A method and device for hierarchical multi-domain routing based on long-short chain cooperation of giant constellations

Abstract

The application discloses a kind of giant constellation hierarchical multi-domain routing method and device based on long-short chain cooperation, it is related to satellite communication field, the method comprises: constructing satellite network hierarchical multi-domain architecture, including: GEO satellite layer, MEO satellite layer and LEO satellite layer;LEO satellite clustering algorithm is used to dynamically cluster LEO satellite network, form multiple distributed clouds;Construct satellite network dynamic topology graph;Based on satellite network dynamic topology graph, using discrete time virtual topology routing algorithm, obtain long-chain global optimal path;For multiple distributed clouds, using deep reinforcement learning model determines short-chain local optimal path;If short-chain local optimal path and corresponding part in long-chain global optimal path are consistent, then long-chain global optimal path is used, if it is not consistent, then short-chain local optimal path replaces corresponding part in long-chain global optimal path, forms actual routing path, the application can solve the conflict problem between high dynamicity and multi-objective of satellite-ground network.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite communication, in particular to a long-short chain cooperative-based mega-constellation hierarchical multi-domain routing method and device. BACKGROUND

[0002] With the gradual maturity of Low Earth Orbit (LEO) satellite constellation technology represented by the American Starlink plan, a large number of LEO satellite clusters will build a new generation of Internet infrastructure with global seamless coverage, high-bandwidth communication, and low transmission delay, providing efficient and continuous network access services for global users. Due to the deployment scale of future LEO satellite networks expected to exceed the level of thousands, and the types of network services it carries will be more diversified, in such a large-scale heterogeneous multi-service satellite network, the LEO satellite routing method needs to ensure the real-time and reliability of multi-service data transmission through efficient topology management and path planning mechanism, therefore, the LEO satellite routing method has become one of the key technologies of LEO satellite network research and has attracted widespread attention.

[0003] However, due to the dynamic changes of the topology structure of LEO satellite communication network, the time-varying of inter-satellite link state, the uneven distribution of network traffic, the short orbit running period and other problems, these problems bring great challenges to the design of LEO satellite routing method. Dynamic topology changes make it impossible for traditional static or semi-static routing methods to capture the changes of network structure in real time, resulting in delayed or failed routing decisions; the time-varying of inter-satellite link state increases the difficulty of link prediction and stability evaluation, making it difficult to guarantee the reliability of routing selection; the unevenness of network traffic distribution may cause overload of some nodes or links, affecting the overall performance and resource utilization efficiency of the network. In view of the differentiated performance requirements of satellite-ground network, due to the differences in topology structure, channel and transmission resources between satellite network segment and ground network segment, as well as the long transmission distance and large delay of satellite-ground link, the satellite-ground network routing strategy is easy to conflict and lack of efficient coordination mechanism, which will lead to cross-domain path shock and resource competition, which will lead to the decline of global routing efficiency and transmission resource utilization. In addition, the coupling effect of high dynamic topology and multi-objective conflict will also lead to a significant decline in satellite-ground network routing efficiency and resource utilization, which requires a hierarchical cooperative optimization mechanism to balance real-time and global optimality, breaking through the limitations of traditional single-objective optimization. At the same time, the high dynamic nature of satellite-ground link requires the routing protocol to quickly complete topology update and path re-planning, and traditional centralized routing algorithms are difficult to meet the real-time requirements due to high computational complexity and large cross-domain communication overhead. SUMMARY

[0004] The purpose of the present application is to provide a long-short chain cooperative-based mega-constellation hierarchical multi-domain routing method and device, which can solve the conflict between high dynamic nature and multi-objective of satellite-ground network.

[0005] To achieve the above object, the application provides the following scheme:

[0006] In a first aspect, the application provides a giant constellation hierarchical multi-domain routing method based on long-short chain cooperation, comprising:

[0007] constructing a satellite network hierarchical multi-domain architecture; the satellite network hierarchical multi-domain architecture comprises a GEO satellite layer, a MEO satellite layer and a LEO satellite layer; the GEO satellite layer is deployed with an SDN controller for maintaining a global view and calculating and generating a long-chain global optimal path;

[0008] adopting a LEO satellite clustering algorithm to dynamically cluster the LEO satellite network to form multiple distributed clouds;

[0009] constructing a satellite network dynamic topology graph;

[0010] adopting a discrete-time virtual topology routing (DT-DVTR) algorithm based on the satellite network dynamic topology graph to obtain a long-chain global optimal path;

[0011] for the multiple distributed clouds, adopting a deep reinforcement learning model to determine a short-chain local optimal path;

[0012] comparing the long-chain global optimal path with the short-chain local optimal path, if the corresponding parts of the short-chain local optimal path and the long-chain global optimal path are consistent, the long-chain global optimal path is adopted, if not, the short-chain local optimal path replaces the corresponding part of the long-chain global optimal path to form an actual routing path;

[0013] the short chain collects satellite node dynamic data in real time and sends it to the long chain; the dynamic data includes link state, queue depth and energy reserve;

[0014] the long chain optimizes the global optimal path based on the dynamic data to form a short-chain perception-long-chain optimization-strategy counterproductive closed-loop cooperation mechanism.

[0015] Optionally, the MEO satellite layer is deployed with a lightweight SDN agent; the LEO satellite layer realizes intra-cluster autonomy through a clustering algorithm.

[0016] Optionally, the adoption of the LEO satellite clustering algorithm to dynamically cluster the LEO satellite network to form multiple distributed clouds comprises two stages, specifically: a cluster initialization stage and a cluster dynamic maintenance stage, the cluster initialization stage specifically comprises the following steps:

[0017] S1: Obtain ephemeris data of LEO satellites, and divide the entire running period into multiple time slices;

[0018] S2: calculating average connectivity of the satellite node based on the ephemeris data and the number of time slices;

[0019] S3: judging whether there is a neighbor node of the satellite node based on the average connectivity, and if not, marking as a free node;

[0020] S4: if there is a neighbor node, marking as an undetermined node, and checking whether there is a determined cluster head node within its communication range;

[0021] S5: if there is a cluster head node, calculating a priority;

[0022] S6: selecting an optimal cluster head based on the priority;

[0023] S7: adding the undetermined node in step S4 to the optimal cluster head, and updating the undetermined node as a cluster member node;

[0024] S8: if there is no cluster head node, comparing the average connectivity of the undetermined node itself with its neighbor nodes in step S4;

[0025] S9: selecting a satellite node with locally maximum average connectivity, upgrading the satellite node as a cluster head node and creating a new cluster; the new cluster follows a single-cluster single-cluster head node constraint condition, and an initial cluster size is set as 1, and if the condition is not met, resetting the node state to enter the next round of determination;

[0026] S10: repeating steps S1-S9 until all node states in the LEO satellite network are determined.

[0027] Optionally, the dynamic maintenance phase of the cluster specifically includes: cluster head node state updating, cluster member node state updating, free node state updating, and undetermined node state updating;

[0028] The cluster head node state updating specifically includes: when the number of cluster members of the cluster head node is 1 and exceeds a preset time limit, triggering state degradation, converting the cluster head node into an undetermined node, and broadcasting a state change message to a neighboring area to trigger a cluster reorganization process;

[0029] The cluster member node state updating: when the cluster member node detects that the hop number between it and the cluster head node exceeds a preset threshold, triggering a state updating process, the cluster member node first initiates a cluster leaving request to the current cluster head node, judges whether there is a neighbor node within a detection range, if not, directly converting into a free state; if there is a neighbor node, degrading into an undetermined state and participating in a new round of dynamic clustering process again;

[0030] The free node state updating: when the free node finds that there is a connection between itself and other nodes, converting into an undetermined node and participating in a new round of dynamic clustering process again;

[0031] The undetermined node state is updated: the undetermined node selects to join the optimal cluster head as a cluster member node or the connectivity of the undetermined node reaches a preset condition to upgrade to a cluster head node.

[0032] Optionally, the constructing the dynamic topology graph of the satellite network specifically comprises the following steps:

[0033] The satellite ephemeris data is preprocessed to obtain a standardized satellite coordinate matrix and a fully connected link state tensor;

[0034] The standardized satellite coordinate matrix and the fully connected link state tensor are converted into a graph structure; each satellite node in the graph structure is assigned a feature vector, and the feature vector is formed by splicing the standardized satellite coordinate matrix and the satellite link state tensor to form a node feature matrix;

[0035] Based on the graph structure, the Euclidean distance between the satellite nodes is calculated and combined with the geometric visibility condition to generate an edge weight matrix; wherein the weight is defined as the product of the exponential decay function of the distance and the visibility indicator;

[0036] Based on the weight matrix, a space-time graph sequence is output;

[0037] Taking the space-time graph sequence as input, a space-time graph convolution network ST-GCN is used to extract a space-time feature vector of the satellite dynamic topology;

[0038] Based on the space-time feature vector, an inter-satellite connection probability at a future time is predicted to generate a global topology probability matrix;

[0039] The global topology probability matrix is processed to generate a stable topology, i.e., a dynamic topology graph of the satellite network; the dynamic topology graph of the satellite network includes: link survival probability, node position, load, and energy consumption state.

[0040] Optionally, based on the dynamic topology graph of the satellite network, a discrete-time virtual topology routing DT-DVTR algorithm is used to obtain a long-chain global optimal path, specifically comprising the following steps:

[0041] Filtering feasible paths that meet the link survival, node energy, and load upper limits;

[0042] For each of the feasible paths, a comprehensive utility function is calculated:

[0043]

[0044] Wherein, D p , B p , L p represent the current delay, bandwidth, and load value of the path p, D p,min , Bp,max , L p,max respectively represent the theoretical minimum latency, the theoretical maximum bandwidth and the maximum load threshold corresponding to the path p, and α, β, γ represent weight factors respectively, and the sum equals 1;

[0045] The long-chain global optimal path is determined based on the comprehensive utility function of each feasible path by using the Dijkstra algorithm.

[0046] Optionally, the long-short chain collaborative hierarchical mega-constellation multi-domain routing method further comprises, after the step of "long-chain based on the dynamic data optimization global optimal path, forming a short-chain perception-long-chain optimization-strategy counterproductive closed-loop collaborative mechanism":

[0047] acquiring a typical service type; the typical service type includes: a latency-sensitive service, a computation-intensive service and a bandwidth-intensive service;

[0048] determining a service priority based on the service type;

[0049] adjusting the weight factors α, β, γ in the comprehensive utility function in real time according to the service priority, so as to maximize the comprehensive utility function;

[0050] selecting the best path according to the adjusted comprehensive utility function.

[0051] In a second aspect, the application provides a long-short chain collaborative hierarchical mega-constellation multi-domain routing device, which comprises:

[0052] a satellite network hierarchical multi-domain architecture construction module, configured to construct a satellite network hierarchical multi-domain architecture; the satellite network hierarchical multi-domain architecture comprises a GEO satellite layer, a MEO satellite layer and a LEO satellite layer; the GEO satellite layer is deployed with an SDN controller, configured to maintain a global view and calculate and generate a long-chain global optimal path;

[0053] a dynamic clustering module, configured to dynamically cluster the LEO satellite network by using a LEO satellite clustering algorithm, to form a plurality of distributed clouds;

[0054] a satellite network dynamic topology graph construction module, configured to construct a satellite network dynamic topology graph;

[0055] a long-chain global optimal path determination module, configured to obtain a long-chain global optimal path by using a discrete-time virtual topology routing (DT-DVTR) algorithm based on the satellite network dynamic topology graph;

[0056] a short-chain local optimal path generation module, configured to determine a short-chain local optimal path for the plurality of distributed clouds by using a deep reinforcement learning model;

[0057] An actual routing path generation module is configured to compare the long-chain global optimal path and the short-chain local optimal path, and if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are consistent, the long-chain global optimal path is adopted, and if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are inconsistent, the short-chain local optimal path is adopted to replace the corresponding parts of the long-chain global optimal path to form an actual routing path.

[0058] A dynamic data acquisition module is configured to acquire satellite node dynamic data in real time in a short chain and send the dynamic data to a long chain, wherein the dynamic data includes link state, queue depth and energy reserve.

[0059] A path optimization module is configured to optimize a global optimal path based on the dynamic data in the long chain to form a short-chain perception-long-chain optimization-strategy feedback closed-loop collaborative mechanism.

[0060] According to the specific embodiments provided in the application, the application has the following technical effects:

[0061] The application provides a long-short chain collaborative based mega-constellation hierarchical multi-domain routing method and device, constructs a distributed cloud based on distributed computing power of LEO satellites, proposes a hierarchical reinforcement learning framework, solves the conflict problem between high dynamics and multiple targets of a satellite-ground network through a "long-chain global planning-short-chain real-time adjustment" collaborative mechanism, wherein the long chain is based on a space-time graph convolution network ST-GCN and an attention mechanism to predict a network topology in real time and generate a Pareto optimal path set, i.e., a global optimal path, which mainly optimizes three targets of delay, energy consumption and load balancing; the short chain relies on a distributed cloud to deploy a lightweight deep reinforcement learning DRL model, adopts a federated reinforcement learning framework, enables distributed cloud satellite nodes to collaboratively share model parameters, jointly trains a distributed cloud global model, generates a local optimal path, and realizes hierarchical multi-domain collaborative routing of a satellite-ground network.

[0062] Based on a service type priority classification and a distributed weight dynamic adjustment mechanism, a dynamic service weight distribution strategy is designed to adjust the priority weight of services in real time, accurately matches the QoS demand of services and the supply of satellite network resources, and theoretically analyzes and proves the performance boundary of delay, throughput and bandwidth resource occupation of different service types, thereby providing theoretical support for constructing a high-reliability and self-adaptive space-ground integrated network for a large-scale LEO constellation. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0064] Figure 1 A flowchart of a long-short chain cooperative super-constellation hierarchical multi-domain routing method according to an embodiment of the present application is provided.

[0065] Figure 2 A schematic diagram of a large-scale satellite network hierarchical multi-domain architecture according to an embodiment of the present application is provided.

[0066] Figure 3 A long-chain global planning flowchart according to an embodiment of the present application is provided.

[0067] Figure 4 A short-chain local planning flowchart according to an embodiment of the present application is provided.

[0068] Figure 5 A schematic diagram of a long-short chain cooperative super-constellation hierarchical multi-domain routing device structure according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0070] The above purposes, features and advantages of the present application will be more apparent and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0071] In an exemplary embodiment, as shown in Figure 1 A long-short chain cooperative super-constellation hierarchical multi-domain routing method is provided, including the following steps 201 to 208. Among them:

[0072] Step 201: Construct a satellite network hierarchical multi-domain architecture; the satellite network hierarchical multi-domain architecture includes a GEO satellite layer, a MEO satellite layer and a LEO satellite layer. The specific architecture is as shown in Figure 2

[0073] Among them, the GEO satellite layer (centralized control): deploy a software defined network (SDN) controller, which can maintain a global view on a long time scale, including global topology, resource state, threat intelligence, etc. This layer is mainly composed of 3-5 geosynchronous orbit satellites, and the main functions are centralized control of global topology management, resource scheduling, security policy distribution, and synchronization of global state (bandwidth, delay, load) based on the SDN controller, with an update period of hours.

[0074] ​MEO satellite layer (auxiliary management): responsible for regional management and relay, each MEO satellite covers multiple LEO satellite clusters, the coverage radius is ≤5000km, runs a lightweight SDN agent, receives GEO satellite instructions and distributes them to LEO satellite clusters. This layer mainly consists of 20-30 medium earth orbit satellites, the main functions are regional relay and auxiliary management, receiving GEO satellite instructions, coordinating communication between LEO satellite clusters, providing cross-domain path relay in the case of LEO satellite failure, deploying lightweight SDN agents, running LEO satellite clustering algorithm for real-time clustering of LEO satellites, local topology update period is minutes.

[0075] LEO satellite layer (clustered autonomy): dynamically clustered based on the proposed LEO satellite clustering algorithm, the number of nodes in a cluster is ≤50, the cluster head node is responsible for local routing decision, inter-cluster communication is relayed through MEO satellites or directly uses inter-satellite link. This layer mainly consists of thousands of low earth orbit satellites, the number of satellites in a single cluster is ≤50. The main functions are dynamic intra-cluster routing, inter-cluster routing, edge computing, short chain real-time optimization adjustment, etc. Each satellite node is equipped with a lightweight DRL model, the topology awareness period is seconds.

[0076] Among them, the whole layered multi-domain architecture is designed based on a multi-layer graph model G=(G GEO ,G MEO ,G LEO ,E inter ), G GEO is a high orbit satellite dynamic weighted graph, G MEO is a medium orbit satellite dynamic weighted graph, G LEO is a low orbit satellite dynamic weighted graph, and E inter is an inter-layer connection edge.

[0077] The i-th layer network is modeled as a dynamic weighted graph G i (t) = (V i (t), E i (t), W i (t)), where V i (t) is a set of satellite nodes, E i (t) is a set of satellite links, and W i (t) is a weight, which specifically includes delay, bandwidth, credibility and other parameters. In addition, inter-layer cooperation needs to consider the weight of cross-layer edges, which is modeled as W inter (t) = w1·T delay (t) + w2·C cost (t), where T delay (t) is the delay, C cost (t) is the energy consumption, w1 and w2 are dynamic weight coefficients of the task, and the sum is equal to 1.

[0078] The hierarchical multi-domain architecture realizes efficient cooperation of star-ground integration by decoupling hierarchical functions and dynamic allocation of cross-layer weights, and is suitable for heterogeneous multi-layer mega-constellation systems.

[0079] Step 202: A LEO satellite clustering algorithm is used to dynamically cluster the LEO satellite network to form multiple distributed clouds.

[0080] After the construction of the hierarchical multi-domain architecture of the mega-constellation, a LEO satellite clustering algorithm is used to construct multiple distributed clouds, and the LEO satellite network is clustered based on node state to cover the entire LEO satellite network with the least number of clusters, thereby improving the robustness of regional autonomy, task decomposition and routing optimization of the LEO satellite network under dynamic topology.

[0081] The clustering algorithm mainly includes two stages: cluster initialization stage and cluster maintenance stage. The main task of the initialization stage is to divide the state of all LEO satellite nodes at the initial time and designate appropriate nodes as cluster head nodes. The cluster maintenance stage is responsible for maintaining the cluster structure when the cluster network topology changes to ensure that all nodes are in the cluster structure.

[0082] The initialization stage is mainly completed by the ground station. After the ground station calculates the weight of each node based on the cluster information, the node state is divided.

[0083] After the initialization stage is completed, the LEO satellite cluster network immediately enters the cluster maintenance stage, and each node changes its state through certain information interaction.

[0084] The node state change mainly includes four types: cluster head node state change, cluster member node state change, free node state change, and undecided node state change.

[0085] According to the system requirements, the LEO satellite cluster is divided into multiple sets. Ideally, the entire network is covered with the least number of cluster heads. In order to reduce the additional control overhead caused by the clustering algorithm and improve the actual utilization rate of network capacity, the clustering algorithm should ensure a low algorithm complexity as much as possible.

[0086] The specific clustering algorithm process is as follows:

[0087] Initialization stage: The core task of the initialization stage is to complete the node state division and elect cluster head CH nodes at the initial time of LEO satellite network startup. This stage is executed by the GS and includes the following steps:

[0088] S1: Obtain the ephemeris data of the LEO satellite, and divide the entire running period into multiple time slices.

[0089] S2: Calculate the average connectivity of the satellite nodes based on the ephemeris data and the number of time slices.

[0090] GS first calculates the average connectivity of each satellite based on LEO satellite constellation information, and then divides the entire running period into multiple time slices, and assumes that the network topology remains stable within each time slice. The average connectivity is defined as follows:

[0091]

[0092] wherein, is the average connectivity of satellite node v i , N T is the number of divided time slices, θ i (t) represents the number of neighbor nodes of node v i within the tth time slice, ι ij =1 indicates that node v i is connected to node v j within the tth time slice, and vice versa, if ι ij =0, it indicates that node v i is not connected to node v j within the tth time slice.

[0093] The algorithm divides the LEO satellite node state into four types, namely CH node, ordinary CM node, undetermined node and free node. After the GS calculates the average connectivity of each node, the state division will be executed in stages.

[0094] S3: Based on the average connectivity, it is judged whether there is a neighbor node of the satellite node. If there is no neighbor node, it is marked as a free node.

[0095] S4: If there is a neighbor node, it is marked as an undetermined node, and it is checked whether there is a determined cluster head node within its communication range.

[0096] That is, it is first determined whether there is a neighbor node of the satellite node. If there is no direct neighbor node, it is classified as a free node. If it has a connection relationship, it is marked as an undetermined node, and it is searched whether there is a determined cluster head node within its hop communication range.

[0097] S5: If there is a cluster head node, the priority is calculated.

[0098] Specifically, if there is a CH node, the priority is calculated based on the following formula to select the optimal cluster head, and the node is included in the corresponding cluster structure and updated to a CM node. The priority calculation is as follows:

[0099]

[0100] wherein, represents the priority measurement of node v i joining the cluster head v j , denotes the average connectivity degree of cluster head node v j , k denotes the number of nodes in the cluster; denotes the hop number of node v i from cluster head node v j .

[0101] S6: Select the optimal cluster head based on the priority.

[0102] S7: Add the undetermined node in step S4 to the optimal cluster head, and update the undetermined node as a cluster member node.

[0103] S8: If there is no cluster head node, compare the average connectivity degree of the undetermined node itself with that of its neighbor nodes in step S4.

[0104] S9: Select the satellite node with the locally maximum average connectivity degree, upgrade the satellite node to a cluster head node, and create a new cluster; the new cluster follows the single cluster head node constraint condition, and the initial cluster size is set to 1, and if the condition is not met, the node state is reset to enter the next round of determination.

[0105] That is, if no CH node is detected, the average connectivity degree of the undetermined node itself is compared with that of its neighbor nodes, and when the average connectivity degree of the node is locally maximum, the node is upgraded to a CH node and a new cluster is created, the cluster follows the single cluster single CH node constraint condition, and the initial cluster size is set to 1, and if the condition is not met, the node state is reset to enter the next round of determination.

[0106] S10: Repeat steps S1-S9 until the state of all nodes in the LEO satellite network is determined.

[0107] The entire initialization process adopts a multi-round iteration mechanism until the state of all nodes in the network is determined. After the initialization phase ends, the LEO satellite clustering network enters the dynamic cluster maintenance phase.

[0108] Dynamic maintenance phase: The cluster maintenance phase realizes state updating and structure optimization through periodic information interaction of satellite nodes. The main task of this phase is to monitor and respond to the dynamic changes of network topology in real time. When the node position or connection relationship changes, the system will adaptively adjust the cluster structure according to the latest topology information, aiming to ensure that all satellite nodes are effectively included in the cluster structure, thereby maintaining the continuity and reliability of network communication.

[0109] Among them, the node state update mainly includes CH node state update, CM node state update, free node state update and undetermined node state update, and the specific maintenance strategy is as follows:

[0110] The cluster head node state updating specifically includes: when the number of cluster members of the cluster head node is 1 and exceeds a preset time limit, triggering state degradation, converting the cluster head node into an undetermined node, and broadcasting a state change message to a neighboring area to trigger a cluster reorganization process.

[0111] Specifically, generally, the node state of the CH has high stability, and only when the number of nodes in the cluster is 1 and exceeds a certain time limit, the CH node triggers a state degradation process. In this scenario, the CH node is converted into an undetermined node, and a state change message is broadcast to the neighboring area to trigger a cluster reorganization process.

[0112] The cluster member node state updating specifically includes: when the cluster member node detects that the hop number between itself and the cluster head node exceeds a preset threshold, triggering a state updating process. The cluster member node first initiates a cluster leaving request to the current cluster head node, and then evaluates the neighborhood condition to determine whether there is an available node in the detection range, that is, whether there is a neighbor node in the detection range. If there is no neighbor node, the cluster member node is directly converted into a free state; otherwise, if there is a neighbor node, the cluster member node is degraded into an undetermined state and participates in a new round of dynamic clustering process.

[0113] The free node state updating: when the free node finds that there is a connection between itself and other nodes, the free node is converted into an undetermined node and participates in a new round of dynamic clustering process.

[0114] Specifically, the free node actively monitors the connection state with other nodes through periodic link detection. When the free node finds that there is a connection between itself and other nodes, the free node is converted into an undetermined node, thereby participating in the cluster construction process again.

[0115] The undetermined node state updating: when the undetermined node chooses to join an optimal cluster head as a cluster member node or the connection degree of the undetermined node reaches a preset condition, the undetermined node is upgraded to a cluster head node.

[0116] Specifically, the undetermined node is a temporary state and needs to make a state decision in the shortest time. The undetermined node chooses to join an optimal cluster as a CM node or is upgraded to a CH node under certain conditions.

[0117] The above clustering algorithm is used for clustering of a LEO satellite constellation, and the generated cluster structure has fewer clusters, good stability, effectively solves the dynamic clustering problem in a large-scale LEO satellite network, thereby optimizing network management and providing a basis for subsequent establishment of a hierarchical federation framework.

[0118] Step 203: constructing a dynamic topology graph of the satellite network.

[0119] For long-chain global planning, the dynamic topology of the satellite network needs to be obtained first. The acquisition of the topology needs to go through multiple stages, including data preprocessing, spatio-temporal graph construction, neural network training, and topology generation and processing stages. The detailed process is shown in Figure 3 The specific steps are as follows:

[0120] Data preprocessing stage: First, the input satellite ephemeris data is time-aligned and coordinate-normalized. Specifically, multi-source orbit data from TLE ephemeris is obtained, converted to UTC standard timestamp, and synchronized with all satellite sampling times. The three-dimensional coordinates in the Earth-Centered Inertial (ECI) coordinate system are normalized by linear transformation to scale the X, Y, and Z axis coordinate values to the [-1, 1] interval, generating a standardized coordinate matrix.

[0121] At the same time, based on inter-satellite link data (including distance, round-trip delay, etc.) and geometric visibility calculation results, a multi-dimensional link state feature vector (containing delay, bandwidth utilization, inter-satellite distance, and binary visibility flag) is constructed for each satellite. The link state tensor is filled with zero vectors for missing links, and finally the standardized satellite coordinate matrix and fully connected link state tensor are output, providing standardized input for subsequent spatio-temporal graph construction.

[0122] Spatio-temporal graph construction stage: In the spatio-temporal graph construction stage of dynamic topology prediction, the preprocessed satellite normalized coordinates and link state data are first converted into a graph structure. Each satellite node is assigned a feature vector, which is formed by concatenating its standardized position coordinates (XYZ values in the ECI coordinate system) and satellite link state data, forming a node feature matrix.

[0123] Next, by calculating the Euclidean distance between satellites and combining the geometric visibility condition, an edge weight matrix is generated, where the weight is defined as the product of the exponential decay function of the distance and the visibility indicator. Finally, a series of spatio-temporal graph sequences (each time step corresponds to a graph structure) are output, providing input for subsequent graph neural network modeling.

[0124] Neural network training stage: The spatio-temporal graph convolutional network ST-GCN is used to extract the spatio-temporal features of satellite dynamic topology. First, a second-order ChebNet is used to aggregate local spatial information between nodes within a single time step, capturing the dependence of satellite position and link state. Then, a one-dimensional convolution (Conv1D) is applied along the time axis to model the evolution trend of historical trajectories, thereby extracting the temporal correlation of satellite motion patterns.

[0125] To further enhance the feature expression capability, a multi-head attention mechanism is introduced to dynamically calculate the spatio-temporal dependence weight between satellite nodes, adaptively fuse key features at long distances and across time periods, and finally output a feature vector that integrates spatio-temporal information.

[0126] Based on the above ST-GCN extracted spatio-temporal feature vector, the inter-satellite connection probability at the future time is predicted. The existence probability of the link between any two satellites is calculated through a learnable fully connected neural network, and is mapped to the [0, 1] interval through the Sigmoid activation function to generate a global topology probability matrix.

[0127] Each element of the matrix reflects the stability or availability of the corresponding link within the prediction time window, for example, a high probability value may indicate that the two satellites maintain stable connection due to the close orbital period, while a low probability indicates that it will soon be out of the visible range.

[0128] During training, the binary cross-entropy loss function is mainly used to predict the link probability combined with dynamic sample weighting and L2 regularization method, and the AdamW optimizer (initial learning rate = 1e-3) and cosine annealing warm restart learning rate scheduling are used for efficient parameter update. Through continuous forward propagation, back propagation and loss calculation, a perfect model is finally obtained, so as to realize high-precision dynamic topology prediction.

[0129] Topology graph generation and processing stage: the probability matrix is processed in the topology graph generation and processing stage to output a deterministic topology. First, the probability graph is binarized by a fixed threshold to retain high-confidence links to form a sparse adjacency matrix, and then the global connectivity of the network is checked (to ensure that there are no isolated nodes, and the threshold is dynamically adjusted if necessary). The verified topology graph can be directly converted into an adjacency list or edge list format for the downstream path planning algorithm (such as Dijkstra or A*) to calculate the optimal inter-satellite route. The whole process strictly follows the real-time constraint to meet the online decision-making needs of the dynamic constellation network.

[0130] Step 204: using the DT-DVTR algorithm based on the satellite network dynamic topology graph to obtain a long-chain global optimal path.

[0131] After that, the DT-DVTR algorithm predicts the future dynamic topology based on the ST-GCN model in the above step S203 (including link survival probability, node position, load and energy state), and uses a multi-objective optimization strategy to decide the global optimal path.

[0132] For each candidate path Calculate the comprehensive utility function:

[0133]

[0134] Where, D p , B p , L p represent the current delay, bandwidth, load value of path p, D p,min , B p,max , Lp,max respectively represent the theoretical minimum latency, the theoretical maximum bandwidth and the maximum load threshold corresponding to the path p, and a, b, g represent the weight factors respectively, and the sum is equal to 1;

[0135] Filtering the feasible path set that meets the link survivability, node energy and load upper limit, and then converting the multi-objective into a single objective through the comprehensive utility function (combining latency, bandwidth and load balancing), the global optimal path is calculated by using Dijkstra algorithm to ensure the global performance optimization under the predicted topology.

[0136] The optimal path is stored in the LRU strategy management path library, and the path set is stored in a hash table, with the key being the service type hash value and the value being the path attribute tuple (latency, bandwidth, load).

[0137] Through high-priority signaling, the LRU strategy management path library of batch historical services is uploaded to each cluster head in each short chain every 1 hour.

[0138] Step 205: For the plurality of distributed clouds, a deep reinforcement learning model is used to determine the short-chain local optimal path.

[0139] For real-time adjustment of short chains, a deep reinforcement learning (DRL) is used to optimize local routing in real time. The duel double deep Q network (D3QN) algorithm is used to train the policy and value networks, taking into account the exploration and convergence speed to cope with instantaneous link disturbances. The detailed process is shown in Figure 4 The specific steps are as follows:

[0140] Based on Markov decision process, s is defined as the current state set, a is defined as the action set, and r is defined as the reward function.

[0141] By designing a state compression encoder, multi-dimensional transmission resource state (local link quality, neighbor node queue length, remaining energy, topology change flag) is mapped to a low-dimensional state vector s. It mainly uses a three-stage progressive dimension reduction structure: first, the original multi-dimensional state (latency, packet loss rate, etc.) is compressed to low dimension and the space-time features are learned through a deformable convolution layer; second, the Top-3 key features are dynamically selected through the gated attention module; finally, the compression is completed through the multi-layer bottleneck self-encoder. In addition, the cluster head periodically collects the local topology state changes within the cluster.

[0142] The action a is defined as selecting the next hop satellite node. According to the current state, the agent can select its own neighbor node for the next hop routing. Thus, link rate adjustment and priority weight allocation are performed.

[0143] The reward is defined as

[0144] r = a1 · t sur -a2 · t e2e-a3·v shake -a4·e energy ;

[0145] where t sur denotes the link survival time, t e2e denotes the end-to-end delay, v shake denotes the jitter variance, e energy denotes the node energy consumption, and α1, α2, α3, α4 denote the corresponding weight factors, respectively.

[0146] The Critic network (value function) evaluates the state value and calculates the advantage function, and the Actor network (policy function) outputs the action probability distribution according to the current state; the loss function combines the value function error term and the policy entropy regularization term using the Clip objective function, the optimizer selects Adam, and the network parameters are iteratively updated through small batch sampling (batch size = 64-256), and multi-thread asynchronous acquisition of environment interaction data is used to improve the training efficiency.

[0147] After the LEO satellite member nodes complete local model training based on private data, the encrypted model parameters (rather than the original data) are uploaded to the LEO satellite cluster head, and privacy protection is achieved through homomorphic encryption.

[0148] The cluster head performs weighted averaging on the received gradient parameters, generates a global model, and then distributes it to each member node to form a closed loop of "local training → parameter uploading → secure aggregation → global updating → model distribution". The entire process follows the differential privacy constraint to prevent member inference attacks, and an asynchronous communication mechanism is used to balance the convergence speed and communication overhead. The specific algorithm process is as follows:

[0149] Initialize the input satellite clustering set as The intra-cluster node set The cluster model parameter set is

[0150] Traverse each cluster c M , and each satellite node v N in the cluster.

[0151] Calculate the model update weight

[0152] where h n (w; w t ) is the minimization objective function, F n (ω) is the minimization local function of different devices n, is the difference between the parameters of the global model and the client model, and μ is the adjustable parameter. Reducing it can reduce the constraint effect.

[0153] Each LEO node vN sending updated weights to the cluster head node;

[0154] The cluster head node updates the aggregated model, i.e., the cluster head averages the model parameters of each node in the cluster, as follows where N is the number of nodes in the current cluster.

[0155] If the updated global model reaches the preset model accuracy in the test task, it is determined that the training of the routing algorithm is completed. Otherwise, the above steps are repeatedly executed until the global model reaches the preset model accuracy in the test task.

[0156] Step 206: Compare the long-chain global optimal path with the short-chain local optimal path. If the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are consistent, the long-chain global optimal path is adopted. If the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are inconsistent, the short-chain local optimal path is adopted to replace the corresponding parts of the long-chain global optimal path, forming a routing path.

[0157] Step 207: The short-chain collects satellite node dynamic data in real time and sends it to the long-chain. The dynamic data includes link state, queue depth, and energy reserve.

[0158] Step 208: The long-chain optimizes the global optimal path based on the dynamic data, forming a short-chain perception-long-chain optimization-strategy counterproductive closed-loop collaborative mechanism.

[0159] High-altitude orbit satellites (especially geostationary orbit GEO satellites) are often used as global routing calculation cores due to their wide coverage and relatively stable topology. Under stable normal conditions, the long-chain (generated by GEO or MEO layers) may be optimal, and the short-chain is locally optimal. The short-chain is a local optimization adjustment based on the global results of the long-chain. The short-chain can provide perception information results and long-chain initialization splicing for the long-chain. The results generated by the long-chain provide data initialization and reward evaluation basis for local short-chain learning and training.

[0160] On each satellite node, compare the long-chain global path with the local path generated by the short-chain through DRL. If the short-chain result is consistent with the long-chain prediction result or the satellite node load is high at this time, the long-chain result is adopted. If the short-chain result is inconsistent with the long-chain prediction result, the short-chain real-time prediction result is adopted.

[0161] In addition, the short-chain collects multi-dimensional data such as link state (delay, packet loss rate), queue depth, and energy reserve of satellite nodes in real time through distributed sensors, and periodically uploads state snapshots to the ground long-chain in an incremental compression encoding manner.

[0162] Longchain dynamically optimizes the prediction accuracy of the terrestrial satellite network digital simulation platform using this data, while updating the business demand patterns and the Pareto optimal solution set in the global path library, forming a closed-loop collaborative mechanism of "shortchain perception - longchain optimization - strategy feedback". This process employs differential privacy encryption and federated aggregation technology to enhance the global planning capabilities of longchain while ensuring the data security of satellite nodes, ultimately achieving continuous performance optimization of the satellite-ground collaborative network.

[0163] Implementing steps 201 to 208, an LEO satellite clustering algorithm is deployed at the MEO satellite layer. Each LEO satellite is clustered using MEO satellites, resulting in a cluster structure with fewer clusters and better stability. This effectively solves the dynamic clustering problem in large-scale LEO satellite networks, thereby optimizing network management and providing a foundation for the subsequent establishment of a hierarchical federated framework. The short chain collects multi-dimensional data such as link status (latency, packet loss rate), queue depth, and energy reserves of satellite nodes in real time through distributed sensors, and periodically uploads state snapshots to the ground-based long chain using incremental compression encoding. The long chain uses this data to dynamically optimize the prediction accuracy of the ground-based satellite network digital simulation platform, while simultaneously updating the business demand patterns and the Pareto optimal solution set in the global path library, forming a closed-loop collaborative mechanism of "short chain perception - long chain optimization - strategy feedback." This process employs differential privacy encryption and federated aggregation technology to enhance the global planning capabilities of the long chain while ensuring the data security of satellite nodes, ultimately achieving continuous performance optimization of the space-ground collaborative network.

[0164] In another exemplary embodiment of this application, in order to achieve a precise match between service QoS requirements and satellite network resource supply, the method further includes:

[0165] Obtain typical service types; the typical service types include: latency-sensitive services, compute-intensive services, and bandwidth-intensive services;

[0166] Based on the aforementioned business type, determine the business priority;

[0167] The weighting factors α, β, and γ in the comprehensive utility function are adjusted in real time according to the business priority to maximize the comprehensive utility function.

[0168] The optimal path is selected based on the adjusted comprehensive utility function.

[0169] In other words, service QoS requirements are quantified into three typical objectives: latency-sensitive services (such as real-time communication) aim for minimum end-to-end latency, compute-intensive services (such as AI inference) focus on minimizing energy consumption, and bandwidth-intensive services (such as file transfer) aim to maximize bandwidth. By modeling quantitative indicators, the system can dynamically adjust priorities according to service type, providing standardized input for subsequent resource allocation.

[0170] Afterwards, the weight factors a, b, g in the comprehensive utility function are adjusted in real time according to the service priority (such as delay-sensitive, computation-intensive), for example, if it is a delay-sensitive service, the weight factor a is increased, and the weight factors b and g are reduced, so as to maximize the comprehensive utility function, and the best path p is selected according to the utility function score * As shown in the following formula.

[0171] p * = argmax U(p);

[0172] Finally, the short chain periodically downloads the real-time monitored link state, queue depth and other data to the long chain, and the latter calculates the latest weight strategy and then issues it. The service traffic is routed according to the weight priority: the delay-sensitive service takes the low-delay path, the bandwidth-sensitive service takes the high-bandwidth path, and the computation-intensive service tends to high-computation nodes. It ensures that the weight is adaptively adjusted according to the network state, suppresses the path oscillation caused by local congestion, and finally realizes the high-precision matching of QoS and resources.

[0173] Based on the same inventive concept, the embodiments of the present application also provide a long-short chain cooperative mega-constellation hierarchical multi-domain routing device for implementing the long-short chain cooperative mega-constellation hierarchical multi-domain routing method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more long-short chain cooperative mega-constellation hierarchical multi-domain routing device embodiments provided below can be referred to the limitations of the long-short chain cooperative mega-constellation hierarchical multi-domain routing method in the above, which will not be repeated here.

[0174] In one exemplary embodiment, a long-short chain cooperative mega-constellation hierarchical multi-domain routing device is provided, referring to Figure 5 The device comprises:

[0175] A satellite network hierarchical multi-domain architecture construction module is configured to construct a satellite network hierarchical multi-domain architecture; the satellite network hierarchical multi-domain architecture comprises a GEO satellite layer, a MEO satellite layer, and a LEO satellite layer; the GEO satellite layer is deployed with an SDN controller, which is configured to maintain a global view and calculate and generate a long-chain global optimal path;

[0176] A dynamic clustering module is configured to dynamically cluster the LEO satellite network by using a LEO satellite clustering algorithm, to form a plurality of distributed clouds;

[0177] A satellite network dynamic topology graph construction module is configured to construct a satellite network dynamic topology graph;

[0178] The long-chain global optimal path determination module is configured to obtain a long-chain global optimal path based on the dynamic topology graph of the satellite network by using a discrete-time virtual topology routing (DT-DVTR) algorithm.

[0179] The short-chain local optimal path generation module is configured to determine a short-chain local optimal path for the multiple distributed clouds by using a deep reinforcement learning model.

[0180] The actual routing path generation module is configured to compare the long-chain global optimal path and the short-chain local optimal path, and if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are consistent, the long-chain global optimal path is used, and if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are inconsistent, the short-chain local optimal path is used to replace the corresponding parts of the long-chain global optimal path to form an actual routing path.

[0181] The dynamic data acquisition module is configured to collect satellite node dynamic data in real time in a short chain and send the dynamic data to a long chain. The dynamic data includes link state, queue depth, and energy reserve.

[0182] The path optimization module is configured to optimize the global optimal path in a long chain based on the dynamic data to form a short-chain perception-long-chain optimization-strategy feedback closed-loop collaborative mechanism.

[0183] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0184] The principles and implementation modes of the present application are described by using specific examples. The above descriptions of the embodiments are only used to help understand the method and its core idea of the present application. For those skilled in the art, according to the idea of the present application, the specific implementation modes and application ranges can be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A method for hierarchical multi-domain routing based on long-short chain cooperation of giant constellations, characterized in that, The method comprises the following steps: A satellite network hierarchical multi-domain architecture is constructed, which comprises a GEO satellite layer, a MEO satellite layer and a LEO satellite layer; the GEO satellite layer is provided with an SDN controller for maintaining a global view and calculating and generating a long-chain global optimal path; A LEO satellite clustering algorithm is used to dynamically cluster the LEO satellite network to form a plurality of distributed clouds; A satellite network dynamic topology graph is constructed; A discrete-time virtual topology routing (DT-DVTR) algorithm is used based on the satellite network dynamic topology graph to obtain a long-chain global optimal path; A deep reinforcement learning model is used to determine a short-chain local optimal path for the plurality of distributed clouds; The long-chain global optimal path and the short-chain local optimal path are compared, and if the corresponding parts of the short-chain local optimal path and the long-chain global optimal path are consistent, the long-chain global optimal path is used, and if not, the short-chain local optimal path is used to replace the corresponding parts of the long-chain global optimal path to form an actual routing path; Short-chain real-time satellite node dynamic data is collected and sent to a long chain; the dynamic data includes link state, queue depth and energy reserve; The long chain optimizes the global optimal path based on the dynamic data to form a short-chain perception-long-chain optimization-strategy feedback closed-loop collaborative mechanism.

2. The method of claim 1, wherein the method is based on long-short chain cooperation of giant constellation hierarchical multi-domain routing. The MEO satellite layer is provided with a lightweight SDN agent; and the LEO satellite layer is autonomously controlled by the clustering algorithm.

3. The method of claim 2, wherein the method is based on long-short chain coordination. The LEO satellite clustering algorithm comprises two stages, i.e., a cluster initialization stage and a cluster dynamic maintenance stage, and the cluster initialization stage comprises the following steps: S1: Obtain ephemeris data of LEO satellites and divide the entire operation cycle into a plurality of time slices; S2: Calculate the average connectivity of satellite nodes based on the ephemeris data and the number of time slices; S3: Determine whether there is a neighbor node based on the average connectivity, and if not, mark it as a free node; S4: If there is a neighbor node, mark it as an undetermined node and check whether there is a determined cluster head node within its communication range; S5: If there is a cluster head node, calculate the priority; S6: Select the optimal cluster head based on the priority; S7: Add the undetermined node in step S4 to the optimal cluster head and update the undetermined node to a cluster member node; S8: If there is no cluster head node, compare the average connectivity of the undetermined node itself and its neighbor nodes in step S4; S9: Select a satellite node with the maximum average connectivity and upgrade it to a cluster head node to create a new cluster; the new cluster follows a single cluster head node constraint condition, and the initial cluster size is set to 1, and if the condition is not met, the node state is reset for the next round of determination; S10: Repeat steps S1-S9 until the state of all nodes in the LEO satellite network is determined.

4. The method of claim 3, wherein the method is based on long-short chain cooperation of giant constellation hierarchical multi-domain routing. The cluster dynamic maintenance stage comprises cluster head node state updating, cluster member node state updating, free node state updating and undetermined node state updating. The cluster head node state updating specifically comprises: when the number of cluster members of the cluster head node is 1 and exceeds a preset time limit, triggering state degradation, converting the cluster head node into an undetermined node, and broadcasting a state change message to a neighboring area to trigger a cluster reorganization process; The cluster member node state updating: when the cluster member node detects that the number of hops between the cluster member node and the cluster head node exceeds a preset threshold, triggering a state updating process, the cluster member node first initiates a cluster leaving request to the current cluster head node, judges whether there is a neighbor node in a detection range, if not, directly converts into a free state; if there is a neighbor node, degrades into an undetermined state and participates in a new round of dynamic clustering process again; The free node state updating: when the free node finds that there is a connection between itself and other nodes, converts into an undetermined node and participates in a new round of dynamic clustering process again; The undetermined node state updating: when the undetermined node selects to join an optimal cluster head as a cluster member node or the connection degree of the undetermined node reaches a preset condition, upgrades into a cluster head node.

5. The method of claim 1, wherein the method is based on long-short chain coordination of super-constellation hierarchical multi-domain routing. The construction of the satellite network dynamic topology map specifically comprises the following steps: Pretreat the satellite ephemeris data to obtain a standardized satellite coordinate matrix and a fully connected link state tensor; Convert the standardized satellite coordinate matrix and the fully connected link state tensor into a graph structure; each satellite node in the graph structure is assigned a feature vector, which is formed by splicing the standardized satellite coordinate matrix and the satellite link state tensor to form a node feature matrix; Based on the graph structure, calculate the Euclidean distance between the satellite nodes and generate an edge weight matrix in combination with a geometric visibility condition; wherein, the weight is defined as the product of the exponential decay function of the distance and the visibility indicator; Output a space-time graph sequence based on the weight matrix; Take the space-time graph sequence as input, and use a space-time graph convolution network ST-GCN to extract a space-time feature vector of the satellite dynamic topology; Based on the space-time feature vector, predict the inter-satellite connection probability at a future time, and generate a global topology probability matrix; Process the global topology probability matrix to generate a stable topology, i.e. a satellite network dynamic topology map; the satellite network dynamic topology map comprises: link survival probability, node position, load and energy consumption state.

6. The method of claim 1, wherein the method is based on long-short chain coordination of super-constellation hierarchical multi-domain routing. Based on the satellite network dynamic topology map, a discrete time virtual topology routing DT-DVTR algorithm is used to obtain a long-chain global optimal path, specifically comprising the following steps: Screen feasible paths that meet the link survival, node energy and load upper limits; For each feasible path, calculate a comprehensive utility function: where D p , B p , L p represent the current delay, bandwidth, load value of path p, D p,min , B p,max , L p,max represent the theoretical minimum delay, theoretical maximum bandwidth and maximum load threshold value of path p respectively, and a, b, g represent weight factors respectively, and add up to 1; Based on the comprehensive utility function of each feasible path, determine the long-chain global optimal path by using Dijkstra algorithm.

7. The method of claim 6, wherein the method is based on long-short chain cooperation of giant constellation hierarchical multi-domain routing. The giant constellation hierarchical multi-domain routing method based on long-short chain cooperation further comprises the following steps after the step "long-chain global optimal path is optimized based on the dynamic data to form a short-chain perception-long-chain optimization-strategy counter-acting closed-loop cooperation mechanism": Obtain typical service types; the typical service types comprise: delay-sensitive service, computation-intensive service and bandwidth-intensive service; Determine service priority based on the service types; Adjusting the weight factors α, β, γ in the comprehensive utility function in real time according to the service priority to maximize the comprehensive utility function; Selecting the optimal path according to the adjusted comprehensive utility function.

8. A hierarchical multi-domain routing apparatus based on long-short chain cooperation, characterized in that, The long-short chain cooperation-based giant constellation hierarchical multi-domain routing device comprises: A satellite network hierarchical multi-domain architecture construction module is configured to construct a satellite network hierarchical multi-domain architecture; the satellite network hierarchical multi-domain architecture comprises a GEO satellite layer, a MEO satellite layer and a LEO satellite layer; the GEO satellite layer is deployed with an SDN controller, which is configured to maintain a global view and calculate and generate a long-chain global optimal path; A dynamic clustering module is configured to dynamically cluster the LEO satellite network by using a LEO satellite clustering algorithm to form a plurality of distributed clouds; A satellite network dynamic topology graph construction module is configured to construct a satellite network dynamic topology graph; A long-chain global optimal path determination module is configured to obtain a long-chain global optimal path by using a discrete-time virtual topology routing (DT-DVTR) algorithm based on the satellite network dynamic topology graph; A short-chain local optimal path generation module is configured to determine a short-chain local optimal path for the plurality of distributed clouds by using a deep reinforcement learning model; An actual routing path generation module is configured to compare the long-chain global optimal path and the short-chain local optimal path; if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are consistent, the long-chain global optimal path is used; if the corresponding parts of the long-chain global optimal path and the short-chain local optimal path are inconsistent, the short-chain local optimal path is used to replace the corresponding parts of the long-chain global optimal path to form an actual routing path; A dynamic data acquisition module is configured to collect satellite node dynamic data in real time in a short chain and send the dynamic data to a long chain; the dynamic data comprises link state, queue depth and energy reserve; A path optimization module is configured to optimize the global optimal path based on the dynamic data in a long chain to form a short-chain perception-long-chain optimization-strategy feedback closed-loop cooperation mechanism.

Images