge their characteristics, while the mutation operation randomly changes some parameters of the solution to explore new possibilities;

[0218] Replacement: The generated new solution replaces the worst solution in the original solution set;

[0219] Termination condition: Repeat the above steps until the termination condition is met, such as reaching the maximum number of iterations or the quality of the solution no longer improves;

[0220] After completing the multi-objective optimization solution, the final unloading decision vector is output It can be expressed as:

[0221]

[0222] Among them, j represents the target node, t i It can be further allocated to different time periods based on actual conditions to optimize service response and resource utilization.

[0223] like Figure 2 The flow chart of the multi-objective optimization algorithm decision execution of air-ground-integrated network traffic unloading shows a detailed description of the prediction and optimization method of using the air-ground-integrated network traffic unloading decision method to perform traffic unloading. The process includes the following key steps:

[0224] Start: Starts the process.

[0225] Initialize network status: including real-time network data collection, continuously collect real-time network data, such as communication delay between nodes, packet loss rate, etc.

[0226] Key performance indicator prediction: Use pre-trained graph neural network (GNN) models such as Figure 3 As shown, short-term network performance indicators are predicted based on historical and real-time data.

[0227] Determine the traffic offloading of the integrated air-ground-space network based on the predicted data: Apply multi-objective optimization algorithms to balance various performance indicators, such as minimizing latency and energy consumption, maximizing bandwidth utilization and service quality, and even constrain variables to solve multi-objective optimization problems and obtain the optimal traffic offloading strategy.

[0228] Unloading strategy verification and decision feedback adjustment: The strategy is updated based on the optimization result feedback to achieve a feedback loop and ensure continuous optimization of network performance.

[0229] End: The process ends.

[0230] This process provides a comprehensive solution framework for air-ground integrated network traffic offloading optimization, which is used to optimize data traffic in a complex air-ground integrated network environment in real time and dynamically. Through this process, network performance and resource utilization can be effectively improved while ensuring the quality of user services and network reliability.

[0231] Embodiment 2: In the second aspect, as Figure 4 As shown, in order to achieve the above-mentioned purpose, the present invention discloses an air-ground network traffic unloading decision system based on a graph neural network, comprising:

[0232] The data processing module 11 is used to receive node-related data in the air-ground integrated network, divide the node-related data into time windows, and perform feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes;

[0233] The model optimization module 12 is used to input the node feature data into the pre-established graph neural network model GNN, generate a representation vector for each node by encoding and aggregating the features of the nodes and edges, and train the GNN using the representation vector of each node to obtain a trained GNN;

[0234] The unloading decision module 13 is used to receive the network status and node characteristics in the air-space-ground integrated network, input the network status and node characteristics in the air-space-ground integrated network into the trained GNN, and output the performance prediction results. Based on the performance prediction results, combined with multi-objective optimization technology, the optimal unloading of the overall network traffic can be achieved.

[0235] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0236] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by the processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0237] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0238] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. A decision-making method for air-ground network traffic offloading based on graph neural network, characterized in that: The method comprises the following steps: Receive node-related data in the air-ground integrated network, divide the node-related data into time windows, and perform feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes; The node feature data is input into the pre-established graph neural network model GNN, and the representation vector of each node is generated by encoding and aggregating the features of the nodes and edges. The GNN is trained using the representation vector of each node to obtain the trained GNN. Receive the network status and node characteristics in the air-ground integrated network, input the network status and node characteristics in the air-ground integrated network into the trained GNN, output the performance prediction results, and combine the multi-objective optimization technology based on the performance prediction results to achieve the optimal offloading of the overall network traffic; The network status and node features in the air-ground integrated network are input into the trained GNN, and the output performance prediction results include multiple key performance indicators. The multiple key performance indicators are as follows: Latency: The latency of each node in the network is calculated as the sum of transmission and processing delays, L: L i =t transmit，i +t process，i Among them, τ transmit，i is the time it takes for node i data to be transmitted in the network, τ process，i is the time that the data of node i is processed on the node; Normalized load: The normalized load of a node is expressed as the ratio of the current load to the maximum processing capacity of the node: Among them, the left side of the equation is the normalized load of node i, and the Load on the right side of the equation i Load is the actual load of node i at present, max(capacity) i is the channel capacity of node i; Energy Consumption: The energy consumption of each node is composed of the energy consumption of its data transmission and processing. i : E i =a i D transmit，i +b i L i Among them, α i and β i It is used to balance the amount of data D transmitted by node i transmit，i and i-node processing delay L i The adjustment coefficient of performance weight; Bandwidth Utilization: The bandwidth utilization of node i in the network is calculated as the bandwidth used, B used，i With the total bandwidth B total，i The ratio U i ; The output vector y of the GNN is defined i as follows: Among them, y i is the true result of the network performance of node i, is the prediction result of the network performance of node i output by the GNN model; L i is the delay of node i in the network, Load i is the normalized load of node i, E i is the energy consumption of node i, U i is the bandwidth utilization of node i; f represents the graph neural network function, x i represents the feature vector of node i, A is the adjacency matrix, and θ is the learned model parameter; The process of achieving optimal traffic offloading of the entire network based on performance prediction results combined with multi-objective optimization technology: Establish decision variables and define a set of decision variables x i，j , where i and j represent the data source node and target offloading node respectively, x i，j The value is 0 or 1, where 1 means unloading data from node i to node j, and 0 means not unloading data from node i to node j; If it is decided to offload data from node i to node j at time t, then x i，j (t) = 1; if there is no uninstallation operation, then x i，j (t) = 0; Define the following decision variables to optimize the data traffic offloading path: Node uninstallation decision variable x i : Indicates whether to select node i as the destination node for data unloading; x i =1 indicates that node i is selected as the unloading destination node, x i =0 means not selected; Traffic allocation decision variable f i，j : represents the data flow from node i to node j; Path selection decision variable p i，j : Indicates whether the data is unloaded from node i via the predetermined path j; Time window decision variable t i : The data unloading operation of the specified node i is performed within the predetermined time window; Build a channel capacity model: Data transfer volume D transmit，i Directly affected by the channel capacity, calculated as follows: D transmit，i =r i ×T i where r i is the real-time data transmission rate calculated from the channel capacity, T i is the transmission time; the channel capacity r i Directly determined by the channel model: Among them B i is the spectrum bandwidth of node i’s communication, p i is the transmit power of node i, g i is the channel gain, I i is interference, N0 is noise power spectral density; The objective function of the multi-objective optimization technique is as follows: Minimize service delay F latency : Among them, D transmit，i represents the data transmission volume of node i, B i is the bandwidth of node i, L i is the processing delay of node i, C i It’s processing power; F latency The calculation of is obtained by the channel capacity formula: where p i，j (t) is the signal power sent from node i to node j at time t, g i，j (t) is the channel gain of the link between node i and node j at time t, I i，j (t) is the co-channel interference power received by node i when communicating with node j at time t, and N0 is the noise power spectrum density; Minimize energy consumption energy : Among them, α i and β i denote the energy efficiency coefficients of data transmission and processing of node i, respectively; Maximize Load Balancing F load : in, is the average load of all nodes; Maximize bandwidth utilization F bandwidth : Among them, U i is the actual bandwidth usage of node i; Minimize the packet loss rate F loss : Among them, P loss，i and P total，i denote the loss and total number of packets of node i respectively; maximize the network throughput F throughput : Among them, T i is the transmission time of node i; The final optimization formula integrates the above multiple objective functions and is formalized as follows: Minimize Among them: Minimize means that the optimization goal is to minimize the value of the above function Z, x i，j is a decision variable, indicating whether to offload data traffic from node i to node j; D transmit，i，j The amount of data transferred from node i to node j under the offloading decision; B i is the bandwidth of node i; α i is the unit data transmission energy consumption of node i; L i is the load of node i, is the average load of all nodes; f i，j is the amount of traffic offloaded from node i to node j; P loss，i，j and P total，i In decision x i，j The number of packet losses and total number of packets from node i to j; U i is the actual bandwidth usage of node i; t i，j is the traffic offloading decision x i，j Length of time; λ1 to λ6 represent the weights of each objective function; The constraints include: power consumption constraint, signal-to-noise ratio constraint, service delay constraint, load balancing constraint, flow conservation constraint and node selection constraint; Finally, the genetic algorithm is applied to solve the multi-objective optimization formula of the air-ground integrated network traffic offloading decision method.

2. The air-to-ground network traffic offloading decision method based on graph neural network according to claim 1 is characterized in that: The node-related data in the integrated space-ground network is obtained as follows: the key data is automatically collected by sensors deployed on satellites Sat, UAVs, and ground base stations BS; real-time transmission to the data processing center; collecting key data in the integrated space-ground network, and collecting the real-time location information of each node in the dynamic topological structure of the integrated space-ground network SAGIN. i (t), the communication status between nodes conn i，j (t), as well as the node’s transmission capability and communication qualityi(t); Also includes the transmission power P i (t), signal loss L i，j (t), and the distance d between nodes i，j (t).

3. The air-to-ground network traffic offloading decision method based on graph neural network according to claim 2 is characterized in that: The node-related data in the air-ground integrated network is standardized or normalized, and the standardization process takes into account the dynamic changes in the network: Among them, x i，j，m (t) is the original key data value of the mth indicator between nodes i and j at time t, x′ i，j，m (t) is the value of x i，j，m (t) is the normalized value; considering the interaction between nodes i and j, μ i，j，m and σ i，j，m They represent the mean and standard deviation of the index m among all nodes respectively; Time window division process: According to the communication characteristics of the integrated air-space-ground network and the data collection frequency, the time series data is divided into several time windows. Each time window includes data within a fixed time length, reflecting the changes in network topology caused by the movement of nodes in space; among them, the time window length setting is adjusted according to the network dynamics and the communication characteristics of the nodes.

4. The air-to-ground network traffic offloading decision method based on graph neural network according to claim 1 is characterized in that: The pre-established graph neural network model GNN includes nodes V∈V and edges e∈E, wherein each node represents a network device and the edge represents the communication connection between nodes; Each node is represented by a feature vector x v express; The graph convolution layer is used in the model construction to process the node features and the adjacency matrix A to learn the implicit representation of the node. The graph convolution layer updates the node status by utilizing the topological structure between nodes. The mathematical form of graph convolution is: Among them, H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the next layer l+1, is the adjacency matrix A plus the identity matrix I N , yes The degree matrix, W (l) is the weight of the lth layer, and σ is the nonlinear activation function ReLU.

5. The air-to-ground network traffic offloading decision method based on graph neural network according to claim 1 is characterized in that: In the process of training GNN using the representation vector of each node, the mean square error MSE is used to construct the loss function Loss: in, is the predicted value of the model, yi is the true value, and N is the number of samples; The new loss function Loss* after introducing regularization technology and adding penalty terms to weights is: Among them, λ is the regularization parameter, W (l) is the weight of the lth layer.

6. The air-ground network traffic offloading decision system based on graph neural network is characterized by: The air-to-ground network traffic offloading decision method based on a graph neural network as described in any one of claims 1 to 5 is adopted, comprising: A data processing module is used to receive node-related data in the air-ground integrated network, divide the node-related data into time windows, and perform feature extraction to obtain node feature data, wherein the node-related data includes node connection status data and real-time traffic data between nodes; The model optimization module is used to input the node feature data into the pre-established graph neural network model GNN, generate the representation vector of each node by encoding and aggregating the features of the nodes and edges, and train the GNN using the representation vector of each node to obtain the trained GNN; The offloading decision module is used to receive the network status and node characteristics in the air-ground integrated network, input the network status and node characteristics in the air-ground integrated network into the trained GNN, and output the performance prediction results. Based on the performance prediction results, combined with multi-objective optimization technology, the optimal offloading of the overall network traffic can be achieved; The network status and node features in the air-ground integrated network are input into the trained GNN, and the output performance prediction results include multiple key performance indicators. The multiple key performance indicators are as follows: Latency: The latency of each node in the network is calculated as the sum of transmission and processing delays L i : L i =t transmit，i +t process，i Among them, τ transmit，i is the time it takes for node i data to be transmitted in the network, τ process，i is the time that the data of node i is processed on the node; Normalized load: The normalized load of a node is expressed as the ratio of the current load to the maximum processing capacity of the node: Among them, the left side of the equation is the normalized load of node i, and the Load on the right side of the equation i Load is the actual load of node i at present, max(capacity) i is the channel capacity of node i; Energy Consumption: The energy consumption of each node is composed of the energy consumption of its data transmission and processing. i : E i =a i D transmit，i +b i L i Among them, α i and β i It is used to balance the amount of data D transmitted by node i transmit，i and i-node processing delay L i The adjustment coefficient of performance weight; Bandwidth Utilization: The bandwidth utilization of node i in the network is calculated as the bandwidth used, B used，i With the total bandwidth B total，i The ratio U i ; The output vector y of the GNN is defined i as follows: Among them, y i is the true result of the network performance of node i, is the prediction result of the network performance of node i output by the GNN model; L i is the delay of node i in the network, Load i is the normalized load of node i, E i is the energy consumption of node i, U i is the bandwidth utilization of node i; f represents the graph neural network function, x i represents the feature vector of node i, A is the adjacency matrix, and θ is the learned model parameter; The process of achieving optimal traffic offloading of the entire network based on performance prediction results combined with multi-objective optimization technology: Establish decision variables and define a set of decision variables x i，j , where i and j represent the data source node and target offloading node respectively, x i，j The value is 0 or 1, where 1 means unloading data from node i to node j, and 0 means not unloading data from node i to node j; If it is decided to offload data from node i to node j at time t, then x i，j (t) = 1; if there is no uninstallation operation, then x i，j (t) = 0; Define the following decision variables to optimize the data traffic offloading path: Node uninstallation decision variable x i : Indicates whether to select node i as the destination node for data unloading; x i =1 indicates that node i is selected as the unloading destination node, x i =0 means not selected; Traffic allocation decision variable f i，j : represents the data flow from node i to node j; Path selection decision variable p i，j : Indicates whether the data is unloaded from node i via the predetermined path j; Time window decision variable t i : The data unloading operation of the specified node i is performed within the predetermined time window; Build a channel capacity model: Data transfer volume D transmit，i Directly affected by the channel capacity, calculated as follows: D transmit,i =r i ×T i where r i is the real-time data transmission rate calculated from the channel capacity, T i is the transmission time; the channel capacity r i Directly determined by the channel model: Among them B i is the spectrum bandwidth of node i’s communication, p i is the transmit power of node i, g i is the channel gain, I i is interference, N0 is noise power spectral density; The objective function of the multi-objective optimization technique is as follows: Minimize service delay F latency : Among them, D transmit，i represents the data transmission volume of node i, B i is the bandwidth of node i, L i is the processing delay of node i, C i It’s processing power; F latency The calculation of is obtained by the channel capacity formula: where p i，j (t) is the signal power sent from node i to node j at time t, g i，j (t) is the channel gain of the link between node i and node j at time t, I i，j (t) is the co-channel interference power received by node i when communicating with node j at time t, and N0 is the noise power spectrum density; Minimize energy consumption energy : Among them, α i and β i denote the energy efficiency coefficients of data transmission and processing of node i, respectively; Maximize Load Balancing F load : in, is the average load of all nodes; Maximize bandwidth utilization F bandwidth : Among them, U i is the actual bandwidth usage of node i; Minimize the packet loss rate F loss : Among them, P loss，i and P total，i denote the loss and total number of packets of node i respectively; maximize the network throughput F throughput : Among them, T i is the transmission time of node i; The final optimization formula integrates the above multiple objective functions and is formalized as follows: Minimize Among them: Minimize means that the optimization goal is to minimize the value of the above function Z, x i，j is a decision variable, indicating whether to offload data traffic from node i to node j; D transmit，i，j The amount of data transferred from node i to node j under the offloading decision; B i is the bandwidth of node i; α i is the unit data transmission energy consumption of node i; L i is the load of node i, is the average load of all nodes; f i，j is the amount of traffic offloaded from node i to node j; P loss，i，j and P total，i In decision x i，j The number of packet losses and total number of packets from node i to j; U i is the actual bandwidth usage of node i; t i，j is the traffic offloading decision x i，j Length of time; λ1 to λ6 represent the weights of each objective function; The constraints include: power consumption constraint, signal-to-noise ratio constraint, service delay constraint, load balancing constraint, flow conservation constraint and node selection constraint; Finally, the genetic algorithm is applied to solve the multi-objective optimization formula of the air-ground integrated network traffic offloading decision method.

Images