An inter-satellite cooperative task offloading and resource allocation method for a low earth orbit satellite dynamic environment

Abstract

The application provides an intersatellite cooperation task unloading and resource allocation method for a low-orbit satellite dynamic environment. The method first constructs a spaceborne edge computing scene composed of ground users and service satellites, perceives task, topology, link and computing resource states, determines a main star and a candidate auxiliary star set for a user, and establishes a task allocation proportion representation. Secondly, the cooperation unloading process is modeled as a Markov decision process, a discrete cooperation mode is screened through a deep Q network, and a candidate action set is generated. Thirdly, for each candidate action, a grey wolf optimization algorithm is used to continuously optimize the task allocation proportion, and an evaluation function is combined to select the final execution strategy. Finally, the main star completes task splitting, cooperative distribution and result aggregation feedback, which can effectively solve the problems of rapid change of satellite topology, limited computing power of a single satellite, difficulty in selecting a cooperation node and difficulty in optimizing the allocation proportion, reduce the task processing delay, and improve the resource utilization and system load balancing.

Description

Technical Field

[0001] This invention relates to the fields of satellite communication and spaceborne edge computing technology, and in particular to an inter-satellite cooperative offloading method based on DQN-GWO for dynamic environments of low-Earth orbit satellite networks. Background Technology

[0002] With the continuous development of low-Earth orbit (LEO) satellite networks, onboard computing platforms, and integrated space-air-ground information systems, the types of services carried by satellite communication systems have gradually expanded from traditional voice transmission, telemetry and remote control, and data forwarding to various task-intensive scenarios such as remote sensing information processing, emergency communication support, distributed intelligent sensing, vehicle-to-everything (V2X) access, and information services for marine and remote areas. In these scenarios, user services not only place high demands on network coverage continuity but also on task processing latency, computational response speed, and service stability. Especially in LEO satellite systems, due to the low orbit, high speed, and short transit period of satellites, the visibility relationship, communication distance, and link status between satellites and users change rapidly over time, resulting in significant dynamic topology characteristics in satellite networks. Simultaneously, influenced by factors such as free-space propagation loss, the Doppler effect, link switching, service load fluctuations, and differences in available computing power among nodes, both communication and computing resources in satellite networks exhibit strong time-varying and uncertainties. Faced with these characteristics, how to efficiently schedule and rationally allocate user tasks under the constraints of limited onboard resources has become one of the key technical challenges in the fields of satellite communication and onboard edge computing.

[0003] The development of edge computing technology provides a new technical approach to enhance task processing capabilities in satellite networks. By moving some computing power to network edge nodes, user-generated tasks can be processed closer to the service source, thereby reducing transmission latency caused by long-distance backhaul and alleviating the centralized processing pressure on the core network. In satellite network scenarios, satellite nodes naturally possess characteristics such as wide coverage and proximity to the user access side, thus they can serve as onboard edge computing nodes to undertake some service processing tasks. In existing technologies, a common approach is for users to upload tasks to the access satellite, which then uses its local computing resources to complete the task processing. However, the available computing power, storage capacity, and communication bandwidth of a single access satellite are relatively limited. When multiple user tasks arrive simultaneously, the processing load of local hotspot satellites can easily increase rapidly, leading to problems such as increased task queuing, increased processing latency, and decreased service quality. Conversely, some neighboring satellites in the same area may still have some idle computing resources, but due to the lack of an effective collaborative scheduling mechanism, these dispersed resources cannot participate in task processing in a timely manner, resulting in low overall system resource utilization and difficulty in simultaneously meeting latency control and load balancing requirements.

[0004] To address these challenges, researchers have begun exploring multi-satellite collaborative processing mechanisms in satellite networks. After user tasks are uploaded to an access satellite, the access satellite acts as the master control node, further transferring some tasks to neighboring satellites for collaborative processing based on the current network topology, link conditions, and satellite resource status. This type of method typically establishes a collaborative relationship between the master and auxiliary satellites, distributing the processing load originally concentrated on a single satellite to multiple satellite nodes, thereby improving overall task capacity and resource reuse efficiency. In this process, the master satellite usually needs to determine a set of candidate satellites that can participate in collaboration based on user task requirements and the status of surrounding satellites. Then, it selects auxiliary satellite nodes from the candidate satellites to actually participate in task processing, dividing the user task into multiple sub-tasks and allocating them to the master and auxiliary satellites according to a certain ratio. After each participating satellite completes its assigned sub-task, the processing results are returned and aggregated for output. Inter-satellite collaborative task processing is not simply a matter of multiple nodes executing in parallel; it involves a joint decision-making process encompassing multiple stages, including selecting collaborative nodes, determining task splitting ratios, resource scheduling, and result aggregation. Its implementation difficulty is significantly higher than that of single-satellite local processing.

[0005] On the other hand, there are still significant shortcomings in practical applications for dynamic low-Earth orbit (LEO) satellites. First, the selection of a primary satellite from a secondary satellite is a discrete decision problem, while the task proportion undertaken by each participating satellite is a continuous optimization problem. These two problems are coupled, causing the decision space to expand rapidly with the increase in the number of users and candidate satellites. Traditional exhaustive methods or fixed-rule methods are insufficient to meet the real-time solution requirements in dynamic environments. Second, in high-speed LEO satellite scenarios, link rates, visibility relationships, and satellite load states all continuously change. If task allocation is based solely on static thresholds or a single heuristic rule, it is usually difficult to simultaneously achieve reduced total latency, improved collaboration efficiency, and balanced system load. Third, some existing solutions focus on single-objective optimization and lack fine-grained control over the degree of task migration between primary satellite local processing and secondary satellite collaborative processing. Therefore, it is difficult to form stable, effective, and adaptive collaboration strategies in complex, time-varying scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide an inter-satellite cooperative task offloading and resource allocation method for dynamic low-Earth orbit satellite environments. This method addresses key technical problems in existing satellite-based edge computing scenarios, such as high task processing latency, low resource utilization, and uneven load distribution among satellites, caused by rapid changes in satellite topology, limited single-satellite processing capacity, difficulty in selecting cooperative nodes, and challenges in dynamically optimizing task allocation ratios. By introducing a primary-secondary satellite hierarchical cooperative mechanism and combining it with deep reinforcement learning and continuous resource optimization methods, the method achieves joint optimization of the inter-satellite cooperative mode selection and task allocation process. This improves the overall task processing efficiency of the system, reduces task completion latency, and enhances the collaborative processing and adaptive capabilities of the satellite network in highly dynamic environments.

[0007] This invention proposes a comprehensive solution combining a primary-secondary satellite hierarchical collaboration mechanism, a deep reinforcement learning-based collaboration mode selection method, and a continuous task ratio optimization method. This solution dynamically senses the current satellite network status, including user task requirements, link transmission rates between the primary and candidate secondary satellites, and available computing resources for each satellite. Based on this status information, it generates multiple candidate inter-satellite collaboration modes. Subsequently, for each candidate collaboration mode, it continuously optimizes the task allocation ratio between the primary and secondary satellites to obtain the corresponding task splitting results. Finally, it comprehensively evaluates the latency and collaboration costs under different candidate modes, selects the optimal collaboration mode and its corresponding task allocation scheme, and generates the final inter-satellite collaboration offloading strategy. This achieves efficient joint optimization of the task offloading and resource allocation process while ensuring that the system's collaborative processing capabilities and resource constraints are met.

[0008] The method in this embodiment of the invention constructs a hierarchical collaborative architecture between a primary satellite and auxiliary satellites for dynamic low-Earth orbit satellite networks. In this architecture, user tasks first access the primary satellite, which then determines whether to initiate auxiliary satellite collaboration based on the current topology, link status, and auxiliary satellite resource availability, and further completes task decomposition and collaborative processing. Based on this architecture, the primary satellite undertakes both user-side access and global control functions, and also acts as a local computing node to process some tasks. Auxiliary satellites participate in task sub-block processing under the scheduling of the primary satellite, thereby forming a collaborative computing link between users, the primary satellite, and auxiliary satellites.

[0009] In this embodiment of the invention, the system consists of a set of ground users and a set of service satellites. For any user i, let its ground location be denoted as . The ground projection position of the j-th satellite is If the satellite's orbital altitude is H, then the three-dimensional link distance between the user and the satellite is expressed as: ; The satellite with the closest distance to the user is selected as the user's primary access satellite, denoted as . ; Here, S represents the set of satellites in the system. Further, the satellites are sorted from closest to furthest from the user, and the closest satellites are selected to form a candidate auxiliary satellite set. This completes the determination of the primary star access relationship and the range of candidate auxiliary stars.

[0010] In this architecture, after a user task is uploaded to the primary satellite, it can be processed independently by the primary satellite, or the processing load can be shared by the primary satellite and auxiliary satellites. To uniformly describe these two processing methods, the task data volume of user i is defined as D_i, and its task allocation ratio vector on the service satellite set is represented as... .in, This indicates the proportion of tasks that are processed locally on the primary planet. This represents the proportion of tasks assigned to the k-th auxiliary star, and satisfies... ; when When, it means the task is handled entirely independently by the main star; when When this occurs, it indicates that the primary satellite will transfer some tasks to the secondary satellites for collaborative processing. Based on this proportional vector, the task chunk size of the user task on each participating satellite can be expressed as: ; Simultaneously, the primary star combines its local state and the states of candidate auxiliary stars to form local observation information related to collaboration, and determines whether to initiate auxiliary star collaboration and the task proportion undertaken by each auxiliary star. After the task is split, the primary star activates the set of auxiliary stars actually participating in collaboration according to the allocation vector. The actual collaborative auxiliary star set for user i is defined as: ; The set of services participating in user i's task processing within the current time slot can be represented as: ; After establishing the primary-secondary satellite cooperation relationship, this invention further employs a DQN-GWO joint mechanism to implement inter-satellite cooperation offloading decisions. The inter-satellite cooperation offloading process is modeled as a Markov decision process, where the system state in time slot t is composed of both task requirements and link state. For the i-th user, its state characteristics consist of task data volume, uplink transmission rate, and downlink transmission rate; therefore, the global system state is represented as follows: ; in, and Let N represent the uplink and downlink transmission rates between user i and the access master satellite, respectively, and let N represent the number of users in the system.

[0011] After obtaining the status information, DQN performs initial screening of discrete cooperation modes. The system pre-sets a finite number of discrete cooperation actions for each user, where each discrete action represents the cooperation mode currently used by the task. The continuous allocation ratio of the task between the primary and secondary satellites is further solved after the candidate actions are determined. The combined discrete actions of all users form the joint action for the current time slot, represented as follows: .

[0012] Simultaneously, combined with the given current state DQN outputs the value estimation of each discrete action. Where θ represents the network parameters. In terms of network architecture, DQN employs a multi-layer fully connected network for forward propagation and combines it with an ε-greedy mechanism to generate candidate actions. When the random variable is less than the preset exploration probability ε, several actions are randomly selected from the action space as candidates; otherwise, the top K actions are selected based on their Q-values ​​to form a candidate set. ; ; After obtaining the set of candidate actions, for any candidate action Further determination of the continuous allocation ratio of user tasks between primary and secondary satellites is needed. In this embodiment of the invention, the GWO algorithm is used to continuously optimize the ratio vector under a given cooperative mode for candidate actions. First, an initial scoring factor is constructed based on the current discrete action, the set of participating satellites, and the computing power of each satellite. The computing power of the k-th node participating in user i's task processing is defined as... The accessibility item corresponding to the user's distance from their ground projection is: Then the initial rating factor can be expressed as ; in, and These represent the weights of the computational capability term and the distance term, respectively. This represents the primary star bias term introduced by discrete actions. Based on the above scoring factors, the initial seed scale vector is obtained through Softmax mapping. ; Here, τ represents the temperature parameter, which can be adjusted and random perturbations introduced. The result is then projected onto the probabilistic simplex space to construct the initial population for the GWO algorithm. This generates multiple feasible task splitting schemes under candidate action constraints.

[0013] After obtaining the initial population, the proportion vector is iteratively updated using GWO. Let the position of the nth gray wolf in the rth iteration be... The three gray wolves with the best fitness in the current population correspond to α, β, and δ, respectively, and their position vectors are respectively , and Then for any dimensional component... ; ; ; ; Where A is the shrinkage coefficient that decreases linearly with each iteration, and C is the random search parameter. After each update, the new position vector is reprojected onto the probabilistic simplex space to ensure that the proportion vector satisfies the non-negativity and normalization constraints. In this way, GWO continuously searches for the task allocation proportion in a cooperative mode where candidate actions are determined.

[0014] After updating the location, an evaluation function based on Lyapunov optimization is used to assess the merits of each candidate action. The evaluation function comprehensively considers task completion latency and the degree of collaborative migration. For user i, its scoring function is expressed as follows: ; in, This indicates the total task completion delay under the current proportional allocation. Let λ represent the migration ratio, and λ represent the migration penalty coefficient. As shown in the formula above, if a certain ratio vector can reduce the total latency, the score will increase, indicating a better quality candidate decision. However, if there is excessive reliance on auxiliary stars, leading to an excessively large migration ratio, the penalty term will inhibit the score from further increasing. Therefore, the GWO search process essentially seeks a set of optimal solutions among multiple feasible ratio vectors that balance latency performance and cooperation costs, rather than simply pursuing the goal of migrating as many tasks as possible to auxiliary stars. After obtaining the optimal proportion vector of all users corresponding to a certain candidate action, the comprehensive evaluation value of the candidate mode can be obtained by summing the scores of all users under the current candidate collaboration mode. Define candidate actions. The optimal ratio for the i-th user is Then its overall score is defined as ; The final action is represented as ; After obtaining the final action, the system executes the cooperative offloading decision for the current time slot, receives an immediate reward, the environment transitions to the next state, forming experience samples and storing them in the replay pool. When training with randomly selected mini-batches of samples, the target Q-value is represented as... ; Where γ represents the discount factor. The loss function of DQN is expressed as: ; Through the above methods, the DQN network parameters are continuously updated, thereby enabling online adaptation to dynamic satellite network environments.

[0015] Once the final collaboration mode and corresponding optimal task allocation ratio are determined, the primary satellite will split the user task into a local processing component and a secondary satellite collaborative processing component, and distribute them to the primary satellite and each activated secondary satellite for computation. After the primary satellite's local subtask and each secondary satellite's subtask are completed, the secondary satellites will return their results to the primary satellite, which will then aggregate them to form a user task result set and provide feedback to the user. Simultaneously, to evaluate the impact of inter-satellite collaboration on the overall system resource utilization, a satellite load model is constructed. The total task load undertaken by the j-th satellite in the current time slot is denoted as […]. It can be approximated as the total amount of mission data allocated to the satellite. ; in, This represents the proportion of user i's tasks assigned to satellite j. If satellite j does not participate in user i's task processing, then... After obtaining the payload of each satellite, the Jain Fairness Index is used to evaluate the load balancing of the system. ; Where M represents the total number of satellites in the system. Thus, this invention, while completing the offloading decision in the order of mission arrival—primary satellite sensing—DQN candidate pattern generation—GWO ratio optimization—primary and secondary satellite collaborative execution—result aggregation and return, can also uniformly evaluate mission latency and load balancing effects, thereby achieving efficient inter-satellite collaborative computation offloading for the dynamic environment of low-Earth orbit satellites. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0017] Figure 1 This is a schematic diagram of the hierarchical cooperative offloading architecture of the primary star and secondary star in an embodiment of the present invention; Figure 2 This is a schematic diagram of the inter-satellite cooperative task offloading and resource allocation method based on DQN-GWO in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the overall process of the inter-satellite collaborative task unloading and resource allocation method in an embodiment of the present invention. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] Figure 1 A schematic diagram of the master-sat-auxiliary star hierarchical collaborative offloading architecture, specifically including: This invention provides a hierarchical collaborative offloading architecture between primary and secondary satellites suitable for dynamic network environments of low-Earth orbit (LEO) satellites. In this architecture, ground users, the primary satellite, and candidate secondary satellites jointly constitute an inter-satellite collaborative computing offloading system. User tasks are first uploaded to the primary satellite, which then determines whether to initiate secondary satellite collaboration based on current task requirements, link status, and secondary satellite resource status, further completing task splitting and distribution. Due to the high-speed motion of LEO satellites, satellite topology, link quality, and available collaborative resources continuously change. Traditional methods relying solely on local processing by a single satellite or static rule allocation are insufficient to balance task processing latency, resource utilization efficiency, and satellite load balancing performance. Therefore, it is necessary to construct a hierarchical collaborative mechanism between primary and secondary satellites for dynamic environments.

[0021] In this architecture, the user's access master satellite and candidate auxiliary satellite range are first determined based on the spatial relationship between the user and the satellites. Specifically, the three-dimensional link distance between the user and each satellite is calculated according to formula (1), and the satellite with the smallest distance is selected as the user's access master satellite according to formula (2). Then, the remaining satellites are sorted from near to far to construct a candidate auxiliary satellite set. When the master satellite receives the user's task, the task allocation ratio vector between the master satellite and each auxiliary satellite is defined according to formula (3), and the task block amount of the user's task on the master satellite and auxiliary satellites is obtained according to formula (4). Subsequently, the master satellite forms local observation information for the current task scheduling by aggregating the local state and the candidate auxiliary satellite states, and then determines the set of auxiliary satellites actually participating in task processing and the actual service set according to formulas (5) and (6). Accordingly, in Figure 1 In the architecture shown, the candidate auxiliary satellite set provides the range of collaborative resources that can be called upon under the current topology, while the actual set of satellites participating in the task processing is dynamically determined by the primary satellite in combination with the current state, so that the inter-satellite collaboration relationship can be reconstructed in real time as the environment changes.

[0022] Figure 2This is a flowchart illustrating a method for offloading and allocating resources for inter-satellite collaborative tasks based on DQN-GWO, specifically including: exist Figure 2 In the process shown, after the main satellite completes the state perception of the current time slot, it first constructs the system state according to formula (7), and uniformly represents the task data volume, uplink transmission rate and downlink transmission rate of each user as the state input of the current time slot, and sends it to the DQN module. The DQN module is used to establish the mapping relationship between the system state and discrete cooperative actions. Its core function is to quickly filter out several candidate cooperative modes with high value in a large discrete action space. According to formula (8), DQN evaluates the value of each discrete action according to the current state, and generates a Top-K candidate action set according to the Q value size rule, thereby completing the preliminary screening of candidate cooperative modes. Compared with directly exhaustively enumerating all discrete actions in the entire action space, this processing method can significantly reduce the search complexity and improve the decision-making efficiency in dynamic scenarios.

[0023] After obtaining the set of candidate cooperation modes, the system enters the continuous task ratio optimization stage. For each candidate action, a continuous ratio optimization problem is first constructed based on the set of participating satellites, the available computing power of the primary and secondary satellites, and the reachability of the link under the current candidate mode. Specifically, the task acceptance ratio of the primary satellite and each secondary satellite can be uniformly represented according to the aforementioned ratio allocation model and task splitting constraints; then, the initial scoring factor is constructed according to formula (9), and the initial seed ratio vector is obtained through Softmax mapping according to formula (10). Subsequently, the GWO module is used to continuously optimize the task ratio vector under the given candidate cooperation mode. Specifically, the three ratio vectors with the best fitness in the current population are selected as α wolf, β wolf, and δ wolf, and the distance terms and guiding positions of the remaining gray wolf individuals relative to the three dominant gray wolves are calculated according to formulas (11a) to (11d), and the ratio vector is updated. After each round of update, the new ratio vector needs to be reprojected onto the probabilistic simplex space to ensure that the ratio constraints are always satisfied.

[0024] After continuous ratio optimization is completed, the system calculates the comprehensive evaluation value corresponding to the candidate action based on the current ratio vector. Specifically, a scoring function that considers both task completion delay and migration ratio is constructed according to formula (12), and the performance of each candidate action under the current optimal ratio vector is evaluated; then the overall score of the candidate action is calculated according to formula (13), and the candidate mode with the highest score is selected as the final execution action of the current time slot according to formula (14).

[0025] After the final action is determined, the system executes the cooperative offloading decision for the current time slot, obtains the corresponding immediate reward and the state of the next time slot, and forms a state transition sample. The sample is then stored in the experience replay pool, and a small batch of randomly selected samples is used for training to weaken the correlation between samples and improve network training stability. The target Q value is calculated according to formula (15), and the loss function is constructed according to formula (16) to update the DQN network parameters. Through this method, the system completes the offloading decision for the current time slot while simultaneously achieving online training and dynamic correction of the candidate cooperative mode generation network, thus enabling the selection of discrete cooperative modes in subsequent time slots to better adapt to changes in task load and link state fluctuations.

[0026] Figure 3 The overall flowchart for the offloading and resource allocation method of inter-satellite cooperation missions includes: After the system starts, it first establishes the primary satellite access relationship and a set of candidate auxiliary satellites based on the spatial relationship between the user and the satellite, and the primary satellite completes the status awareness of the current time slot. Then, the primary satellite... Figure 1 The architecture shown acquires local observation information and, based on... Figure 2 The illustrated method calls the DQN module to generate several candidate collaboration modes. For each candidate collaboration mode, the GWO module is called to optimize the task splitting ratio between the primary satellite and the auxiliary satellite. Based on this, the system outputs the final execution action for the current time slot according to the comprehensive evaluation result. Once the final execution action is determined, the primary satellite divides the user task into a local processing part and an auxiliary satellite collaborative processing part according to the optimal ratio, and distributes them to the primary satellite and the activated auxiliary satellite for execution, respectively.

[0027] During the mission execution phase, the primary satellite completes the processing of its local mission sub-blocks, while the secondary satellites complete the processing of their respective assigned mission sub-blocks. Once the secondary satellites have finished processing, they return the corresponding results to the primary satellite, which then aggregates them and provides feedback on the mission results to the user. Simultaneously, the system receives an immediate reward based on the execution results of the current time slot and completes a status update, providing input for collaborative offloading decisions in the next time slot. After the final execution action is obtained, the total mission load borne by each satellite is calculated according to formula (17), and the Jain fairness index is calculated according to formula (18), thereby evaluating the satellite load distribution and system load balance in the current time slot.

[0028] If the system operates stably after the current time slot task is completed, it will enter the next time slot and continue to execute the state awareness, candidate mode generation, ratio optimization and collaborative processing process. If there are sudden changes in link state, significant fluctuations in task load or unbalanced satellite load distribution, the DQN-GWO joint optimization process will be re-executed to dynamically adjust the inter-satellite cooperation mode and task allocation ratio, thereby ensuring that the system can still maintain good task processing efficiency, resource utilization and load balancing performance in the high dynamic environment of low-Earth orbit satellites.

[0029] Based on the above, Figures 1 to 3 Correspondingly, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for inter-satellite cooperative mission offloading and resource allocation in dynamic environments for low-Earth orbit satellites as described above.

[0030] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0031] This application also provides a computer device, including a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to implement the steps of the inter-satellite cooperative task offloading and resource allocation method for low-Earth orbit satellite dynamic environment according to any of the above embodiments.

[0032] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB ports, card reader ports, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0033] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0034] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software within the physical device.

[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for offloading and allocating resources for inter-satellite cooperative tasks in a dynamic environment for low-Earth orbit satellites, characterized in that, Includes the following steps: S1, Task Arrival and Network Awareness: Construct an onboard edge computing scenario consisting of a set of ground users and a set of service satellites, and perceive user tasks, satellite topology relationships, link status and available satellite computing resources. Determine the primary satellite and candidate auxiliary satellite sets for each user, and establish a representation of the allocation ratio of tasks between the primary satellite and auxiliary satellites. S2, Candidate cooperation mode generation based on DQN: The system state is constructed by sensing the task requirements, link status and candidate auxiliary satellite resource status in the current time slot by the access primary satellite; the inter-satellite cooperation unloading process is modeled as a Markov decision process, and the discrete cooperation modes are initially screened based on the deep Q network to generate a set of candidate actions. S3, GWO-based task allocation ratio optimization: For any candidate action, under the given cooperation mode of the candidate action, the Grey Wolf optimization algorithm is used to continuously optimize the allocation ratio of tasks between the primary star and the secondary star to obtain the optimal task allocation ratio corresponding to each candidate action, and the final execution action is selected based on the evaluation function. S4, Collaborative Execution and Result Return: Based on the final execution action and the corresponding optimal task allocation ratio, the master satellite splits the user task into a local processing part and an auxiliary satellite collaborative processing part. Each auxiliary satellite returns the task processing result to the master satellite, which then aggregates the local processing result and the result returned by the auxiliary satellite and feeds back the final task result to the user.

2. The method according to claim 1, characterized in that, In step S1, the process of determining the primary satellite and candidate auxiliary satellite set for each user is as follows: For any user i, let its ground location be... The ground projection position of the j-th satellite is If the satellite's orbital altitude is H, then the three-dimensional link distance between user i and satellite j is: ； Based on the link distance between the user and each satellite, the satellite with the shortest distance to the user is selected as the user's primary access satellite, i.e.: ； Where S represents the set of satellites in the system, This indicates that user i is connected to the primary satellite; After determining the primary satellite to be connected, the remaining satellites are sorted from nearest to farthest from the user, and the closest satellites are selected to form a candidate set of secondary satellites for that user. .

3. The method according to claim 2, characterized in that, In step S1, the allocation ratio of the task between the primary and secondary stars is specifically expressed as follows: Define the task allocation ratio vector for user i as follows: ,in This indicates the proportion of tasks that are processed locally on the primary planet. This represents the proportion of tasks assigned to the k-th auxiliary star for processing. The number of candidate auxiliary stars corresponding to user i; The task allocation ratio must satisfy the non-negativity constraint and the normalization constraint, that is: 。 4. The method according to claim 1, characterized in that, In step S2, the constructed system state includes the amount of user task data, the uplink transmission rate between the user and the main satellite, and the downlink transmission rate. The system state expression is as follows: ； in, This represents the amount of task data for user i. and Let N represent the uplink and downlink transmission rates between user i and the access master satellite, respectively, and let N represent the number of users in the system.

5. The method according to claim 4, characterized in that, In step S2, the process of selecting discrete cooperation modes based on deep Q-networks is as follows: Combine the discrete collaborative actions of all users within the current time frame into a joint action. Based on the current state The main star uses a deep Q-network to evaluate the value of each candidate collaborative action. ,in Indicates network parameters; Construct the target Q value based on the immediate reward of the current time frame and the maximum action value of the next time frame: ； in, This represents the immediate reward obtained after performing the corresponding discrete cooperative action within the current time frame. Indicates the discount factor. Indicates the system state for the next time frame; A loss function is constructed using the difference between the target Q-value and the current Q-value to complete the iterative update of the network parameters. The loss function is: ； By evaluating the value and updating the parameters of a deep Q-network, various discrete cooperation modes under the current system state are screened, and several candidate cooperation modes with higher value are retained to form a set of candidate actions.

6. The method according to claim 1, characterized in that, In step S3, before continuous optimization using the Grey Wolf optimization algorithm, an initial task allocation ratio is first constructed. The specific process is as follows: For user i, if its total task data volume is The amount of task allocated to the k-th service node is: Initial scoring factors are constructed based on the current discrete actions, the set of participating satellites, and the computing capabilities of each satellite. Define the computing power of the k-th node participating in user i's task processing as: The reachability item corresponding to the user's ground projection distance from this node is: The initial rating factors are: ； in, and These represent the weights of the computational capability term and the distance term, respectively. This represents the primary star bias term introduced by discrete actions; Based on the above scoring factors, the initial seed ratio vector is obtained through Softmax mapping and used as the initial population input for the Grey Wolf optimization algorithm.

7. The method according to claim 1, characterized in that, In step S3, the iterative update process of the Grey Wolf optimization algorithm is as follows: Each gray wolf's position vector is represented as a candidate task allocation ratio solution. In each iteration, based on the position vectors of the three gray wolves with the best fitness in the current population, namely α wolf, β wolf, and δ wolf, the position vectors of the other gray wolves are updated. The update process is as follows: ； ； ； ； in, This represents the position vector of the nth gray wolf during the r-th iteration. , and Let represent the position vectors of the three gray wolves with the best fitness in the current population. , and The shrinkage coefficient, , and For random search parameters; After each iteration, the updated proportion vector is normalized to satisfy the non-negativity constraint and the normalization constraint. After the iteration terminates, the optimal task allocation proportion corresponding to each candidate cooperation mode is obtained.

8. The method according to claim 1, characterized in that, In step S3, the process of selecting the final action based on the evaluation function is as follows: For user i, construct an evaluation function that simultaneously considers task completion latency and task migration ratio: ； in, This represents the task completion delay for user i under the current task allocation ratio. Indicates the percentage of tasks migrated. Indicates the migration penalty coefficient; For the k-th candidate collaboration mode, the evaluation functions of all users are summed to obtain the comprehensive evaluation value of the candidate mode: ； Based on the comprehensive evaluation value of all candidate cooperation modes, the candidate mode with the highest evaluation value is selected as the final action to be executed in the current time frame: ； in, This represents the set of candidate cooperative patterns generated by the deep Q-network within the current time frame. This indicates the optimal collaboration mode that was ultimately selected.

9. The method according to claim 1, characterized in that, It also includes the system load balancing evaluation steps: Define the total mission load undertaken by the j-th satellite as: ； in, This represents the proportion of task processing allocated by user i to satellite j; The Jain Fairness Index is used to evaluate the load balancing level of the system, and its expression is: ； Where M represents the total number of satellites in the system, and J represents the load balancing degree of the system. The closer the J value is to 1, the better the load balancing of the system.

Images