Terminal inter-satellite link switching method based on satellite-ground heterogeneous data

Abstract

The invention belongs to the technical field of satellite communication, and particularly relates to a terminal inter-satellite link switching method based on satellite-ground heterogeneous data. The method comprises the following steps: constructing an LEO satellite mobile communication network and a graph structure thereof according to an LEO satellite, a user terminal, a ground gateway station and a communication link thereof; and acquiring satellite orbit and link state data, ground gateway station data, meteorological data and user behavior data in real time, and inputting the data into the trained inter-satellite link switching prediction model of the terminal after data preprocessing and feature coding to obtain an optimal inter-satellite link switching strategy of the terminal. The prediction model is based on a multi-layer GNN model and introduces a weight matrix, a time sequence enhancement and time sequence attention mechanism and a dynamic adjustment mechanism, so that comprehensive evaluation and subsequent decision-making of various data of switching candidate satellites can be realized, and an optimal terminal inter-satellite link switching strategy is obtained. According to the invention, the success rate and efficiency of inter-satellite link switching are effectively improved, the risks of switching delay and communication interruption are reduced, and the communication quality is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of satellite communication, and particularly relates to a method for switching inter-satellite links of terminals based on space-ground heterogeneous data. Background Art

[0002] Currently, satellite network technology has developed rapidly, and can provide global coverage and persistent communication services. By using satellite relays to achieve communication between mobile users and fixed users, satellite communication has gradually become an important part of global mobile communication. Due to its wide coverage, low transmission loss, high communication capacity and other advantages, it is widely used in many fields such as military, disaster prevention and relief, field exploration, and personal mobile communication. Low Earth Orbit (LEO) mobile communication uses a group of satellites with much lower orbits than the geostationary equatorial orbit (GEO) to achieve communication between mobile users or between mobile users and fixed users. Through multi-satellite networking, seamless coverage of the earth can be achieved, and users can view multiple satellites simultaneously to ensure the reliability of communication. LEO satellites are distributed at an altitude of about 750 - 1500 kilometers above the ground. The coverage range of a single LEO satellite is relatively small, and the coverage time is only about ten minutes. The satellite networking is complex. LEO satellites rotate at a high speed relative to the earth's surface. In order to achieve persistent communication services, it is necessary to continuously and frequently switch connections between LEO satellites.

[0003] Traditional LEO satellite switching technologies mainly rely on single inter-satellite link data for switching decisions. These data usually include basic parameters such as signal strength, link delay, and bit error rate. However, LEO satellite communication networks are known for their high mobility and complex dynamic environments, making it difficult to comprehensively reflect the actual state and communication quality of the network only relying on this single-dimensional inter-satellite link information.

[0004] The rapid movement of LEO satellite communication networks leads to frequent changes in the state of inter-satellite links, and traditional technologies cannot capture these changes in a timely manner, greatly reducing the accuracy of handover prediction. Traditional technologies ignore important factors such as ground gateway station data, user requirements, and weather conditions, which have a significant impact on the success rate of inter-satellite link switching and communication quality. Summary of the Invention

[0005] The object of the present invention: By introducing multi-source heterogeneous data, such as coordination data between ground stations and satellites, user requirement data, and weather data, etc. Combining multi-dimensional data, the present invention can more comprehensively evaluate the situation of inter-satellite handover, thereby improving the accuracy of handover prediction. At the same time, a more reasonable and efficient handover strategy is formulated according to the weights and supports of heterogeneous data, optimizing the utilization of network resources and improving the overall communication performance.

[0006] To solve the above technical problems, the present invention proposes a method for terminal inter-satellite link switching based on space-ground heterogeneous data, which includes:

[0007] S1: Construct a LEO satellite mobile communication network and its graph structure according to LEO satellites, user terminals, ground gateways, and their communication links;

[0008] S2: In response to a request from a user terminal, real-time obtain satellite orbit and link status data, ground gateway data, meteorological data, and user behavior data, perform data preprocessing and feature encoding on them, and obtain the corresponding processed data;

[0009] S3: Based on the corresponding processed data and the graph structure G of the LEO satellite mobile communication network, extract the feature vectors of each node and each edge;

[0010] S4: Input the feature vectors of each node and each edge into the trained terminal inter-satellite link switching prediction model to obtain the optimal terminal inter-satellite link switching strategy.

[0011] Further, the training process of the terminal inter-satellite link switching prediction model includes:

[0012] S101: Obtain a historical data set of the LEO satellite mobile communication network, which includes historical satellite orbit and link status data, historical ground gateway data, historical meteorological data, historical user behavior data, and historical switching results;

[0013] S102: Perform preprocessing and feature encoding on the historical data set of the LEO satellite mobile communication network to obtain the corresponding processed data;

[0014] S103: Calculate the support degree of each data source according to the historical data set of the LEO satellite mobile communication network, and initialize the weight matrix W of each data source in combination with expert knowledge 0 ;

[0015] S104: Based on the processed data and the graph structure of the LEO satellite mobile communication network, in combination with the support degree of each data source and the weight matrix W 0 , extract the feature vectors of each node and each edge, and input them into the prediction model;

[0016] S105: Use multiple layers of GCN layers to aggregate the features of neighbor nodes to the target node through multiple layers of vectors, and extract the high-order feature vectors of the node;

[0017] S106: Input the high-order feature vectors of the node into the Transformer for position encoding and time information enhancement, and extract the temporal feature vectors of the node;

[0018] S107: Input the timing feature vector of the node into the LSTM to obtain the node features after timing enhancement;

[0019] S108: Input the node features after timing enhancement and the feature vectors of each edge into the fully connected layer, and calculate the comprehensive score S of each satellite node sat , select the satellite node i* with the highest comprehensive score for switching to obtain the predicted switching result;

[0020] S109: Calculate the loss between the predicted switching result and the historical switching result, and backpropagate through this loss to optimize the terminal inter-satellite link switching prediction model.

[0021] The present invention utilizes multi-source heterogeneous data and deep learning technology. It not only collects inter-satellite link data, but also integrates multi-source heterogeneous data such as ground gateway station feedback, real-time weather conditions, and user behavior patterns, greatly enriching the input dimension of the model. Specifically, the ground gateway station data provides real-time quality feedback on satellite signal reception, weather data (such as cloud cover, precipitation, wind speed) reveals the impact of environmental factors on signal transmission, and user behavior data reflects the dynamic fluctuations of network demands. These data together constitute a comprehensive and detailed network state portrait.

[0022] The key to the technological innovation lies in, firstly, simplifying non-traditional data sources such as weather data and user behavior through feature encoding so that they can be used as inputs for the graph neural network (GCN). Secondly, the present invention introduces a weight matrix to reflect the importance and reliability of different data sources and features in the decision-making process. To enhance the model's perception ability of time information, the present invention adopts timing feature encoding, such as sine and cosine function encoding methods, to integrate time information into the node features. Thirdly, the present invention introduces a timing attention mechanism. By calculating the similarity of neighbor nodes at the current time step, timing attention weights are generated to highlight the features at important time steps and improve the model's attention to key time information. Through multi-layer GCN feature integration and timing enhancement, a node feature representation containing rich local and global graph information is obtained. Based on these features, the feature scores of each candidate satellite are calculated, and the satellite with the highest score is selected for switching by comprehensively considering various data scores, ensuring the performance and stability of the communication network.

[0023] This innovation not only enriches the input of the prediction model, but also significantly improves the accuracy of the comprehensive performance evaluation of candidate satellites and the level of decision-making intelligence, ensuring efficient and stable communication in a complex and changing satellite network environment. In addition, this method is particularly suitable for scenarios such as remote areas, the ocean, and emergency disaster relief, demonstrating broad application prospects and important social value.

[0024] Advantages of the present invention:

[0025] The present invention expands the data dimension, performs fine feature encoding, introduces temporal feature encoding and an attention mechanism to enhance the model's perception ability of time information, evaluates the performance of handover candidate satellites through deep learning based on the switching decision support degrees and influence weights of heterogeneous data pairs, selects the optimal handover candidate satellite according to the comprehensive data features, significantly improves the accuracy and real-time performance of inter-satellite link handover, aims to improve the success rate and efficiency of inter-satellite handover, and reduce the risks of handover delay and communication interruption. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of the steps of an embodiment of the present invention;

[0027] Figure 2 is a system schematic diagram of a LEO satellite mobile communication network in an embodiment of the present invention;

[0028] Figure 3 is a schematic flowchart of data preprocessing and feature encoding in an embodiment of the present invention;

[0029] Figure 4 is a flowchart of the training steps of a terminal inter-satellite link handover prediction model in an embodiment of the present invention;

[0030] Figure 5 is a schematic flowchart of satellite handover decision-making in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] Figure 1 is a flowchart of the steps of an embodiment of the present invention.

[0033] An embodiment of the present invention proposes a terminal inter-satellite link handover method based on space-ground heterogeneous data. Referring to Figure 1 as shown, the method includes:

[0034] S1: Construct a LEO satellite mobile communication network and its graph structure according to LEO satellites, user terminals, ground gateway stations, and their communication links;

[0035] S2: In response to the request of the user terminal, real-time obtain satellite orbit and link status data, ground gateway station data, meteorological data, and user behavior data, perform data preprocessing and feature encoding on them, and obtain the corresponding processed data;

[0036] S3: Based on the processed corresponding data and the graph structure G of the LEO satellite mobile communication network, extract the feature vectors of each node and each edge.

[0037] S4: Input the feature vectors of each node and each edge into the trained terminal inter-satellite link handover prediction model to obtain the optimal terminal inter-satellite link handover strategy.

[0038] Figure 2 It is a schematic diagram of the system of the LEO satellite mobile communication network in the embodiment of the present invention.

[0039] Refer to Figure 2 As shown, in a Low Earth Orbit (LEO) satellite mobile communication network, there are k LEO satellites orbiting the Earth's surface at high speed, which are distributed at an altitude of approximately 750 - 1500 kilometers from the ground. In the same target area, there are numerous users U (user) and ground gateway stations. The coverage area of a single LEO satellite is relatively small, and the coverage time is only about ten minutes. In order to achieve persistent communication services, users in the same target area need to continuously and frequently switch communication links between LEO satellites. The communication link switching can be carried out between the user-satellite service link or the ground gateway station-satellite feeder link.

[0040] Usually, LEO satellite handover mainly relies on single inter-satellite link data for handover decisions, but important factors such as ground gateway station data, user requirements, and weather conditions also have a significant impact on the success rate and communication quality of inter-satellite link handover. In the embodiment of the present invention, these important factors are introduced into the inter-satellite communication link handover to obtain a more reasonable and efficient handover strategy, optimize network resource utilization, and improve the overall communication performance.

[0041] The present invention uses 4 types of related data sets for satellite handover prediction, specifically: satellite orbit and link status data, ground gateway station data, meteorological data, and user behavior data.

[0042] In a preferred embodiment, the satellite orbit and link status data include the remaining service time, orbital altitude, position coordinates, speed, and orbital inclination of the satellite; the ground gateway station data include the satellite signal reception quality, satellite signal strength, and satellite signal arrival time; the meteorological data include cloud cover, precipitation, wind speed, temperature, humidity, and air pressure; the user behavior data include user behavior categories, the frequency of user behavior, the duration of user behavior, and the average interval of user behavior.

[0043] Specifically, obtain satellite orbit and link status data in real time. Obtain data from the satellite control center in real time, including basic information such as the remaining service time, orbital altitude, position coordinates (latitude and longitude), speed, and orbital inclination of the satellite. These data are crucial for predicting the position of the satellite and adjusting the communication link.

[0044] During the inter-satellite handover process, the ground gateway station can act as a relay station for data transmission to ensure the continuity and integrity of data for user terminals during the handover process. Especially in methods that rely on the ground gateway station for data forwarding and handover decisions, the role of the ground gateway station is even more prominent.

[0045] Specifically, obtain ground gateway station data in real time. Obtain current data forwarding status, data transmission volume, and other metrics of the satellite and the ground gateway station in real time, including satellite signal reception quality, signal strength, signal arrival time, etc. to evaluate the network load situation.

[0046] In a low Earth orbit (LEO) satellite communication network, since the satellite is relatively close to the Earth's surface, its signals are more vulnerable to various meteorological conditions in the atmosphere. Specifically, cloud thickness and precipitation will directly cause signal attenuation. Especially in LEO satellite communications in the high-frequency bands (such as Ka and Ku bands), the rain attenuation phenomenon is particularly significant, which may greatly reduce the signal strength and even cause communication interruptions.

[0047] Specifically, obtain meteorological data in real time. Integrate detailed meteorological data such as cloud cover, precipitation, wind speed, temperature, humidity, and air pressure provided by meteorological satellites, as well as relevant real-time meteorological information from ground meteorological stations. By introducing these detailed meteorological data, the LEO satellite communication system can more accurately evaluate the specific impact of weather changes on satellite communications, including but not limited to the degree of signal attenuation, communication link stability, and potential network fault points. Based on these evaluation results, the system can take countermeasures in advance, such as adjusting satellite power, switching communication bands or paths, optimizing network configuration, etc., to effectively reduce communication interruptions and data loss caused by weather factors, especially in high-risk areas.

[0048] User behavior data reflects the dynamic changes in network load and user needs, and is of great significance for predicting network load, optimizing resource allocation, and formulating handover strategies.

[0049] Specifically, obtain user behavior data in real time. Obtain information such as the geographical location, service usage data, and traffic consumption patterns of users in real time, and obtain user behavior data from this information, including: user behavior categories (such as browsing, downloading, uploading, etc.), the frequency and duration of user behavior, etc. By introducing user behavior data, the system can more accurately grasp the changing trends of user needs, thereby making handover decisions that are closer to user needs and improving user experience and service quality.

[0050] Data preprocessing aims to clean and format the original data to improve data quality and ensure the accuracy and efficiency of subsequent analysis.

[0051] In a preferred embodiment, the specific processes of the data preprocessing and feature encoding sequentially include: data cleaning, data verification, feature encoding, and normalization processing.

[0052] Figure 3 This is a schematic diagram of the specific process of data preprocessing and feature encoding in the embodiments of the present invention.

[0053] Preferably, referring to Figure 3 As shown, in the embodiments of the present invention, data preprocessing and feature encoding are performed on 4 types of related data sets, and the specific processes sequentially include: data cleaning, data verification, feature encoding, and normalization processing.

[0054] Data cleaning: Identify and delete errors and outliers, such as correcting unreasonable location data or removing obvious noise signals.

[0055] Encoding meteorological data: Sort the weather conditions according to the degree of influence on communication and perform ordered encoding.

[0056] Exemplarily, if only one-dimensional encoding is required, integer values can be directly assigned according to the degree of influence: sunny: 1 (least influence); cloudy: 2 (medium influence); rainy: 3 (medium influence); thunderstorm: 4 (greatest influence).

[0057] Feature encoding of user behavior data: Perform feature encoding on user behavior categories (such as browsing, downloading, uploading, etc.). Encoding can be performed according to features such as the frequency and duration of user behavior for subsequent analysis of user behavior patterns.

[0058] First, assign a unique integer identifier to each user behavior, which can be regarded as the basic encoding of the behavior. Exemplarily, browsing: 1; downloading: 2; uploading: 3; video call: 4; (other behaviors can be added in sequence).

[0059] Secondly, for each user behavior, further encoding can be performed according to features such as the frequency and duration of the user behavior. For simplicity, a normalization method can be used to convert these feature values to a certain range (between 0 and 1).

[0060] Frequency of user behavior: The number of times the user performs this behavior within a period of time.

[0061] Duration of user behavior: The average duration of each time the user performs this behavior.

[0062] Time distribution of user behavior: The time period distribution during which the user performs this behavior (such as morning, noon, evening, etc.), which can adopt one-hot encoding or sequential encoding.

[0063] Average interval of user behavior: The average time interval between two consecutive times the user performs this behavior, normalized to between 0 and 1.

[0064] Finally, combine the behavior category encoding and the behavior feature encoding to form a more comprehensive user behavior feature vector.

[0065] For example, for a user's "download" behavior within a certain period of time, the feature vector may be as follows:

[0066] Behavior category encoding: 2;

[0067] Frequency (normalized): 0.6 (assuming that 60% of the user's behaviors during this period are downloads)

[0068] Duration (normalized): 0.8 (assuming that the average duration of the user's download behavior is relatively long, accounting for 80% of the overall duration);

[0069] Time distribution: [0.1, 0.3, 0.6] (assuming that the user's downloads account for 10% in the morning, 30% at noon, and 60% in the evening)

[0070] Average interval (after normalization): 0.4 (assuming that the user performs a download every 40% of the time period on average);

[0071] Furthermore, the "download" behavior feature vector of this user can be: [2, 0.6, 0.8, [0.1, 0.3, 0.6], 0.4].

[0072] This encoding method takes into account both the category of user behavior and features such as the frequency and duration of the behavior, and can more comprehensively reflect the user's behavior pattern. In subsequent analysis, these feature vectors can be used to identify the user's behavior habits, preferences, and potential problems, thereby optimizing the service quality and user experience of the satellite communication network.

[0073] Data normalization: Standardize or normalize data from different sources to eliminate the influence of dimensions so that the data can be effectively compared.

[0074] Adopt min-max normalization: This method scales the data proportionally to a specified minimum and maximum value (between 0 and 1). Its formula is:

[0075]

[0076] where, x normDenote the data after standardization as [[ID=]], x represents the input data, min(x) represents the minimum value of the data, max(x) represents the maximum value of the data, and min new represents the minimum value in the target range of scaling, and max new represents the maximum value in the target range of scaling.

[0077] The present invention uses a graph structure to represent a LEO satellite mobile communication network.

[0078] In a preferred embodiment, the graph structure G of the LEO satellite mobile communication network is represented as:

[0079] G = (V, E)

[0080] Among them, V represents the set of nodes, where the nodes include satellites, ground gateway stations, and user terminals; E represents the set of edges, where each edge connects nodes, and each edge represents a communication link.

[0081] Specifically, each node v (v ∈ V) of the graph structure G has a corresponding feature vector, which includes:

[0082] For the feature vector of the satellite node, it may specifically include: orbital parameters of the satellite (such as position, speed), operating state of the satellite (such as remaining service time, operating mode), communication quality indicators (such as signal strength, bandwidth), current load, etc.

[0083] For the feature vector of the ground gateway station, it may specifically include: geographical location of the ground gateway station, device state, communication quality indicators, etc.

[0084] For the feature vector of the user terminal, it may specifically include: geographical location of the user terminal, service usage data (such as traffic consumption pattern), device type, etc.

[0085] For satellite and ground station nodes, weather data features are added, such as cloud thickness, precipitation, wind speed, etc., and these features are encoded as part of the node features after encoding.

[0086] For user terminal nodes, user behavior data features are added, such as behavior category encoding, frequency, duration, time distribution, behavior interval, etc., and these features are also encoded as part of the node features after encoding.

[0087] For the feature vector f of each edge e (e ∈ E) e , it may specifically include:

[0088] Features such as bandwidth, transmission delay, and historical packet loss rate of the link.

[0089] For some edges (especially the edges involving the ground station and the satellite), the characteristics of the impact of weather on the link can be added, such as the degree of signal attenuation, expected communication stability, etc.

[0090] In a preferred embodiment, the feature vectors of each node v and each edge e are specifically as follows:

[0091] Feature vector H of the satellite node sat :

[0092] H sat = [orbital parameters, remaining service time, communication quality index, current load, weather data characteristics];

[0093] Feature vector H of the ground gateway station node gnd :

[0094] H gnd = [geographical location, device status, communication quality index, weather data characteristics];

[0095] Feature vector H of the user terminal node user :

[0096] H user = [geographical location, service usage data, device type, user behavior data];

[0097] Feature vector E of the edge ij :

[0098]

[0099] where i represents the i-th node, i ∈ E, j represents the j-th node, j ∈ E, and i ≠ j.

[0100] Integrating and encoding different types of node data into a unified feature vector provides a consistent input format for the graph neural network. The feature vectors including weather and user behavior data enhance the richness of node information and contribute to subsequent feature aggregation and decision-making processes.

[0101] The terminal inter-satellite link handover prediction model adopted by the present invention (hereinafter can be abbreviated as "prediction model") is an improvement on the existing GCN model, and it includes multiple GCN layers, Transformer layers, LSTM layers and fully connected layers.

[0102] Figure 4 It is the training step flow chart of the terminal inter-satellite link handover prediction model in the embodiment of the present invention.

[0103] In a preferred embodiment, referring to Figure 4 as shown, the training process of the terminal inter-satellite link handover prediction model includes:

[0104] S101: Obtain the historical dataset of the LEO satellite mobile communication network, which includes historical satellite orbit and link state data, historical ground gateway data, historical meteorological data, historical user behavior data, and historical handover results.

[0105] Specifically, obtain the historical dataset of the LEO satellite mobile communication network from an open-source data platform, and use the processed historical dataset as the training data for the prediction model. In this step, the historical satellite orbit and link state data, historical ground gateway data, historical meteorological data, and historical user behavior data are respectively similar in content, data type, data structure, etc. compared with the real-time obtained satellite orbit and link state data, ground gateway data, meteorological data, and user behavior data, which will not be elaborated here.

[0106] The historical handover result is to record the actual handover target satellite and path of the user at different time steps. Encode the historical handover result into a label, and use it as the label data during the training of the prediction model to ensure that the training samples correspond one-to-one with the feature inputs (4 types of data sources).

[0107] S102: Preprocess and feature-encode the historical dataset of the LEO satellite mobile communication network to obtain the corresponding processed data.

[0108] The data preprocessing and feature encoding in this step are the same as the methods for preprocessing and feature encoding the 4 types of measured data in the above model prediction stage, which will not be elaborated here.

[0109] Although the prediction model in the embodiments of the present invention uses 4 types of related data (historical satellite orbit and link state data, historical ground gateway data, historical meteorological data, historical user behavior data), in fact, the contribution of each type of data source data to the prediction result is different, and it is necessary to analyze and process this.

[0110] S103: Calculate the support degree of each data source according to the historical dataset of the LEO satellite mobile communication network, and initialize the weight matrix W of each data source in combination with expert knowledge 0 .

[0111] Specifically, calculate the support degree of each data source and the weight of the feature through the historical data performance, feature importance analysis, and expert knowledge of the historical dataset, ensuring that important and reliable data sources have a greater influence during the feature aggregation and decision-making process.

[0112] The support degree of each data source reflects the credibility of the data source and affects the weight adjustment of subsequent feature aggregation.

[0113] The calculation formula for the support degree Support(u) is:

[0114]

[0115] Among them, correct(u, t) represents the historical performance data of each data source u, indicating the number of times that the data source u provided correct data decisions at the t-th moment, and T is the number of past decisions.

[0116] The weight matrix W is the feature weight for each feature of each data source, and is used to measure the importance of the feature for node representation learning and switching decisions. W is fixed and does not change with model training.

[0117] The initial value w of the weight matrix 0 is set according to the following factors:

[0118] Historical data performance: Use statistical methods or machine learning models to evaluate the contribution of each feature in past decisions.

[0119] Feature importance: Determine the importance of the feature through feature correlation analysis (such as Pearson correlation coefficient, information gain).

[0120] Domain expert knowledge: Combine the experience of communication domain experts to make a preliminary setting of the weights.

[0121] For example, weights are set for 4 types of data sources and their features respectively:

[0122] Satellite orbit features: remaining service time (weight 0.4), orbital altitude (weight 0.3), etc.;

[0123] Features of the ground gateway station: network load (0.3), signal reception quality (0.2), signal transmission delay (0.2), etc.

[0124] Meteorological features: cloud thickness (weight 0.25), wind speed (weight 0.15), etc.;

[0125] User behavior features: traffic pattern (weight 0.2), time distribution (weight 0.1), etc.

[0126] The support reflects the credibility of the data source, and W 0 represents the weights of each feature of the data source. Therefore, combining the support of the data source with the feature weights of the data source can comprehensively represent the comprehensive weighted feature influence of each data feature of each data source on the node.

[0127] S104: Based on the processed corresponding data and the LEO satellite mobile communication network graph structure, combine the support of each data source and the weight matrix W0, extract the feature vectors of each node and the feature vectors of each edge, and input them into the prediction model.

[0128] Specifically, the weighted fusion of feature vectors is as follows: For each node v in the graph structure G, according to its affiliated data source (such as satellite, ground gateway station, user terminal) and its features, the original feature vector Hv is first extracted.

[0129] Multiply each feature element by the corresponding data source support degree Support(u) and the weight value in the weight matrix W 0 to obtain the weighted feature vector H' v , and its calculation formula is:

[0130]

[0131] where represents the element-wise multiplication operation.

[0132] Edge feature vector processing: For each edge e(i,j) in the graph structure G, extract its original feature vector F e . Similar to the node feature processing, multiply the edge feature vector F e by the corresponding data source support degree Support(u) and the weight value in the weight matrix W 0 to obtain the weighted edge feature vector F' e , and its calculation formula is:

[0133]

[0134] Perform normalization processing on the weighted node feature vector H' v and the edge feature vector F' e to ensure that the numerical ranges between different feature dimensions are consistent. Input the normalized node feature vector H' v and the edge feature vector F' e into the prediction model: Input the processed node feature vector and edge feature vector into the terminal inter-satellite link handover prediction model for training according to the connection relationship of the graph structure G.

[0135] Ensure that the batch and order of the input data are consistent with the model training stage to avoid data distribution differences between the training and prediction stages. Combine the support degree of multi-source heterogeneous data and the weight matrix W 0 , perform weighted fusion and normalization processing on the feature vectors of nodes and edges to ensure that the influence of different data sources in the feature extraction process is reasonably reflected

[0136] S105: Use multiple GCN layers to aggregate the features of neighbor nodes to the target node through multiple layers of vectors and extract the high-order feature vectors of the nodes.

[0137] Specifically, feature aggregation synthesizes the features of neighbor nodes into the target node through a graph convolutional network (GCN). The node feature vector is gradually enriched by aggregating the features of neighbor nodes through each layer of the GCN, enabling the capture of higher-order graph structure information.

[0138] And the support degree and weight influence of the above data sources are added. It more accurately reflects the importance of each data source and its features in the decision-making process and fully utilizes the link information.

[0139] The formula for multi-layer neighbor feature aggregation is as follows:

[0140]

[0141] In the formula, represents the feature representation of any node v at the l + 1 layer, which combines the features of neighbor node u. This feature vector has undergone feature aggregation through multiple layers of GCN and is the high-order feature vector of node v, which includes rich local and global graph information; l represents the number of network layers, and v represents the node. ReLU(.) represents the activation function used to ensure the non-linearity of the output features; each layer of GCN further processes and integrates the features through a non-linear activation function (such as ReLU), capturing more complex relationships and high-order dependencies between nodes layer by layer while avoiding the vanishing gradient problem. represents the feature vector of node u at the l layer, N(v) is the set of neighbor nodes of node v, and W uv represents the weight between node v and its neighbor node u. Support(u) is the support degree of neighbor node u, and Bias(v) represents the bias term of node v.

[0142] By aggregating the features of neighbor nodes through multi-layer weighted aggregation and combining the support degree and weight matrix, the feature representation of the target node is updated, which includes rich local and global graph information.

[0143] S106: Input the high-order feature vector of the node into the Transformer for position encoding and time information enhancement to extract the temporal feature vector of the node.

[0144] Specifically, for time series data, position encoding is introduced to enhance the model's perception ability of time information. The same sine and cosine function encoding method as that of the Transformer can be adopted. It uses sine and cosine functions to generate position encoding, so that the model can understand the order of elements in the sequence. The formula is as follows:

[0145]

[0146]

[0147] Among them, PE (pos,2i)Denote the encoded value at position pos in dimension 2i, PE (pos,2i+1) Denote the encoded value at position pos in dimension 2i + 1, where pos represents the position index in the sequence (starting from 0), i represents the index of the dimension, sin(.) represents the sine function, cos(.) represents the cosine function, and d model Denote the dimension of the model.

[0148] Add the encoded feature time series feature PE (pos) to the feature vector of each node, which is specifically represented as:

[0149] For the node features at each time step, through calculate the similarity of neighbor nodes at the current time step, generate the temporal attention weight a i,j , highlighting the features at important time steps and enhancing the model's attention to key time information.

[0150] The temporal attention weight a i,j is calculated as follows:

[0151]

[0152] Among them, score(H(vi),H(vj)) uses similarity measurement methods such as dot product or cosine similarity.

[0153] Then perform weighted feature aggregation: Use the temporal attention weight to perform weighted aggregation on the features of neighbor nodes:

[0154]

[0155] In the formula, represents the temporal feature vector of the node.

[0156] S107: Input the temporal feature vector of the node into the LSTM to obtain the node features after temporal enhancement.

[0157] Specifically, input the temporal feature vector of the node into the LSTM to capture the dynamics of the time series, which is represented as:

[0158]

[0159] Among them, k is the time window size of the temporal model.

[0160] The overall formula for temporal enhancement and weighted feature aggregation is:

[0161]

[0162] Among them, Represents the node feature vector after temporal enhancement.

[0163] S108: Input the node features after temporal enhancement and the feature vectors of each edge into a fully connected layer, and calculate the comprehensive score S of each satellite node sat , select the satellite node i* with the highest comprehensive score for switching to obtain the predicted switching result.

[0164] Specifically, calculate the feature scores of each data source, and then obtain the candidate satellite features to be switched, providing a quantitative basis for the final decision, and ensuring a reasonable balance of the importance and reliability of different data sources during the decision-making process.

[0165] The feature score S of each data source i The calculation formula is:

[0166]

[0167] Among them, Fcature ij represents the j-th feature in data source i, and W ij represents the weight of feature j in data source i. After obtaining the feature weighted scores of each data source, the support scores S of all data data sources of each candidate satellite can be obtained sat , S sat This is also the comprehensive score of each candidate satellite node, and its calculation formula is:

[0168]

[0169] In the formula, S sat represents the comprehensive score of each candidate satellite node, which is the total score after accumulating the weighted scores of the features of each feature source of each candidate satellite node. Support(i) represents the support degree (i.e., credibility) of data source i, and TimeAdjust(i) represents the temporal adjustment factor of data source i, which is adjusted according to the learning definition in the temporal enhancement step and reflects its adaptability to the current time step.

[0170] Figure 5 This is the flow diagram of satellite handover decision-making in the embodiment of the present invention.

[0171] Refer to Figure 5 As shown, when the user terminal initiates a handover request for the current serving satellite, 4 types of relevant data are obtained in real time. After data preprocessing, feature encoding, etc., the comprehensive score S of each candidate satellite node is calculated through the prediction model sat .

[0172] When calculating the feature score S for each candidate satellite node satWhen considering the impact of the ground gateway station as a node and the edges between nodes along with weights on the comprehensive score of candidate satellite features, only the candidate satellite with the highest feature score needs to be considered as the optimal handover satellite at this time.

[0173] In a preferred embodiment, referring to Figure 5 as shown, the comprehensive score S of the satellite node sat needs to meet the satellite handover conditions, specifically:

[0174] When the comprehensive score S of the satellite node sat is greater than the satellite handover threshold Threshold sat (i.e., the candidate satellite meets the handover conditions), a set of satellite nodes that meet the handover conditions is filtered out:

[0175]

[0176] where C represents the set of candidate satellite nodes that meet the satellite handover conditions, and i represents the serial number of the satellite node;

[0177] When C≠0, the satellite node with the highest comprehensive score is selected:

[0178]

[0179] At the same time, a path selection threshold Threshold of the ground gateway station is set for the node data source of the ground gateway station gnd , when the feature score S provided by the ground gateway station data in the candidate satellite comprehensive score S sat ≥Threshold sat(gnd) gnd then the handover communication path requires the ground gateway station to perform data transfer (i.e., information is transferred through the ground gateway station to complete the communication link handover); when S sat(gnd) gnd <Threshold gnd then the handover communication path is directly completed through the handover satellite (i.e., information is directly transferred between satellites to complete the communication link handover).

[0180] Since when calculating the feature score S sat for each candidate satellite node, the impact of the ground gateway station as a node and the edges between nodes along with weights on the comprehensive score of candidate satellite features has been considered, therefore, the feature score S sat provided by the ground gateway station data in the candidate satellite comprehensive score S sat(gnd) can be calculated.

[0181] Finally, the satellite handover decision Decision is expressed as:

[0182]

[0183] S109: Calculate the loss between the predicted handover result and the historical handover result, and through this loss, perform backpropagation to optimize the terminal inter-satellite link handover prediction model.

[0184] In a preferred embodiment, after the steps S1 - S3, the feature vectors of each node v in the graph structure G and the feature vectors of each edge are obtained, and they are input into the trained terminal inter-satellite link handover prediction model. The prediction process includes:

[0185] Utilize multiple layers of GCN layers to aggregate the features of neighbor nodes to the target node through multiple layers of vector aggregation, and extract the high-order feature vectors of the nodes;

[0186] Input the high-order feature vectors of the nodes into the Transformer for position encoding and time information enhancement, and extract the temporal feature vectors of the nodes;

[0187] Input the temporal feature vectors of the nodes into the LSTM to obtain the node features after temporal enhancement;

[0188] Input the node features after temporal enhancement into the fully connected layer to calculate the support scores S of all data data sources of each satellite node sat , and use it as the comprehensive score S of each satellite node sat , when the comprehensive score S of the satellite node sat meets the satellite handover condition, then the optimal terminal inter-satellite link handover strategy is obtained.

[0189] Since the prediction process of the terminal inter-satellite link handover prediction of the present invention has the same or similar steps as the training process, only the input data is different. Therefore, it will not be elaborated here.

[0190] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium can include: ROM, RAM, disk, or optical disc, etc.

[0191] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A terminal inter-satellite link switching method based on satellite-ground heterogeneous data, characterized in that: include: According to LEO satellites, user terminals, ground gateways and their communication links, a LEO satellite mobile communication network and its graph structure are constructed; Respond to requests from user terminals, obtain satellite orbit and link status data, ground gateway data, meteorological data, and user behavior data in real time, perform data preprocessing and feature encoding on them, and obtain the processed corresponding data; Extracting a feature vector of each node and a feature vector of each edge based on the processed corresponding data and the graph structure G of the LEO satellite mobile communication network; The feature vector of each node and the feature vector of each edge are input into the trained terminal intersatellite link switching prediction model to obtain the optimal terminal intersatellite link switching strategy.

2. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1 is characterized in that: The training process of the terminal intersatellite link switching prediction model includes: Obtain historical data sets of the LEO satellite mobile communication network, including historical satellite orbit and link status data, historical ground gateway data, historical meteorological data, historical user behavior data, and historical switching results; Preprocessing and feature encoding the historical data set of the LEO satellite mobile communication network to obtain processed corresponding data; The support of each data source is calculated based on the historical data set of the LEO satellite mobile communication network, and the weight matrix W0 of each data source is initialized based on expert knowledge; Based on the processed data and the LEO satellite mobile communication network graph structure, combined with the support and weight matrix W0 of each data source, extract the feature vector of each node and the feature vector of each edge, and input them into the prediction model; Using multiple GCN layers, the features of neighbor nodes are aggregated to the target node through multiple layers of vectors, and the high-order feature vector of the node is extracted; The high-order feature vector of the node is input into the Transformer to perform position encoding and time information enhancement, and extract the time series feature vector of the node; Inputting the time series feature vector of the node into LSTM to obtain the node feature after time series enhancement; The node features after time series enhancement and the feature vector of each edge are input into the fully connected layer to calculate the comprehensive score S of each satellite node sat , select the satellite node i* with the highest comprehensive score for switching, and get the predicted switching result; The loss between the predicted switching result and the historical switching result is calculated, and the terminal intersatellite link switching prediction model is optimized through back propagation of the loss.

3. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1, characterized in that: The feature vector of each node v and the feature vector of each edge are input into the trained terminal intersatellite link switching prediction model, and the prediction process includes: Aggregate the features of neighbor nodes to the target node through GCN to obtain the updated node feature vector; Input the updated node feature vector into the multi-layer GCN to extract the high-order feature vector of the node; The high-order feature vector of the node is input into the Transformer to perform position encoding and time information enhancement, and extract the time series feature vector of the node; Input the node's time series feature vector into LSTM to obtain the node feature after time series enhancement; The node features after time series enhancement are input into the fully connected layer to calculate the comprehensive score S of each satellite node sat , when the comprehensive score S of the satellite node sat When the satellite switching conditions are met, the optimal terminal intersatellite link switching strategy is obtained.

4. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 3 is characterized in that: The comprehensive score S of the satellite node sat The satellite switching conditions that need to be met are as follows: When the comprehensive score of the satellite node S sat Greater than the satellite switching threshold Threshold sat When , the satellite node i* with the highest comprehensive score is selected, which is specifically expressed as: C={i|S sat ( i) >Threshold sat }, Wherein, C represents the set of candidate satellite nodes that meet the satellite switching conditions, and i represents the sequence number of the satellite node; When C≠0, select the satellite node i* with the highest comprehensive score, which is expressed as:

5. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1, characterized in that: The satellite orbit and link status data include the satellite's remaining service time, orbital altitude, position coordinates, speed, and orbital inclination; the ground gateway data include satellite signal reception quality, satellite signal strength, and satellite signal arrival time; the meteorological data include cloud coverage, precipitation, wind speed, temperature, humidity, and air pressure; the user behavior data include user behavior category, user behavior frequency, user behavior duration, and average interval of user behavior.

6. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1, characterized in that: The specific process of data preprocessing and feature encoding includes: data cleaning, data verification, feature encoding and normalization processing.

7. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1, characterized in that: The graph structure G of the LEO satellite mobile communication network is expressed as: G=(V,E) Among them, V represents the node set, where nodes include satellites, ground gateways and user terminals; E represents the edge set, where each edge connects nodes and each edge represents a communication link.

8. The terminal inter-satellite link switching method based on satellite-ground heterogeneous data according to claim 1, characterized in that: The eigenvector of each node v and the eigenvector of each edge are specifically: The characteristic vector H of the satellite node sat : H sat =[orbital parameters, remaining service time, communication quality indicators, current load, weather data characteristics]; The characteristic vector H of the ground gateway node gnd : H gnd =[geographic location, device status, communication quality indicators, weather data characteristics]; The feature vector H of the user terminal node user : H user =[geographic location, service usage data, device type, user behavior data]; The eigenvector E of the edge ij : E ij =[Bandwidth ij , transmission delay ij , Historical packet loss rate ij , Signal attenuation ij , Expected communication stability ij ]Wherein, i represents the i-th node, i∈E, j represents the j-th node, j∈E, and i≠j.

Images