Satellite edge computing task offloading method and system

Abstract

The application provides a satellite edge computing task offloading method and system, which comprises the following steps: uploading a computing task to a target low-orbit satellite node for a ground user terminal; obtaining an offloading target node of the computing task based on a deployed first reinforcement learning agent for the target low-orbit satellite node; when the offloading target node is a non-target low-orbit satellite node in a low-orbit satellite node group or a ground server, the non-target low-orbit satellite node cooperatively obtains an offloading path of the computing task based on a deployed second reinforcement learning agent, and offloads the computing task to the offloading target node based on the offloading path. The satellite edge computing task offloading method and system of the application realize satellite edge computing task offloading based on a double-layer reinforcement learning architecture, and realize dual optimization of time delay and energy consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, and in particular to a satellite edge computing task offloading method and system. BACKGROUND

[0002] With the continuous development of satellite communication and on-orbit computing technology, satellite edge computing has become an important direction to improve the intelligent level of non-ground network. Traditional satellite networks mainly undertake data relay and forwarding functions, and the computing power is mainly concentrated on the ground center. However, this centralized computing model faces two challenges: on the one hand, with the continuous growth of the number of ground users and data density, returning all data to the ground center for processing will cause congestion and high delay of the satellite-ground link bandwidth; on the other hand, in some remote or disaster scenarios, the ground center is inaccessible or the response delay is too large, which cannot meet the demand of real-time processing tasks.

[0003] In recent years, satellites have gradually possessed strong on-orbit processing capabilities, making it possible to perform intelligent computing tasks such as data analysis, image recognition, and target detection on the satellite side, which has promoted the proposal and development of the concept of "satellite edge computing". By deploying edge computing resources on satellite nodes, the task return delay can be significantly reduced, the response speed can be improved, and the satellite-ground communication bottleneck can be alleviated. Satellite edge computing makes satellites not only a relay pipeline, but also a computing node with processing capabilities.

[0004] However, the satellite edge computing system still faces many challenges in operation. First, due to the limitations of orbital height and satellite layout, the visible window between low-orbit satellites changes dynamically, and the inter-satellite link (ISL) has intermittency and directionality, and the link delay and stability are affected by the space topology structure. Second, the on-orbit resources of satellites are limited, and the computing power, power, and cache space are all bottlenecks, so how to efficiently allocate tasks to the most suitable nodes becomes a key problem.

[0005] The goal of task scheduling is not only to select a node with processing capabilities, but also to consider multiple constraints and performance indicators. Since the computing operation itself consumes energy, and the transmission of tasks in inter-satellite or satellite-ground links also accompanies transmission energy consumption and communication delay, the scheduling system needs to consider multiple objectives such as computing resource allocation, transmission path selection, energy consumption minimization, and task response delay minimization.

[0006] In addition, different nodes (such as satellite nodes on board and ground stations) also have differences in computing power, processing accuracy, and task adaptability. For example, some high-precision perception or computing tasks may have to be offloaded to the ground high-performance computing center for execution due to the limitations of on-orbit computing power or algorithm models, which puts higher requirements on offloading paths and node selection.

[0007] To address these challenges, some recent research works have attempted to adopt task offloading and collaborative scheduling strategies. Common modes include.

[0008] (1) Tasks are directly sent by ground users to the coverage satellite and completed on the satellite (local processing mode).

[0009] (2) If the task requires extremely high computing resources, or there is no suitable node in the network to process it, the task will be forwarded to the ground station (remote backhaul mode).

[0010] However, existing research generally lacks in-depth exploration of the "inter-satellite offloading" capability. That is, when the current satellite is in a state of resource shortage, whether the task can be offloaded to other reachable satellites in the network for processing, so as to effectively balance the task load, improve resource utilization, and reduce overall energy consumption and communication overhead. This is a key direction that needs to be further explored. Inter-satellite collaborative offloading as an intermediate layer offloading path selection mechanism can significantly alleviate the task hotspot concentration problem and improve the overall scheduling flexibility and resilience of the system.

[0011] In the above modes, the key problem is how to select the optimal offloading target node and the most suitable path to meet the comprehensive performance requirements of the task delay, energy consumption, and processing success rate. Since the system state evolves dynamically over time, including inter-satellite link accessibility, network load distribution, and node state, traditional static scheduling or heuristic algorithms are difficult to adapt to optimal decision-making in a dynamic environment.

[0012] Therefore, some research has begun to introduce reinforcement learning-based scheduling algorithms, enabling satellites to learn task offloading strategies based on historical experience and real-time observations. The reinforcement learning framework continuously optimizes offloading actions through interaction with the environment, and has adaptability and generalization ability in unknown environments. However, most existing work focuses on local scheduling decisions (such as satellite-ground link or intra-satellite routing selection), and the study of end-to-end cross-node offloading path selection for tasks is still insufficient.

[0013] Therefore, how to make dynamic decisions based on system state after the task is uploaded from the ground user to the network, and select the processing node (local or ground station) and offloading path, is one of the key technologies that need to be broken through in the current satellite edge computing system. SUMMARY

[0014] In view of the above problems, the purpose of the present application is to provide a satellite edge computing task offloading method and system, which realizes satellite edge computing task offloading based on a double-layer reinforcement learning framework, and realizes dual optimization of delay and energy consumption.

[0015] In a first aspect, the present application provides a satellite edge computing task offloading method, comprising the following steps: for a ground user terminal, uploading a computing task to a target low-orbit satellite node; for the target low-orbit satellite node, obtaining an offloading target node of the computing task based on a deployed first reinforcement learning agent; when the offloading target node is a non-target low-orbit satellite node in a low-orbit satellite node group or a ground server, the non-target low-orbit satellite node cooperatively obtains an offloading path of the computing task based on a deployed second reinforcement learning agent, so as to offload the computing task to the offloading target node based on the offloading path.

[0016] In an implementation form of the first aspect, the method further comprises: training an offloading policy based on the first reinforcement learning agent to obtain the offloading target node of the computing task based on the offloading policy; and training a path policy based on the second reinforcement learning agent to cooperatively obtain the offloading path of the computing task based on the path policy.

[0017] In an implementation form of the first aspect, training the offloading policy based on the first reinforcement learning agent to obtain the offloading target node of the computing task based on the offloading policy comprises the following steps:

[0018] obtaining a computing task requirement and a low-orbit satellite node state; the low-orbit satellite node state comprises one or more combinations of a cache state, a computing queue, a remaining resource, a remaining energy, a link load, and a neighboring satellite node topology reachability of the low-orbit satellite node; the task requirement comprises a computing delay requirement, a precision requirement, and an energy consumption requirement;

[0019] training the offloading policy based on the task requirement and the low-orbit satellite node state in the first reinforcement learning agent to make the offloading policy make a decision on the offloading target node.

[0020] In an implementation form of the first aspect, training the path policy based on the second reinforcement learning agent to cooperatively obtain the offloading path of the computing task based on the path policy comprises the following steps:

[0021] the second reinforcement learning agents respectively deployed on a plurality of non-target low-orbit satellite nodes in the low-orbit satellite node group form a distributed multi-agent system;

[0022] for the plurality of non-target low-orbit satellite nodes, training the path policy based on a link state in the second reinforcement learning agents to determine a next-hop low-orbit satellite node based on the path policy; wherein the link state comprises one or more combinations of a link delay, a queue length, an energy consumption level, a link reachability, a queuing waiting time, a current load, and a hop limit.

[0023] In one implementation of the first aspect, the first reinforcement learning agent trains the offloading policy using a deep reinforcement learning method with an Actor-Critic architecture, and the second reinforcement learning agent trains the path policy using a multi-agent proximal policy optimization algorithm; the first reinforcement learning agent and the second reinforcement learning agent update the offloading policy and the path policy based on reward function feedback.

[0024] In one implementation of the first aspect, the reward function ,in This represents the normalized task completion delay. Indicates normalized task energy consumption, Indicates the normalized task success rate, This represents the normalized link load balancing degree. This represents the non-negative weight coefficient, and .

[0025] In one implementation of the first aspect, the unloading strategy selects the candidate node with the largest utility function as the unloading target node, wherein the utility function... , Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, This represents the normalized link load balancing degree. Represents the weighting coefficient, and , n Indicates the first n There are 10 candidate nodes; the weight coefficients are set according to the task type of the computation task, which includes latency-sensitive, energy-sensitive and computationally intensive tasks.

[0026] In one implementation of the first aspect, the path strategy is further used to trigger a global path reconstruction mechanism when low-Earth orbit satellite node links become congested; and to automatically trigger the nearest available low-Earth orbit satellite node for offloading path reconstruction when the computing task forwarding fails.

[0027] In one implementation of the first aspect, the path strategy is further used to plan the unloading path by combining the link prediction state obtained from the trajectory prediction model, and the reward function used is... ,in Delay Indicates task completion delay, Energy Indicates normalized task energy consumption, Success This represents the normalized task success rate. Indicates the probability that the link will be available in the future. The value represents the weighting coefficient, and the probability of future link availability is obtained from the trajectory prediction model.

[0028] In a second aspect, the present application provides a satellite edge computing task offloading system, comprising a ground user terminal, a group of low-orbit satellite nodes, and a ground server.

[0029] The ground user terminal is configured to upload a computing task to a target low-orbit satellite node.

[0030] The target low-orbit satellite node in the group of low-orbit satellite nodes is configured to obtain an offloading target node of the computing task based on a deployed first reinforcement learning agent.

[0031] When the offloading target node is a non-target low-orbit satellite node in the group of low-orbit satellite nodes or the ground server, the non-target low-orbit satellite node is configured to cooperatively obtain an offloading path of the computing task based on a deployed second reinforcement learning agent, and offload the computing task to the offloading target node based on the offloading path.

[0032] As described above, the satellite edge computing task offloading method and system of the present application have the following beneficial effects.

[0033] (1) By introducing an "inter-satellite cooperative offloading" mechanism, the hierarchical agent jointly decides the task processing node and the offloading path, effectively improving the system resource utilization and scheduling flexibility, and ultimately achieving dual optimization of energy consumption and time delay.

[0034] (2) The task processing node selection and offloading path planning are effectively optimized cooperatively. Compared with the traditional static scheduling mechanism, it can adapt to the dynamic topology and load change environment in real time, significantly reducing the overall task processing delay and transmission energy consumption.

[0035] (3) The system load balancing and fault tolerance are enhanced, making the satellite edge computing have stronger scalability and robustness, and being suitable for intelligent task scheduling requirements under large-scale constellation, providing intelligent offloading strategies for future large-scale satellite internet scenarios.

[0036] (4) Through the distributed deployment and local autonomous learning mechanism, it has good scalability and robustness, and is suitable for future large-scale satellite constellation and task-intensive space computing scenarios, and has broad practical application prospects and industrialization value. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A flowchart showing the satellite edge computing task offloading method of the present application in an embodiment.

[0038] Figure 2 A frame diagram showing the satellite edge computing task offloading method of the present application in an embodiment.

[0039] Figure 3 The diagram shown is an architectural schematic of a two-layer reinforcement learning task scheduling method in one embodiment of the present invention.

[0040] Figure 4 The diagram shown is a structural schematic of a satellite edge computing task offloading system according to an embodiment of the present invention. Detailed Implementation

[0041] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0042] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0043] The satellite edge computing task offloading method and system of this invention starts with the uploading of computing tasks from ground user terminals, allowing computing tasks to dynamically select the optimal strategy among three methods: on-board local processing, inter-satellite forwarding processing, or back-to-ground server processing. To balance scheduling flexibility and global performance, the satellite edge computing task offloading method and system of this invention adopts a two-layer architecture: the upper layer is a single agent that selects offloading nodes, and the lower layer is a multi-agent collaborative optimization of the offloading path. This achieves comprehensive optimization of task energy consumption, communication latency, and computing resources, realizing intelligent closed-loop control of terminal tasks from uploading, offloading target selection, routing forwarding to completion feedback. This significantly improves the overall computing task processing efficiency, reduces energy consumption, and enhances the system's adaptability to dynamic network structures.

[0044] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] like Figure 1 As shown, in one embodiment, the satellite edge computing task offloading method of the present invention includes steps S1-S3.

[0046] Step S1: For ground user terminals, upload the computing task to the target low-Earth orbit satellite node.

[0047] Specifically, in this invention, the ground user terminal generates a computing task, which uploads the computing task to the currently connected low-Earth orbit satellite node within the coverage area, i.e., the target low-Earth orbit satellite node.

[0048] Step S2: For the target low-Earth orbit satellite node, obtain the offloading target node of the computing task based on the deployed first reinforcement learning agent.

[0049] Specifically, the low-Earth orbit (LEO) satellite node cluster comprises multiple LEO satellite nodes. All LEO satellite nodes communicate with the target LEO satellite node via an InterSatellite Link (ISL) to establish a dynamic inter-satellite communication topology, and communicate with a ground server via a radio frequency (RF) link. The ground server possesses strong computing power. Each LEO satellite node is equipped with a first reinforcement learning agent. After training as a single agent, this agent is capable of making decisions regarding unloading the target node based on current mission requirements and the status of each LEO satellite node. The unloading target node can be the target LEO satellite node, other non-target LEO satellite nodes, or the ground server.

[0050] In this invention, for the target low-Earth orbit satellite node, after receiving the computing task, a decision is made to unload the target node based on the first reinforcement learning agent deployed on it.

[0051] Step S3: When the unloading target node is a non-target LEO satellite node or a ground server in the LEO satellite node group, the non-target LEO satellite node obtains the unloading path of the computing task based on the deployed second reinforcement learning agent, so as to unload the computing task to the unloading target node based on the unloading path.

[0052] Specifically, a second reinforcement learning agent is also deployed in each low-Earth orbit satellite node. After being trained in a multi-agent manner, the second reinforcement learning agent can collaborate with other second reinforcement learning agents to construct an offloading path, so that the computing task is offloaded to the offloading target node via the offloading path, thereby completing the offloading of the satellite edge computing task.

[0053] like Figure 2 As shown, after the computation task is uploaded from the ground user terminal, the first reinforcement learning agent on the target low-Earth orbit satellite node in the upper layer decides the unloading target (local, other satellites or ground server), and then multiple second reinforcement learning agents in the lower layer select the unloading path. Figure 2 It clearly describes the transmission path of computing tasks in the network and the interaction logic between decision nodes.

[0054] In the present application, the first reinforcement learning agent and the second reinforcement learning agent need to be trained to make decisions on the offloading target node and the offloading path, respectively. The offloading strategy is trained based on the first reinforcement learning agent to obtain the offloading target node of the computing task based on the offloading strategy. The path strategy is trained based on the second reinforcement learning agent to obtain the offloading path of the computing task based on the path strategy.

[0055] In an embodiment, training the offloading strategy based on the first reinforcement learning agent to obtain the offloading target node of the computing task based on the offloading strategy comprises the following steps.

[0056] a) Obtain the computing task requirements and the low-orbit satellite node state; the low-orbit satellite node state includes one or more combinations of the cache state, the computing queue, the remaining resources, the remaining energy, the link load, and the adjacent satellite node topology accessibility of the low-orbit satellite node; the task requirements include the computing delay requirement, the accuracy requirement, and the energy consumption requirement.

[0057] b) Based on the task requirements and the low-orbit satellite node state, the first reinforcement learning agent trains the offloading strategy to make decisions on the offloading target node.

[0058] Preferably, the low-orbit satellite node state is normalized before training the offloading strategy, thereby improving the stability and generalization ability in a dynamic environment.

[0059] In an embodiment, training the path strategy based on the second reinforcement learning agent to obtain the offloading path of the computing task based on the path strategy comprises the following steps.

[0060] a) The second reinforcement learning agents respectively deployed on a plurality of non-target low-orbit satellite nodes in the low-orbit satellite node group constitute a distributed multi-agent system.

[0061] b) For the plurality of non-target low-orbit satellite nodes, the path strategy is trained in the second reinforcement learning agent coordination based on the link state to determine the next-hop low-orbit satellite node based on the path strategy, thereby obtaining the offloading path from the target low-orbit satellite node to the offloading target node. The link state includes one or more combinations of link delay, queue length, energy consumption level, link accessibility, queuing waiting time, current load, and hop limit.

[0062] The second learning agent trains the path policy using the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm.

[0063] The second reinforcement learning agent feeds back the training results to the first reinforcement learning agent. The first reinforcement learning agent constructs a reward function based on its own training results and updates the unloading strategy and path strategy of both the first and second reinforcement learning agents based on the reward function, i.e., updating the model parameters. The reward function is based on weighted task completion latency, task energy consumption, task success rate, and link load balancing. In one embodiment, the reward function... ,in Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, This represents the normalized link load balancing degree. This represents the non-negative weight coefficient, and To ensure the task completion time is within a certain timeframe. D Task energy consumption E Task success rate S and link load balancing L To enable comparisons within the same dimension, the aforementioned indicators were first normalized. The non-negative weighting coefficients reflect the importance of different indicators in the overall optimization objective; for example, they can be appropriately increased in low-latency scenarios. In energy-constrained scenarios, it can improve The negative sign indicates that the optimization direction for latency and energy consumption is to minimize, while the optimization direction for success rate and load balance is to maximize. Through this weighted mechanism, multi-objective unified optimization of offloading and path strategies can be achieved, ensuring that the first and second reinforcement learning agents take into account latency, energy consumption, success rate, and load balance during training.

[0064] like Figure 3 As shown in the figure, a low-Earth orbit satellite node with computing and communication capabilities is represented by the symbol "". o "This represents the observation information that the intelligent agent obtains from the environment, including node status, link status, and task requirements, etc." a "Represents the decision action (Action) output by the intelligent agent." r"The reward value is calculated based on the action feedback, combined with energy consumption, latency, and task completion status, and is used to optimize the strategy. The task scheduling process is divided into an upper-layer offloading target selection stage and a lower-layer routing path optimization stage, corresponding to the upper-layer reinforcement learning agent and the lower-layer multi-agent collaborative learning system shown in the figure, respectively. After the task arrives at the satellite node, the upper-layer agent shown at the top of the figure first receives the state information input. This agent generates an offloading decision based on the current network environment state, deciding whether the task should be executed locally, forwarded to other satellite nodes via inter-satellite links, or directly downlinked to the ground server. This decision-making process is a closed loop of "state → policy network → offloading action". If the task chooses a non-local processing path, it enters the lower-layer collaborative path optimization stage, which is driven by multiple lower-layer satellites shown at the bottom of the figure." A distributed system composed of star node agents participates in the process. Each agent independently perceives its local link status, cache status, and neighboring node status through a reinforcement learning architecture, and makes next-hop path decisions based on the globally shared state, ultimately forming a multi-hop forwarding path that meets transmission latency and energy consumption constraints. This process is manifested as "local observation → policy network → next-hop action output". The entire scheduling process relies on the feedback mechanism provided by the environment module to complete the training loop: when a task is completed or fails, the system obtains reward information from the environment, including but not limited to task completion latency, forwarding node energy consumption, link utilization, etc., and feeds it back to the two levels of agents to update their policy network and value network. Therefore, the satellite edge computing task offloading method of this invention is particularly suitable for task-intensive edge computing application scenarios in large-scale dynamic satellite networks.

[0065] In one embodiment, the offloading strategy determines the weights of the computational tasks to be offloaded to local, non-target LEO satellite nodes, or ground servers based on task type. The task types include latency-sensitive, energy-sensitive, and computationally intensive tasks. It also supports heterogeneous node capability modeling, including a matching mechanism between task requirements and node inference accuracy and processing capabilities, ensuring that high-precision tasks are preferentially processed on ground servers, while ordinary tasks are offloaded to nearby LEO satellite nodes. Specifically, in the offloading strategy, the candidate node with the highest utility function is selected as the offloading target node, where the utility function... , Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, This represents the normalized link load balancing degree. This represents the weighting coefficient, which is set according to the task type (latency-sensitive, energy-sensitive, computationally intensive). , n Indicates the first n There are candidate nodes. Therefore, the target node for unloading is... Different types of tasks are differentiated by adjusting the weights, for example, time-sensitive tasks increase , energy-sensitive tasks increase , and compute-intensive tasks maintain the weight distribution of success rate and balance.

[0066] The global path reconstruction mechanism is triggered when the link of the low-orbit satellite node is congested. When the computing task forwarding fails, the path reconstruction is automatically triggered to unload the nearest available low-orbit satellite node. Preferably, the path strategy can also plan the unloading path in combination with the link prediction state obtained by the orbit prediction model. Specifically, the future link availability / delay / prediction quantity obtained by the orbit prediction model is incorporated into the second intelligent learning agent as an additional feature. The reward function imposes a penalty on the future interruption risk, encouraging the selection of more stable links within the prediction window. For example, the reward function wherein Delay represents the task completion delay, Energy represents the normalized task energy consumption, Success represents the normalized task success rate, represents the future link availability probability, represents the weight coefficient. The future link availability probability is obtained by the orbit prediction model.

[0067] Meanwhile, the present application also has a strategy hot switching function, which can load a pre-trained strategy to ensure the continuity of task scheduling when detecting that the network topology change reaches a threshold. The strategy hot switching is essentially a fault-tolerant and rapid recovery mechanism. When the network topology changes dramatically (e.g., multiple satellites fail or large-scale link interruption), the reinforcement learning agent of online learning needs a certain time to re-converge, which may cause task interruption. To avoid this problem, when the topology change exceeds the threshold, the system will call and load a pre-trained strategy parameter that has been trained in a similar topology from the strategy library. In this way, the following effects can be achieved: first, replace the parameters of the current agent (such as Actor / Critic network weights) to make it immediately have stable scheduling capability; second, ensure that the task scheduling process does not interrupt, achieving so-called "hot switching"; third, the system can continue to fine-tune or retrain under the new topology to gradually restore optimal performance. The role of loading the pre-trained strategy is to quickly switch to a set of existing reinforcement learning model parameters when an anomaly is detected, which is equivalent to "using existing experience in emergency", avoiding the cold start delay of online learning.

[0068] As shown in Figure 4 , in an embodiment, the satellite edge computing task offloading system of the present application includes a ground user terminal 41, a low-orbit satellite node group 42, and a ground server 43.

[0069] The ground user terminal 41 is configured to upload a computing task to a target low-orbit satellite node.

[0070] The target low-orbit satellite node in the low-orbit satellite node group 42 is configured to obtain an offloading target node of the computing task based on a first reinforcement learning agent.

[0071] When the offloading target node is a non-target low-orbit satellite node in the low-orbit satellite node group or the ground server 43, the non-target low-orbit satellite node is configured to cooperatively obtain an offloading path of the computing task based on a second reinforcement learning agent, and offload the computing task to the offloading target node based on the offloading path.

[0072] The above embodiments only illustrate the principles and effects of the present application, and are not intended to limit the present application. Any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical ideas disclosed by the present application should be covered by the claims of the present application.

Claims

1. A method for offloading satellite edge computing tasks, characterized in that: The method includes the following steps: For ground user terminals, the computing tasks are uploaded to the target low-Earth orbit satellite node; For the target low-Earth orbit satellite node, the offloading target node for the computing task is obtained based on the deployed first reinforcement learning agent; When the unloading target node is a non-target LEO satellite node or a ground server in the LEO satellite node group, the non-target LEO satellite node collaboratively obtains the unloading path of the computing task based on the deployed second reinforcement learning agent, so as to unload the computing task to the unloading target node based on the unloading path; wherein, each second reinforcement learning agent independently perceives its local link status, cache status and adjacent node status through reinforcement learning architecture, and makes next-hop path decision based on the globally shared status, forming a multi-hop forwarding path that satisfies the constraints of transmission latency and energy consumption; It also includes training an unloading strategy based on a first reinforcement learning agent to obtain the unloading target node of the computing task based on the unloading strategy; and training a path strategy based on a second reinforcement learning agent to collaboratively obtain the unloading path of the computing task based on the path strategy. The unloading strategy selects the candidate node with the highest utility function as the unloading target node, where the utility function... Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, The normalized link load balancing degree is represented by α, β, γ, and δ, which represent weight coefficients, and α+β+γ+δ=1, where n represents the nth candidate node. The weight coefficients are set according to the task type of the computation task, which includes latency-sensitive, energy-sensitive, and computationally intensive tasks.

2. The satellite edge computing task offloading method according to claim 1, characterized in that: The process of training an unloading strategy based on a first reinforcement learning agent to obtain the unloading target node for a computational task includes the following steps: Obtain the computation task requirements and the status of low-Earth orbit (LEO) satellite nodes; the LEO satellite node status includes one or more combinations of the LEO satellite node's cache status, computation queue, remaining resources, remaining energy, link load, and topological reachability of neighboring satellite nodes; the task requirements include computation latency requirements, accuracy requirements, and energy consumption requirements; Based on the task requirements and the state of the low-Earth orbit satellite nodes, the offloading strategy is trained in the first reinforcement learning agent so that the offloading strategy makes decisions on offloading target nodes.

3. The satellite edge computing task offloading method according to claim 1, characterized in that: The process of collaboratively obtaining the unloading path for computational tasks based on the training path strategy of the second reinforcement learning agent includes the following steps: The second reinforcement learning agents deployed on each of the multiple non-target low-Earth orbit satellite nodes in the low-Earth orbit satellite node group constitute a distributed multi-agent system. For the plurality of non-target LEO satellite nodes, the path policy is trained in the second reinforcement learning agent collaboration based on the link state, so as to determine the LEO satellite node for the next hop based on the path policy; wherein the link state includes one or more combinations of link latency, queue length, energy consumption level, link reachability, queuing waiting time, current load and hop count limit.

4. The satellite edge computing task offloading method according to claim 1, characterized in that: The first reinforcement learning agent trains the offloading policy using a deep reinforcement learning method with an Actor-Critic architecture, and the second reinforcement learning agent trains the path policy using a multi-agent proximal policy optimization algorithm. The first reinforcement learning agent and the second reinforcement learning agent update the unloading policy and the path policy based on the reward function feedback.

5. The satellite edge computing task offloading method according to claim 4, characterized in that: The reward function in Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, Represents the normalized link load balancing degree, α, β, γ and δ represent non-negative weighting coefficients, and α+β+γ+δ=1.

6. The satellite edge computing task offloading method according to claim 1, characterized in that: The path strategy is also used to trigger a global path reconstruction mechanism when low-Earth orbit satellite node links become congested; and to automatically trigger the nearest available low-Earth orbit satellite node for offloading path reconstruction when the computing task forwarding fails.

7. The satellite edge computing task offloading method according to claim 1, characterized in that: The path strategy is also used to plan the unloading path by combining the link prediction state obtained from the track prediction model, and the reward function used is r = -α·Delay-β·Energy-ξ·(1-AvailProb future )+γ·Success, where Delay represents the task completion delay, Energy represents the normalized task energy consumption, Success represents the normalized task success rate, and AvailProb future The probability of future link availability is represented by α, β, γ, and ξ, which are weighting coefficients. The probability of future link availability is obtained from the orbit prediction model.

8. A satellite edge computing task offloading system, characterized in that: The system includes ground user terminals, a low-orbit satellite node cluster, and a ground server. The ground user terminal is used to upload computing tasks to the target low-Earth orbit satellite node; The target low-Earth orbit satellite nodes in the low-Earth orbit satellite node group are used to obtain the offloading target nodes of the computing task based on the deployed first reinforcement learning agent. When the unloading target node is a non-target LEO satellite node in the LEO satellite node group or the ground server, the non-target LEO satellite node is used to collaboratively obtain the unloading path of the computing task based on the deployed second reinforcement learning agent, so as to unload the computing task to the unloading target node based on the unloading path; wherein, each second reinforcement learning agent independently perceives its local link status, cache status and adjacent node status through the reinforcement learning architecture, and makes the next-hop path decision based on the globally shared status, forming a multi-hop forwarding path that satisfies the constraints of transmission latency and energy consumption; It also includes training an unloading strategy based on a first reinforcement learning agent to obtain the unloading target node of the computation task based on the unloading strategy; The second reinforcement learning agent is used to train a path strategy to collaboratively obtain the unloading path of the computation task based on the path strategy. The unloading strategy selects the candidate node with the highest utility function as the unloading target node, where the utility function... Indicates the normalized task completion delay, Indicates normalized task energy consumption, Indicates the normalized task success rate, The normalized link load balancing degree is represented by α, β, γ, and δ, which represent weight coefficients, and α+β+γ+δ=1, where n represents the nth candidate node. The weight coefficients are set according to the task type of the computation task, which includes latency-sensitive, energy-sensitive, and computationally intensive tasks.

Images