A routing method for a heterogeneous low earth orbit satellite network and an electronic device

Abstract

This invention discloses a routing method and electronic device for heterogeneous low-Earth orbit satellite networks. The method includes: constructing a domain-specific network architecture and establishing a hierarchical architecture comprising a data plane, a control plane, and a meta-learning layer; establishing an end-to-end latency model and an energy consumption model; defining a multi-objective optimization function based on KL divergence; and executing meta-learning routing decisions, which includes a meta-optimization stage and an agent optimization stage. In the meta-optimization stage, an adaptive learning objective is generated through the meta-network; in the agent optimization stage, the routing method is updated according to the learning objective, and routing actions are executed. This invention solves the problems of poor adaptability of routing methods and difficulties in multi-objective trade-offs in heterogeneous networks, achieving adaptive generation and efficient collaboration of routing methods, and improving the overall performance of the network.

Description

Technical Field

[0001] This invention relates to the field of satellite communication network technology, and more specifically, to a routing method and electronic equipment for heterogeneous low-Earth orbit satellite networks, particularly a technique for adaptive routing decision-making using a meta-learning mechanism. Background Technology

[0002] With the increasing global demand for communication, low-Earth orbit (LEO) satellite networks, with their advantages of global coverage and low latency, are becoming the core infrastructure for building the next generation of seamless information networks. To effectively manage the massive constellations of tens of thousands of satellites, a domain-based management architecture has emerged. By dividing the constellation into multiple logical or physical control domains, it achieves decentralized and scalable network management, and has become the mainstream trend in current LEO satellite network design.

[0003] However, while domain-based management improves management efficiency, it also increases the frequency of inter-satellite link switching and the computational burden of path calculation. Different control domains may cover geographical areas with vastly different service characteristics. For example, a control domain covering densely populated urban areas faces high concurrent traffic and frequent link congestion, while a control domain covering vast oceans or deserts may have lower link loads but face complex space environment interference. This inherent heterogeneity leads to significant differences in network topology dynamics, link quality distribution, and service quality requirements among different control domains. Therefore, routing mechanisms must not only operate efficiently within a single domain, but more importantly, they must possess the ability to flexibly migrate and quickly adapt between heterogeneous domains.

[0004] To address the complex and dynamic environment of satellite networks, reinforcement learning-based intelligent routing methods have become a hot research topic. These methods model satellite nodes as intelligent agents, learning and optimizing routing methods through continuous interaction with the network environment. However, when applied to heterogeneous low-Earth orbit satellite networks with domain-specific management, traditional reinforcement learning methods reveal their inherent limitations. First, traditional reinforcement learning algorithms typically rely on fixed, manually designed learning rules (such as fixed learning rates and update directions). When faced with control domains exhibiting vastly different characteristics, they cannot dynamically adjust their learning logic, resulting in poor adaptability, slow convergence, and even the potential for catastrophic forgetting during cross-domain routing. Second, traditional reinforcement learning methods often employ a single scalar reward function to guide optimization. This is insufficient in complex scenarios requiring a trade-off between multiple performance metrics such as latency and energy consumption, making it difficult to balance convergence speed under highly dynamic topologies with the customized needs of heterogeneous tasks. It struggles to accurately express the subtle differences in performance preferences among different services or control domains, ultimately leading to routing methods that fail to meet the heterogeneous and multidimensional quality of service requirements in practical applications.

[0005] In summary, existing technologies lack a routing scheme that can systematically solve the problems of heterogeneous subnet collaboration, cross-domain adaptive generation, and multi-objective optimization, making it difficult to meet the requirements of efficient, flexible, and autonomous operation of future heterogeneous low-Earth orbit satellite networks. Summary of the Invention

[0006] The main objective of this invention is to provide a routing method and electronic device for heterogeneous low-Earth orbit satellite networks, in order to solve the problems in the prior art where routing methods are difficult to adapt to network heterogeneity, traditional reinforcement learning algorithms have poor generalization ability, and multi-objective optimization is difficult.

[0007] To achieve the above objectives, this invention provides a routing method for heterogeneous low-Earth orbit (LEO) satellite networks, characterized by comprising: constructing a domain-based network architecture, dividing the LEO satellite constellation into multiple heterogeneous control domains, each of which is managed by a control node; establishing a hierarchical architecture based on the heterogeneous control domains, the hierarchical architecture including a data plane composed of satellite nodes, a control plane composed of control nodes, and a meta-learning layer containing a meta-network, the meta-network being used to generate learning objectives adapted to different heterogeneous control domains; establishing an end-to-end latency model and energy consumption model in the heterogeneous LEO satellite network environment to quantitatively evaluate the latency and energy consumption of routing; and defining a multi-objective optimization function to balance latency and energy consumption, wherein the multi-objective... The optimization function is expressed as evaluating the KL divergence between the empirical performance distribution guided by the routing method of the satellite agent and a potential preference distribution representing an ideal trade-off between latency and energy consumption; performing meta-learning routing decisions, including a meta-optimization phase and an agent optimization phase; in the meta-optimization phase, the satellite agent interaction trajectories collected from multiple heterogeneous control domains are processed by the meta-network to generate the adaptive learning objective, and the parameters of the meta-network are updated using the meta-gradient method; in the agent optimization phase, each satellite agent in the heterogeneous control domain updates its own routing method network parameters according to the learning objective generated by the meta-network, and selects the next-hop satellite node to perform routing actions based on the updated routing method.

[0008] Furthermore, the empirical performance distribution is composed of a set of historical performance index vectors collected by the agent during its interaction with the environment, and each historical performance index vector contains observations in two dimensions: latency and energy consumption; the potential preference distribution is a pre-defined target distribution whose shape reflects the current control domain's trade-off preference for the two performance indicators of latency and energy consumption.

[0009] Furthermore, the potential preference distribution is constructed as a Gaussian mixture model; the mean vector of the Gaussian mixture model corresponds to a preset service quality target anchor point, the service quality target anchor point includes a first anchor point, which represents the performance index combination under the latency-first scenario, and a second anchor point, which represents the performance index combination under the energy-first scenario; the covariance matrix of the Gaussian mixture model reflects the tolerance for deviation from the service quality target anchor point, wherein a smaller covariance value indicates a more stringent performance requirement for the corresponding anchor point.

[0010] Furthermore, the agent's interaction trajectory includes a preset time window. Within the network, the system contains the agent's state sequence, the sequence of routing actions performed, and the sequence of performance metric feedback obtained. The meta-network is configured to receive the agent's interaction trajectory as input and output the learning objective, which is a gradient direction vector that guides the updating of the routing method's network parameters.

[0011] Furthermore, the meta-network adopts a recurrent neural network architecture based on gated recurrent units to handle the temporal dependencies in the agent's interaction trajectory; in the meta-optimization stage, the meta-gradient method is used to update the parameters of the meta-network, specifically: calculating the gradient of the sum of performance improvements generated by the learning objective after guiding the agent to optimize in multiple heterogeneous control environments relative to the parameters of the meta-network, and updating the parameters of the meta-network along the gradient direction.

[0012] Furthermore, the latency model comprehensively considers the transmission latency, propagation latency, queuing latency, and processing latency of the inter-satellite link, as well as the transmission latency, propagation latency, and processing latency of the satellite-to-ground link; the energy consumption model comprehensively considers the energy consumption of the communication transceiver module and the routing calculation module of the satellite node.

[0013] Furthermore, in the layered architecture, the data plane consists of satellite nodes and is responsible for forwarding data packets; the control plane consists of control nodes in each of the heterogeneous control domains and is responsible for executing routing methods within the domain; the meta-learning layer is deployed at ground stations or core gateway nodes and is responsible for generating and distributing global learning rules.

[0014] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory, characterized in that, when the processor executes the program, it implements the method described above.

[0015] The present invention also provides a computer program product, including computer instructions, characterized in that the instructions, when executed by a processor, implement the method described above.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described above.

[0017] This invention achieves refined management and global collaboration of heterogeneous networks by constructing a domain-specific meta-learning architecture. It utilizes a meta-optimization framework based on discovery reinforcement learning to replace traditional fixed learning rules, enabling routing methods to adaptively generate based on environmental characteristics, significantly improving the agent's adaptability across heterogeneous domains. Furthermore, it employs a multi-objective optimization function based on KL divergence, transforming routing decisions from maximizing a single reward to distributed matching, thus more flexibly and accurately meeting the diverse performance trade-offs in heterogeneous networks. Compared to existing technologies, this invention offers advantages such as strong adaptive routing methods, high network collaboration efficiency, and flexible multi-objective optimization. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a structural block diagram of an electronic device provided in an embodiment of the present invention; Figure 2 This is a flowchart of a routing method for heterogeneous low-Earth orbit satellite networks provided in an embodiment of the present invention; Figure 3 This is a detailed flowchart of the meta-learning routing decision method provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the layered network architecture provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1

[0022] This embodiment provides a routing method for heterogeneous low-Earth orbit satellite networks, which can be applied to, for example... Figure 1The electronic device 100 shown can be a satellite node, ground station, or core gateway node that performs control functions in a satellite network. Specifically, the electronic device 100 includes a processor 101, a memory 102, and a communication interface 103. The memory 102 stores computer program instructions, and the processor 101 executes the program stored in the memory 102 to implement the routing method described in this embodiment. The communication interface 103 is used for data interaction with other nodes in the network (such as other satellites or ground stations). The system also includes a power module that provides power to the various components.

[0023] Reference Figure 2 The routing method in this embodiment mainly includes the following steps: Step S201: Construct a domain-based network architecture.

[0024] To address the management complexity and network heterogeneity of large-scale LEO satellite constellations, this embodiment constructs a heterogeneous domain-based network architecture. This architecture divides the entire LEO satellite constellation into multiple logically non-overlapping heterogeneous control domains. For example, these domains can be divided based on the satellites' geographical coverage, service load characteristics, or orbital parameters. Each control domain is managed by a control node, which is preferably the satellite with the best computing power and energy status within the domain. This control node is responsible for collecting and maintaining the status information of the satellites within the domain in real time and executing routing decisions based on globally distributed learning objectives.

[0025] Based on the predefined heterogeneous control domains, a layered architecture integrating software-defined networking concepts is established, such as... Figure 4 As shown. The architecture, from bottom to top, includes a data plane 401, a control plane 402, and a meta-learning layer 403. This layered architecture clearly defines the correspondence between each layer and network entities: Data plane 401 consists of all satellite nodes in the network, and its core responsibility is to perform specific packet forwarding operations according to the routing method issued by the control plane.

[0026] Control plane 402 consists of control nodes within various heterogeneous control domains. Each control node acts as a local "brain" of its domain, responsible for collecting network status information, performing routing calculations, and coordinating the behavior of satellite nodes within its domain. This distributed control approach reduces the complexity of global management and enables scalable control of large low-Earth orbit constellations.

[0027] The meta-learning layer 403 is the core of this architecture and is typically deployed on ground stations or core gateway nodes where computing resources are more abundant. This layer contains a core meta-network whose function is "learning how to learn," that is, by analyzing data collected from various heterogeneous control domains, it automatically generates learning objectives or update rules that can adapt to the characteristics of different domain environments, and distributes these rules to various control nodes in the control plane to guide them in optimizing their local routing methods.

[0028] In terms of network modeling, this embodiment can use the Walker-Delta constellation model for topology construction, and its parameters can be expressed as follows: ,in Represents the total number of satellites. Indicates the number of orbital planes. The phase factor. Each satellite can be configured to maintain up to four inter-satellite links, including, for example, two stable intra-plane links and two dynamically changing inter-plane links. The operation time is discretized into... A length of A uniform time slot, the time slot set is represented as = {1, 2, ..., The set of entities in a network is defined as follows: User set = { , , ..., Gateway collection = { , ,..., }, satellite node set = { , , ..., Inter-satellite link set = {( , ) | , ∈ , ≠ The set of user requests is denoted as}. = { , , ..., The constellations are divided into... There are control domains, and the set of domains is represented as... = { , , ..., For any control domain Its internal node set and link set are respectively represented as and The set of inter-domain links connecting different domains is represented as .

[0029] Specifically, the meta-network adopts a two-layer GRU architecture. Its input vector consists of the current cycle's state sequence s, action sequence a, and feedback delay / energy consumption scalars. The input dimension is set according to the upper limit of the satellite node's neighbor nodes. The adaptive learning objective output by the meta-network is mapped as a gradient direction tensor with the same dimension as the agent's policy network parameters.

[0030] Step S202: Establish end-to-end latency model and energy consumption model.

[0031] These two models are used to quantify and calculate latency and energy consumption during data transmission, providing a computational foundation for subsequent evaluation of routing performance and multi-objective optimization. This embodiment establishes a refined end-to-end latency and energy consumption model.

[0032] 1. End-to-end delay model

[0033] The delay model comprehensively considers various delays in the transmission of data packets on inter-satellite links and satellite-to-ground links.

[0034] For inter-satellite links, in time slots ,satellite and The signal-to-noise ratio of the links can be calculated. In a preferred embodiment, this calculation logic is implemented using the following formula:

[0035] in, and They are satellites The transmit power and antenna gain, It is a satellite The receiving antenna gain, Boltzmann's constant, The system noise temperature, It is the allocated link bandwidth. This indicates the loss of free space.

[0036] Based on the signal-to-noise ratio, the maximum transmission rate of the link can be calculated:

[0037] The latency on inter-satellite links mainly includes: Transmission delay, or the time it takes to send data packets, is calculated as follows: ,in For the request The amount of data.

[0038] Propagation delay, which is the time it takes for a signal to propagate through space, is calculated as follows: ,in The distance between the two satellites. It is the speed of light.

[0039] Queuing delay, which is the time it takes for a data packet to wait for forwarding in a node's buffer, is related to the degree of network congestion.

[0040] Processing latency, which is the time it takes for a satellite node to process data packets, is calculated as follows: ,in It is to handle requests Number of CPU cycles required It is a satellite Currently available computing power (CPU cycles per second).

[0041] Therefore, the routing delays of an intra-domain link and an inter-domain link can be expressed as follows:

[0042]

[0043] For satellite-to-ground links, the delay model is similar to that of inter-satellite links, but weather conditions (such as rain attenuation) need to be taken into account. The influence of factors such as ) on the signal-to-noise ratio (SNR) can be expressed as follows: The SNR of the uplink and downlink can be expressed as follows:

[0044]

[0045] in, and These are gateways The transmit power and antenna gain, It is a satellite The receiving antenna gain; and They are satellites The transmit power and antenna gain, It is a gateway. The receiving antenna gain; and It is the allocated link bandwidth. and This indicates the loss of free space.

[0046] The processing latency of satellite-to-ground links is relatively small and can usually be ignored. Therefore, uplink and downlink latency mainly consist of transmission latency and propagation latency.

[0047]

[0048] in, and These represent the uplink and downlink transmission delays between the gateway and the satellite, respectively. and These represent the propagation delays of the uplink and downlink between the gateway and the satellite, respectively.

[0049] Ultimately, for a user originating from the access gateway... via satellite path Reaching the destination gateway Then, the request from the target user. The total end-to-end routing delay is the sum of the delays of each segment. In a preferred embodiment, this calculation logic is implemented using the following formula:

[0050] in, and These represent intra-domain and inter-domain inter-satellite link delays, respectively. It is a constant representing a fixed delay in the ground segment.

[0051] 2. End-to-end energy consumption model

[0052] Energy consumption mainly includes communication energy consumption and computing energy consumption.

[0053] Communication energy consumption, which is the energy consumed by a satellite in transmitting and receiving data packets, can be expressed as:

[0054] in, and These are the transmit and receive powers, respectively. This refers to the satellite link transmission delay.

[0055] Computational energy consumption, that is, the energy consumed by a satellite in performing processing tasks such as route calculations, can be expressed as:

[0056] in, For the satellite's processing power, To handle satellite latency.

[0057] ask The total energy consumption of a satellite network segment is the sum of the communication energy consumption of all links along the path and the computing energy consumption of all nodes:

[0058] in, This represents the total energy consumption of satellite links within the domain. This represents the total energy consumption of inter-domain satellite links.

[0059] Step S203: Define a multi-objective optimization function to balance and optimize latency and energy consumption.

[0060] Traditional reinforcement learning routing methods typically maximize a scalar reward value, struggling to handle the complex trade-offs between latency and energy consumption. This embodiment abandons this paradigm, reformulating the routing optimization problem as a distribution matching problem. Specifically, the optimization objective is defined as minimizing the Kullback-Leibler (KL) divergence between the empirical performance distribution guided by the routing method and a latent preference distribution representing an ideal compromise.

[0061] Among them, empirical performance distribution It is a distribution composed of a series of performance indicator vectors (latency, energy consumption) actually observed by the intelligent agent (satellite node) during its interaction with the environment. Latent preference distribution This is a pre-defined target distribution representing the current control domain's preference for trade-offs between latency and energy consumption. For example, for a control domain carrying real-time communication services, its preference distribution may have a higher probability density in the low-latency region; while for an energy-constrained control domain, its preference distribution may be biased towards the low-energy region.

[0062] By minimizing the KL divergence between these two distributions, the behavioral patterns learned by the agent's routing method can be guided so that the statistical distribution characteristics of its performance (a combination of latency and energy consumption) can accurately approximate the preset preference target.

[0063] In a preferred embodiment, the potential preference distribution It is constructed as a Gaussian mixture model. The mean vector of this model can be set as preset service quality target anchor points. For example, one anchor point is a combination of (low latency, high energy consumption) representing a latency-priority scenario; another anchor point is a combination of (high latency, low energy consumption) representing an energy-priority scenario. The model's covariance matrix reflects the tolerance for deviations from these anchor points; a smaller covariance value means more stringent requirements for the corresponding performance indicators.

[0064] The complete optimization problem can be expressed as:

[0065] in, Denotes KL divergence, Represents the routing method, the entropy term in the objective function. Logarithmic terms are used for regularization to stabilize the learning process.

[0066] This optimization problem is subject to a series of constraints: C1: (Total delay constraint); C2: (Total energy consumption constraint); C3: (Link capacity constraints); C4: (Each satellite belongs to only one domain); C5: (Control domains do not overlap); C6: (There are upper and lower limits to the number of satellites managed by each domain).

[0067] The potential preference distribution is initialized using a preset set of anchor points. In the heterogeneous control domain, for services with a preference for low latency (such as voice), the anchor points are shifted towards the latency coordinate axis; for services with a preference for energy saving (such as non-real-time telemetry), the anchor points are shifted towards the energy consumption coordinate axis. The covariance matrix is ​​dynamically assigned values ​​based on the fluctuation variance of traffic within the domain.

[0068] Step S204: Perform meta-learning routing decisions.

[0069] The meta-learning mechanism used in this invention is specifically based on the Discovering Reinforcement Learning (DRL) framework, which automatically 'discovers' and generates learning rules suitable for the current environment through a meta-network.

[0070] To solve the aforementioned optimization problem and achieve adaptive routing across heterogeneous domains, this embodiment employs a meta-optimization framework based on discovery reinforcement learning. This framework models the routing decision problem as a Markov decision process, where satellite nodes performing specific routing actions are defined as agents (i.e., satellite agents). The agent's state... It can be defined as a vector containing multiple dimensions, for example... = { } represents the estimated distance from the current node to the destination, the local available computing resources, and the available bandwidth resources of the link, respectively. The agent's actions... From its neighboring node set Select a node As the next jump.

[0071] The decision-making process proceeds through two collaborative optimization phases, such as... Figure 3 As shown: the meta-optimization stage and the agent optimization stage.

[0072] Step S301: Meta-optimization stage.

[0073] The goal of this stage is to learn a universal "learning rule" that enables the agent to learn quickly in any heterogeneous environment. This function is undertaken by the meta-network deployed in the meta-learning layer 403.

[0074] At the start of each meta-iteration (step S302), the system collects the interaction trajectories of agents from multiple heterogeneous control domain environments. An agent's interaction trajectory (step S303) is within a preset time window. Within, the sequence of states experienced by the agent, the sequence of routing actions executed, and the sequence of performance metrics (latency, energy consumption) feedback obtained.

[0075] The meta-network receives these trajectory data as input (step S304) and outputs an adaptive learning objective. This learning objective is no longer a fixed scalar reward, but a complex gradient direction vector that directly guides the agent on how to update its internal routing method network parameters.

[0076] In a preferred embodiment, the meta-network employs a recurrent neural network architecture based on gated recurrent units, which is adept at handling time-dependent sequential data such as agent interaction trajectories.

[0077] The parameters of the meta-network itself The update is performed using the meta-gradient method (step S305). Specifically, the system evaluates the overall performance improvement (i.e., the reduction in KL divergence) resulting from guiding agents in various environments through several optimization steps using the learning objective generated by the current meta-network. Then, this performance improvement is calculated relative to the meta-network parameters. The gradient is calculated and updated along the gradient direction. Its update logic can be approximated as:

[0078] in, It is the expected return guided by the meta-network. Represents a sampling environment. These are the routing method network parameters for the agent. By iteratively repeating this process, the metanetwork can "discover" an efficient learning rule that enables its generated learning objectives to generalize to various unseen heterogeneous low-orbit environments.

[0079] Step S306: Agent optimization stage.

[0080] Within each control domain, the local satellite agent receives learning objectives "tailor-made" for it from the meta-learning layer (step S307).

[0081] Subsequently, based on this learning objective, the satellite agent uses optimization algorithms such as gradient descent to update its own routing method network parameters. (Step S308). The optimization objective is to minimize the difference between the method network output and the learning objective generated by the meta-network. Its loss function can be expressed as:

[0082] in, It is the goal of meta-network generation. It is the output of the agent network. It is an auxiliary loss term used to stabilize the training process.

[0083] After the parameters are updated, the agent, based on the optimized routing method network, inputs the current state, selects the optimal next-hop satellite node to perform the routing action (step S309), and forwards the data packet.

[0084] When a new satellite node enters a control domain, the meta-learning layer directly distributes the pre-trained meta-network parameters to the control node of that node, thereby enabling zero-sample or small-sample rapid initialization of the routing strategy.

[0085] Through alternating iterations of meta-optimization and agent-based optimization, the system can achieve continuous adaptive evolution of routing methods. The meta-network learns general knowledge across environments from a global perspective, while the agent uses this general knowledge to make rapid and refined method adjustments in the local environment, thereby achieving efficient, flexible, and collaborative routing across the entire heterogeneous low-Earth orbit satellite network.

[0086] In summary, this embodiment constructs a domain-specific meta-learning architecture and adopts KL divergence-based distribution matching as the optimization objective. It utilizes meta-networks to automatically generate adaptive learning rules to guide agents in making routing decisions, effectively solving the technical problems of poor adaptability of routing methods and difficulty in balancing multiple objectives in heterogeneous low-Earth orbit satellite networks. This significantly improves the overall performance and intelligence level of the network.

[0087] It should be noted that the various methods described in the embodiments of the present invention can be implemented based on the hardware structure of the electronic device 100. Furthermore, those skilled in the art will understand that the functions of the embodiments of the present invention can be implemented by a processor executing a computer program, which can be stored in a computer-readable storage medium, such as memory 102. This computer program product includes computer instructions, which, when executed by the processor, can implement the methods described in the embodiments of the present invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A routing method for heterogeneous low-Earth orbit satellite networks, characterized in that, include: A domain-based network architecture is constructed, dividing the low-Earth orbit satellite constellation into multiple heterogeneous control domains, each of which is managed by a control node. A hierarchical architecture is established based on the aforementioned heterogeneous control domain; The layered architecture includes a data plane composed of satellite nodes, a control plane composed of control nodes, and a meta-learning layer containing a meta-network. The meta-network is used to generate learning objectives that adapt to different heterogeneous control domains; Establish end-to-end latency and energy consumption models in a heterogeneous low-Earth orbit satellite network environment to quantitatively evaluate the latency and energy consumption of routing. Define a multi-objective optimization function to balance latency and energy consumption; The multi-objective optimization function is expressed as evaluating the KL divergence between the empirical performance distribution and the potential preference distribution, wherein the empirical performance distribution is generated by a routing method guided by the satellite agent, and the potential preference distribution represents an ideal trade-off between latency and energy consumption. Perform meta-learning routing decisions, the meta-learning routing decisions including: In the meta-optimization stage, the satellite agent interaction trajectories collected from multiple heterogeneous control domains are processed by the meta-network to generate adaptive learning targets, and the parameters of the meta-network are updated using the meta-gradient method. During the agent optimization phase, each satellite agent in the heterogeneous control domain receives the learning objective; updates its own routing method network parameters according to the learning objective, and selects the next-hop satellite node to perform routing actions based on the updated routing method.

2. The method according to claim 1, characterized in that, The empirical performance distribution consists of a set of historical performance index vectors collected by the satellite agent during its interaction with the environment. Each historical performance index vector contains observations in two dimensions: latency and energy consumption. The potential preference distribution is a pre-defined target distribution whose shape reflects the current heterogeneous control domain's trade-off preference for the two performance indicators of latency and energy consumption.

3. The method according to claim 2, characterized in that, The potential preference distribution is constructed as a Gaussian mixture model; the mean vector of the Gaussian mixture model corresponds to a preset service quality target anchor point, the service quality target anchor point includes a first anchor point, which represents the performance index combination under the latency-first scenario, and a second anchor point, which represents the performance index combination under the energy-first scenario; the covariance matrix of the Gaussian mixture model reflects the tolerance for deviation from the service quality target anchor point, wherein a smaller covariance value indicates a more stringent performance requirement for the corresponding anchor point.

4. The method according to claim 1, characterized in that, The satellite intelligent agent interaction trajectory includes a preset time window. Within the network, the satellite agent's state sequence, the sequence of routing actions performed, and the sequence of performance metric feedback are included. The meta-network is configured to receive the satellite agent's interaction trajectory as input and output the learning objective, which is a gradient direction vector that guides the satellite agent's own routing method network parameter updates.

5. The method according to claim 4, characterized in that, The meta-network adopts a recurrent neural network architecture based on gated recurrent units to handle the temporal dependencies in the interaction trajectory of the satellite agent. In the meta-optimization stage, the parameters of the meta-network are updated using the meta-gradient method. Specifically, the gradient of the sum of performance improvements generated by the learning objective after guiding the satellite agent to optimize in multiple heterogeneous control environments with respect to the parameters of the meta-network is calculated, and the parameters of the meta-network are updated along the gradient direction.

6. The method according to claim 1, characterized in that, The latency model comprehensively considers the transmission latency, propagation latency, queuing latency, and processing latency of inter-satellite links, as well as the transmission latency, propagation latency, and processing latency of satellite-to-ground links; the energy consumption model comprehensively considers the energy consumption of the communication transceiver module and the routing calculation module of satellite nodes.

7. The method according to claim 1, characterized in that, In the hierarchical architecture, the data plane consists of satellite nodes and is responsible for forwarding data packets; the control plane consists of control nodes in each of the heterogeneous control domains and is responsible for executing the routing methods of the satellite agents within the domain. The meta-learning layer is deployed at ground stations or core gateway nodes and is responsible for the generation and distribution of global learning rules.

8. An electronic device, comprising: A memory, a processor, and a computer program stored on the memory, characterized in that, when the processor executes the program, it implements the method as described in any one of claims 1 to 7.

9. A computer program product comprising computer instructions, characterized in that, When the instructions are executed by the processor, they implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images