Multi-star coordination computing and task scheduling method and system suitable for constellation scale evolution

Abstract

This invention discloses a multi-satellite collaborative computing and task scheduling method and system applicable to constellation-scale evolution. The method includes: acquiring heterogeneous state data of each satellite based on the orbital plane state view of the current time slot, performing unified feature encoding to obtain the feature representation of each satellite, and aggregating them into a global state feature at the orbital plane level; segmenting and encoding each task to be decomposed according to the global state feature to obtain an action vector including the number of sub-task decompositions, the task decomposition scheme, and the satellite allocation scheme for each task to be decomposed; calculating the task completion rate of each priority group based on the action vector and task priority to determine the system-level reward value; updating the state of each satellite according to the computing resources used by each task to be decomposed in the current time slot to generate the orbital plane state view of the next time slot, and then returning to the step of acquiring heterogeneous state data, until all time slots are iterated, and obtaining a task scheduling scheme based on the action vector corresponding to each task to be decomposed in each time slot.

Description

Technical Field

[0001] This invention belongs to the field of satellite mission scheduling technology, specifically relating to a multi-satellite collaborative computing and mission scheduling method and system applicable to constellation-scale evolution. Background Technology

[0002] Currently, with the rapid deployment of low-Earth orbit mega-constellations and the deepening application demands, space-based information systems are facing severe technical bottlenecks. Traditional satellite networks mainly adopt a centralized architecture of "satellite acquisition and ground processing." Raw data needs to be relayed multiple times via inter-satellite links before being sent down to the ground control station for processing. This approach is not only limited by scarce ground station resources and high data return costs, but also fails to meet the stringent near-real-time processing requirements of applications such as disaster emergency response and high-timeliness target identification due to significant return latency. At the same time, the increasing complexity of space missions places higher demands on on-board computing power. However, individual satellites are limited by their inherent size, weight, and power consumption constraints, resulting in extremely limited computing, storage, and energy resources, making it difficult to independently support data-intensive processing tasks. A deeper problem lies in the fact that existing technological systems lack a collaborative perspective that treats the entire constellation as a unified computing resource. Their task scheduling mechanisms are mostly derived from static ground networks or single-satellite optimization, failing to effectively perceive and schedule distributed multi-satellite resources. This leads to a severe imbalance in the computing load within the constellation, creating "resource silos." Furthermore, the statically preset scheduling strategies are ill-suited to adapt to dynamic changes in network topology and the random injection of tasks. Overall system resource utilization, task throughput, and resilience and robustness during scaling are all insufficient. These systemic deficiencies in data transmission, single-satellite computing power, and multi-satellite collaboration constitute a fundamental obstacle to the intelligent and large-scale evolution of space-based information services.

[0003] To address the aforementioned obstacles, the following solutions have been proposed in existing technologies: Patent application number 202411848558.5—"Task Scheduling Method, System, and Equipment for Low-Earth Orbit Constellations to Dynamic Space Targets in a Fully Distributed Satellite System"—proposes a task scheduling method for low-Earth orbit constellations to dynamic space targets in a fully distributed satellite system. This method solves the problems of high cost, low flexibility, and low stability of traditional centralized scheduling through a decentralized architecture. This method uses a Co-evolutionary Genetic Algorithm (CCGA) to decompose the multi-satellite scheduling problem into a single-satellite scheduling problem. Each satellite independently performs local optimization, and global negotiation is conducted through inter-satellite communication in a ring topology. Task benefits and observation effects are used as evaluation indicators to ultimately reach a consensus on the task scheduling scheme. Patent application number 202510613884.6—"A Simulation-Based Dynamic Planning Method for Satellite Scheduling Tasks"—constructs a dynamic environment model, a satellite resource dynamic model, and a multi-satellite collaborative model. Driven by multiple models, a task planning algorithm is used to dynamically schedule single-satellite and collaborative tasks, ultimately generating a satellite task execution plan table. The paper "Task Offloading and Resource Allocation for Satellite-Terrestrial Integrated Networks" addresses the computational offloading and resource allocation problem in satellite-terrestrial integrated networks, aiming to minimize the total computational cost of tasks by jointly optimizing the selection of offloading nodes, the offloading ratio, and the allocation of computational resources. The paper proposes a multi-layered, multi-access edge computing architecture integrating cloud and satellite, and designs a multi-agent reinforcement learning algorithm that combines a dual deep Q-network and a deep deterministic policy gradient to handle discrete and continuous action spaces.

[0004] However, in patents for mission scheduling methods, systems, and devices for low-Earth orbit constellations targeting dynamic space targets in a fully distributed satellite architecture, the optimization process is based on iterative genetic algorithms, which are computationally complex and have slow convergence speeds, making it difficult to meet the rapid scheduling requirements of highly dynamic tasks. The global negotiation mechanism does not adequately address the fault tolerance of satellite failures or communication interruptions, resulting in insufficient system robustness. While a patent for a simulation-based dynamic planning method for satellite scheduling tasks emphasizes the "dynamic" environment and resources, its underlying mission planning algorithm still relies on traditional optimization methods (such as genetic algorithms and particle swarm optimization), lacking agent-scale adaptive capabilities and unable to dynamically handle scale changes from tens to thousands of satellites through a unified strategy. Its multi-satellite collaborative model requires manual design of interaction rules for different scenarios, failing to achieve strategy transfer and incremental optimization, and exhibiting repetitive modeling issues. Furthermore, this patent does not address the end-to-end joint optimization of task decomposition and resource allocation, still treating task allocation and resource scheduling as separate stages. The method proposed in the literature "Task Offloading and Resource Allocation for Satellite-Terrestrial Integrated Networks" suffers from over-simplification of the model and decoupling of optimization. The task processing model in this study is too simplistic, only supporting proportional offloading of tasks between local and single satellite or cloud nodes, and failing to accurately identify and model the task dependencies within complex computing tasks. Therefore, it cannot achieve intelligent decomposition and parallel collaborative processing of tasks across multiple satellite nodes.

[0005] In summary, traditional task scheduling methods face significant challenges, primarily including: firstly, neural networks based on fixed input dimensions struggle to adapt to dynamic changes in the number of satellites; secondly, existing methods often separate task decomposition from resource allocation, leading to inefficient decision-making. Furthermore, traditional reinforcement learning methods face difficulties in training and slow convergence in large-scale scenarios, making them unsuitable for practical applications.

[0006] Therefore, how to provide a multi-star collaborative computing and task scheduling method that is applicable to complex multi-star environments, has high computational resource utilization, is robust, and is suitable for constellation-scale evolution has become an important issue. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a multi-star collaborative computing and task scheduling method and system applicable to constellation-scale evolution.

[0008] The technical problem to be solved by this invention is achieved through the following technical solution:

[0009] In a first aspect, the present invention provides a multi-star collaborative computing and task scheduling method applicable to constellation-scale evolution, the multi-star collaborative computing and task scheduling method comprising:

[0010] Collect network link information and mission request information of each LEO satellite to construct an orbital plane status view in the current time slot;

[0011] After obtaining heterogeneous state data of each satellite based on the orbital plane state view, the heterogeneous state data of each satellite is uniformly feature-encoded to obtain the feature representation of each satellite; the feature representations of each satellite are aggregated into global state features at the orbital plane level; the satellite is a LEO satellite or a MEO satellite.

[0012] Based on the global state features, each task to be decomposed is segmented and encoded to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme for each task to be decomposed.

[0013] The system-level reward value is determined by calculating the task completion rate of each priority group based on the action vector corresponding to each task to be decomposed and the preset task priority.

[0014] Update the satellite status according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane status view of the next time slot based on the updated satellite status. Then return to execute the step of obtaining the heterogeneous status data of each satellite based on the orbital plane status view. After iterating through all time slots, obtain the action vector and system-level reward value corresponding to each task to be decomposed in each time slot.

[0015] The task scheduling scheme is obtained based on the action vectors corresponding to each task to be decomposed in each time slot.

[0016] Secondly, the present invention provides a multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution, the multi-star collaborative computing and task scheduling system comprising:

[0017] The dynamic state perception module is used to collect network link information and mission request information of each LEO satellite in order to construct an orbital plane state view in the current time slot;

[0018] A unified feature encoding module is used to obtain heterogeneous state data of each satellite based on the orbital plane state view, perform unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite; and aggregate the feature representations of each satellite into a global state feature at the orbital plane level; wherein the satellite is a LEO satellite or a MEO satellite.

[0019] The joint strategy decision module is used to segment and encode each task to be decomposed based on the global state characteristics to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme corresponding to each task to be decomposed.

[0020] The environment simulation and execution module is used to calculate the task completion rate of each priority group based on the action vectors corresponding to each task to be decomposed and the preset task priorities to determine the system-level reward value.

[0021] The environment simulation and execution module is also used to update the state of each satellite according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane state view of the next time slot based on the updated satellite state, and then return to execute the step of obtaining the heterogeneous state data of each satellite based on the orbital plane state view, until all time slots are iterated, and the action vector and system-level reward value corresponding to each task to be decomposed in each time slot are obtained.

[0022] The environment simulation and execution module is also used to obtain a task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot.

[0023] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0024] Memory, used to store computer programs;

[0025] When a processor executes a computer program stored in memory, it implements the steps described in any of the above-mentioned multi-star collaborative computing and task scheduling methods and systems applicable to constellation-scale evolution.

[0026] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method steps described in any of the above-mentioned applicable constellation-scale evolution multi-star collaborative computing and task scheduling methods and systems.

[0027] This invention provides a multi-satellite collaborative computing and task scheduling method applicable to constellation-scale evolution. After obtaining heterogeneous state data of each satellite based on the orbital plane state view, the method performs unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite. Then, the feature representations of each satellite are aggregated into global state features at the orbital plane level. This process can achieve adaptive support for the dynamic expansion of the satellite network scale, which is not only suitable for complex multi-satellite environments and enhances robustness, but also overcomes the technical bottleneck of fixed input dimensions in traditional neural networks.

[0028] By segmenting and encoding each task to be decomposed based on global state characteristics, we can simultaneously obtain the number of subtask decompositions, task decomposition schemes, and action vectors for executing satellite allocation schemes for each task to be decomposed. This avoids the information loss and decision bias that exist in traditional multi-stage optimization and effectively improves the utilization rate of computing resources.

[0029] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a multi-star collaborative computing and task scheduling method applicable to constellation-scale evolution provided by an embodiment of the present invention;

[0031] Figure 2 This is a flowchart illustrating another method for multi-star collaborative computing and task scheduling applicable to constellation-scale evolution provided by an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the joint optimization decision-making process for task decomposition and allocation provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the reward function generation process provided in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the architecture of a multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution provided by an embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram of an integrated space-ground satellite network scenario applied in an embodiment of the present invention;

[0036] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0038] To address the shortcomings of existing satellite mission scheduling methods, such as difficulty in adapting to complex multi-satellite environments, low computational resource utilization, and insufficient robustness, this invention provides a multi-satellite collaborative computing and mission scheduling method applicable to constellation-scale evolution. (See also...) Figure 1 , Figure 1 This is a flowchart illustrating a multi-star collaborative computing and task scheduling method applicable to constellation-scale evolution provided by an embodiment of the present invention, specifically including the following steps:

[0039] Step S101: Collect network link information and mission request information of each LEO (Low Earth Orbit) satellite, and construct an orbital plane status view under the current time slot.

[0040] First, establish an integrated space-ground satellite network scenario, see [link / reference] Figure 2 , Figure 2 This is a flowchart illustrating another multi-satellite collaborative computing and task scheduling method applicable to constellation-scale evolution provided by an embodiment of the present invention. The satellite's operating time is divided into T time slots, each time slot being 1 second long. For each time slot, network link information of the corresponding integrated space-ground satellite network is generated.

[0041] The collection of network link information and mission request information of each LEO satellite in the current time slot includes real-time acquisition of its inter-satellite link status and quality parameters through the network link information acquisition unit, and acquisition of multi-dimensional characteristics such as data volume, priority and timeliness requirements of LEO satellites that have received specific missions through mission request information.

[0042] Then, the control satellite systematically collects and integrates the network link information and mission request information of all the aforementioned LEO satellites through the communication links established within the orbital plane, constructing a complete orbital plane status view for the current time slot. The satellites include both LEO and MEO (Medium Earth Orbit) satellites. The network link information of the LEO satellites contains relevant information about the MEO satellites.

[0043] Step S102: After obtaining the heterogeneous state data of each satellite based on the orbital plane state view, perform unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite; aggregate the feature representation of each satellite into a global state feature at the orbital plane level; the satellite is a LEO satellite or a MEO satellite.

[0044] In this embodiment of the invention, heterogeneous status data from satellites across all orbital planes is acquired based on an orbital plane status view. This heterogeneous status data encompasses multi-dimensional information such as resource status, link quality, and mission characteristics. After uniformly encoding the aforementioned information, a variable number of satellites are mapped to a unified feature space using dynamic embedding technology. This allows each satellite to obtain its unique feature representation, supporting subsequent collaborative decision-making. See also... Figure 3 , Figure 3 This is a schematic diagram of the joint optimization decision-making process for task decomposition and allocation provided in an embodiment of the present invention.

[0045] In this embodiment of the invention, feature aggregation can be used to aggregate the feature representations of each satellite into global state features at the orbital plane level, so as to capture the resource distribution and task load features at the system level.

[0046] Step S103: Segment and encode each task to be decomposed according to the global state features to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme corresponding to each task to be decomposed.

[0047] In this embodiment of the invention, each task to be decomposed is segmented and encoded based on global state features to obtain the action vector corresponding to each task to be decomposed, including:

[0048] The number of decomposition steps for each task is determined based on the global state characteristics and each task to be decomposed.

[0049] By utilizing global state features and the number of decompositions for each task to be decomposed, each task to be decomposed is divided into multiple subtasks, thus obtaining the task decomposition scheme corresponding to each task to be decomposed.

[0050] Based on the task decomposition scheme of each task to be decomposed, the execution satellite allocation scheme corresponding to each task to be decomposed is determined.

[0051] In this embodiment of the invention, the optimal number of decompositions for each task can be determined based on global state characteristics combined with the data volume, computational complexity, and real-time requirements of the task to be decomposed. For a satellite receiving a task, the number of satellites to which the task can be distributed is limited, and there is a certain time loss in distributing to these satellites. The computing resources on these satellites are also uncertain. Optimal decomposition involves integrating the states of all satellites that can send the task to determine the best decomposition method and obtain the optimal number of decompositions. Specifically, if there is a task to be decomposed with a data volume of 2GB, a complexity of 2500 operations per bit, and up to 10 surrounding satellites, the policy network will analyze the number of hops and distance required for the existing satellites to reach these 10 satellites. At the same time, it will analyze the specific attributes of each satellite, such as: remaining computing resources, whether other satellites also need to transmit the task, etc., and finally give the optimal number of decompositions that can guarantee the completion of the task.

[0052] Then, for each task to be decomposed, based on its determined number of decompositions and global state characteristics, the key nodes for task segmentation are located, and the task to be decomposed is divided into multiple sub-tasks to form a specific task decomposition scheme.

[0053] Based on the task decomposition scheme of each task to be decomposed, the data scale, computing resource requirements and data dependencies of each sub-task are further calculated to obtain the attribute information of the sub-task, so as to provide a basis for executing the satellite allocation scheme.

[0054] Based on the global state characteristics and the attribute information of each subtask, an execution satellite allocation scheme is generated, and a suitable execution satellite is assigned to each subtask to achieve optimal overall system performance.

[0055] The above decision results, namely the number of sub-task decompositions, task decomposition schemes, and satellite allocation schemes corresponding to the tasks to be decomposed, are encoded into standardized action vectors.

[0056] Step S104: Calculate the task completion rate of each priority group based on the action vector corresponding to each task to be decomposed and the preset task priority to determine the system-level reward value.

[0057] In this embodiment of the invention, the task completion rate of each priority group is calculated based on the action vector corresponding to each task to be decomposed and the preset task priority to determine the system-level reward value, including:

[0058] Calculate the available computation time for each subtask based on the action vector corresponding to each task to be decomposed and multiple subtasks;

[0059] Based on preset task priorities, the subtasks in each task to be decomposed are divided into different priority groups;

[0060] The task completion rate of each priority group is calculated based on the subtasks in each priority group and the available computation time for each subtask.

[0061] The system-level reward value is determined by weighted summation of the task completion rates of each priority group.

[0062] See Figure 4 , Figure 4 This is a schematic diagram of the reward function generation process provided in an embodiment of the present invention. Based on the action vectors corresponding to each task to be decomposed and the information of multiple subtasks, the transmission delay generated by each subtask during transmission is calculated. Propagation delay and queuing delay The sum of these three factors is the total communication delay of the subtask. , expressed as:

[0063] ;

[0064] The maximum allowed end-to-end latency for each subtask is After deducting the latency required for subtask communication Then, the available computation time for that subtask is... :

[0065] ;

[0066] In this embodiment of the invention, all subtasks are divided into different priority groups according to the priority set in the original task. Subtasks within the same group have the same priority to ensure that high-priority tasks are not affected by low-priority tasks in resource allocation. Computational and communication resources are allocated to each group of subtasks in descending order of task group priority to ensure priority scheduling of critical tasks.

[0067] For multiple subtasks derived from the same original task, the principle of "success if all are successful, failure if any one fails" is adopted. In this embodiment of the invention, the available computation time for each subtask is used as the time standard for whether the subtask can be successfully executed, and then the task completion rate of each priority group can be further calculated statistically.

[0068] The task completion rates of each priority group are weighted and summed according to their importance to construct a system-level reward function, which comprehensively reflects the overall task execution effect:

[0069] ;

[0070] in, , Indicates the total number of priority groups; Indicates the first Task completion rate for each priority group; Indicates the first The weights corresponding to each priority group; This represents the system-level reward value.

[0071] Step S105: Update the status of each satellite according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane status view of the next time slot based on the updated satellite status. Then return to execute the step of obtaining the heterogeneous status data of each satellite based on the orbital plane status view. After iterating through all time slots, obtain the action vector and system-level reward value corresponding to each task to be decomposed in each time slot.

[0072] In this embodiment of the invention, based on the sub-tasks that have been allocated and executed in the current time slot, i.e., the action vectors corresponding to each task to be decomposed in the current time slot, the computing resources of the corresponding satellites are deducted, and the status of each satellite is updated, specifically the remaining available computing resources and current load status of each satellite.

[0073] Then, based on the execution results of the subtasks, the executed tasks to be decomposed are marked as "completed" and the resources they occupy are released; for the tasks to be decomposed that fail to execute, their status is updated to "discarded".

[0074] Based on the updated satellite resources and mission status information, the orbital plane status view for the next time slot is re-aggregated and generated for the next decision cycle, i.e., the status input for the next time slot. The clock is advanced one time step to prepare for receiving new mission requests or triggering the next round of mission scheduling and resource allocation. Then, the process returns to the step of obtaining heterogeneous status data of each satellite based on the orbital plane status view, until all time slots are iterated, obtaining the action vector and system-level reward value corresponding to each task to be decomposed in each time slot.

[0075] Step S106: Obtain the task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot.

[0076] In this embodiment of the invention, each satellite encapsulates the state information of its orbital plane in each decision time slot, the generated action vector, the obtained system-level reward value, and the state data of the next time slot into standardized experience units for storage, and transmits them to the ground station through the satellite-to-ground link. After receiving and aggregating the data from all orbital planes, the ground station performs spatiotemporal alignment and splicing processing, and finally archives the complete system experience data into a distributed database to obtain the mission scheduling scheme, providing data support for subsequent analysis and model iteration.

[0077] In one implementation, before obtaining the task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot, the multi-satellite collaborative computing and task scheduling method further includes:

[0078] When the number of time slots in the iteration meets the preset training conditions, the global value estimation technique is used to update the action vectors and system-level reward values ​​corresponding to each task to be decomposed in each time slot, so as to obtain the updated action vectors corresponding to each task to be decomposed in each time slot.

[0079] Then, a task scheduling scheme can be obtained based on the action vectors corresponding to each task to be decomposed in each updated time slot.

[0080] In this embodiment of the invention, when the ground station determines that the number of currently received time slots meets the training conditions, the training process is initiated; otherwise, the process returns to the step of collecting network link information and mission request information of each LEO satellite to construct a heterogeneous state data acquisition system based on the orbital plane state view. The training conditions can be that the number of currently received time slots exceeds a preset number of times. The preset number of times can be set by technicians based on experience and is not limited here.

[0081] The specific training process is as follows:

[0082] The ground station sequentially reads the action vectors and system-level reward values ​​corresponding to each time slot, inputs them into the global value network to estimate the global value; based on the obtained global value estimate, it calculates the generalized advantage estimate (GAE) and further derives the λ-reward; using the action vectors corresponding to each time slot, the system-level reward value, and the λ-reward, it performs multiple rounds of parameter updates on the global value network. Based on the updated network, it re-estimates the global value, recalculates the GAE, and sends the calculated GAE back to the control satellite; the control satellite updates its onboard policy network multiple times based on the GAE and locally stored standardized empirical units. This policy network includes a unified feature encoding module and a joint policy decision module (used to execute the unified feature encoding step and the task decomposition step, respectively). Both the control satellite and the ground station clear their stored data. Step S102 to the step of clearing stored data is executed repeatedly until the model converges in the current integrated space-ground satellite network scenario. The converged model is saved, and the action vectors corresponding to each task to be decomposed in each time slot are obtained after the update. Based on the action vectors corresponding to each task to be decomposed in each time slot, a task scheduling scheme is obtained.

[0083] Furthermore, in order to construct a task scheduling scheme that can continuously evolve with the scale of the satellite network, the multi-satellite collaborative computing and task scheduling method in this embodiment of the invention may also include a step of dynamic evolution of scenario complexity:

[0084] Increase the complexity of satellite system scenarios, such as increasing the number of orbital planes in the low Earth orbit layer, or deploying more satellites on existing orbital planes;

[0085] The course learning control submodule deployed in the ground station adaptively adjusts the following training parameters according to the evolved new scenario: training conditions, number of single training updates for the global value network and policy network, learning rate of both, entropy regularization coefficient, discount factor, and generalized advantage estimation parameters.

[0086] After the parameters are reconfigured, the process returns to step S101 to continue a new round of system training and optimization until the optimized task scheduling scheme is finally obtained.

[0087] To address the issues of low training efficiency and convergence difficulties faced by existing reinforcement learning methods in large-scale scenarios, this invention introduces a course-based training system. Through a progressive complexity evolution strategy, a stable and efficient learning process is achieved. Starting from simple scenarios, intelligent parameter transfer is used to gradually adapt to complex environments, effectively avoiding the low exploration efficiency and policy oscillation problems caused by direct training in large-scale scenarios.

[0088] In this embodiment of the invention, after obtaining the heterogeneous state data of each satellite based on the orbital plane state view, the feature representation of each satellite is obtained by uniformly encoding the heterogeneous state data of each satellite. Then, the feature representation of each satellite is aggregated into a global state feature at the orbital plane level. This process can achieve adaptive support for the dynamic expansion of the satellite network scale, which is not only suitable for complex multi-satellite environments and enhances robustness, but also overcomes the technical bottleneck of fixed input dimension of traditional neural networks.

[0089] By segmenting and encoding each task to be decomposed based on global state characteristics, we can simultaneously obtain the number of subtask decompositions, task decomposition schemes, and action vectors for executing satellite allocation schemes for each task to be decomposed. This avoids the information loss and decision bias that exist in traditional multi-stage optimization and effectively improves the utilization rate of computing resources.

[0090] Based on the same inventive concept, embodiments of the present invention also provide a multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution, see [link to relevant documentation]. Figure 5 , Figure 5 This is a schematic diagram of the architecture of a multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution provided by an embodiment of the present invention. The modules in the system are tightly coupled in function and progressively advance in logic, constructing a closed-loop intelligent decision-making and scheduling system, as detailed below:

[0091] The dynamic state perception module is used to collect network link information and mission request information of each LEO satellite in order to construct an orbital plane state view in the current time slot;

[0092] A unified feature encoding module is used to obtain heterogeneous state data of each satellite based on the orbital plane state view, perform unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite; and aggregate the feature representations of each satellite into a global state feature at the orbital plane level; wherein the satellite is a LEO satellite or a MEO satellite.

[0093] The joint strategy decision module is used to segment and encode each task to be decomposed based on the global state characteristics to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme corresponding to each task to be decomposed.

[0094] The environment simulation and execution module is used to calculate the task completion rate of each priority group based on the action vectors corresponding to each task to be decomposed and the preset task priorities to determine the system-level reward value.

[0095] The environment simulation and execution module is also used to update the state of each satellite according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane state view of the next time slot based on the updated satellite state. Then, it returns to execute the step of obtaining the heterogeneous state data of each satellite based on the orbital plane state view, until iterates through all time slots, and obtains the action vector and system-level reward value corresponding to each task to be decomposed in each time slot.

[0096] The environment simulation and execution module is also used to obtain a task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot.

[0097] In this embodiment of the invention, the dynamic state perception module includes a task request information acquisition unit and a network link information acquisition unit; the task request information acquisition unit is used to acquire the task request information of each LEO satellite, and the network link information acquisition unit is used to acquire the network link information of each LEO satellite.

[0098] In this embodiment of the invention, the unified feature encoding module includes a dynamic embedding layer, a multi-head self-attention layer, and a feature aggregation layer. The dynamic embedding layer is used to obtain heterogeneous state data of each satellite based on the orbital plane state view, and then perform unified feature encoding on the heterogeneous state data of each satellite, using dynamic embedding technology to map a variable number of satellites to a unified feature space; the multi-head self-attention layer is used to obtain the feature representation of each satellite; the feature aggregation layer is used to aggregate the feature representation of each satellite into a global state feature at the orbital plane level.

[0099] In this embodiment of the invention, the joint strategy decision module includes a task decomposition decision submodule and a task allocation decision submodule; the task decomposition decision submodule includes a task decomposition number decision unit, a task decomposition location decision unit and a subtask size calculation unit, and the task allocation decision submodule includes a subtask allocation decision unit.

[0100] The system includes: a task decomposition number decision unit, used to determine the number of decompositions for each task based on global state characteristics and the tasks to be decomposed; a task decomposition location decision unit, used to locate key nodes for task segmentation and form specific task decomposition schemes for each task to be decomposed; a subtask size calculation unit, used to further calculate the data size, computing resource requirements, and data dependencies of each subtask based on the generated decomposition scheme, providing a basis for task allocation; and a subtask allocation decision unit, used to determine the execution satellite allocation scheme for each task to be decomposed based on the task decomposition scheme for each task to be decomposed.

[0101] In one implementation, the environment simulation and execution module includes an environment state management submodule, a resource state update submodule, a task execution submodule, and a reward function generation submodule. The environment state management submodule is used to re-aggregate and generate a global environment state representation at the orbital plane level based on the updated node resource and task state information. The resource state update submodule is used to deduct the computing resources of the corresponding satellite nodes according to the allocated and executed sub-task information, and update the remaining available computing units and current load status of each node. The task execution submodule is used to mark completed tasks as "completed" and release their occupied resources based on the sub-task execution results; for failed tasks, its status is updated to "discarded". The reward function generation submodule is used to calculate the system-level reward value.

[0102] In one implementation, the multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution also includes a task decomposition and allocation joint action generation module. This module serves as a key link between strategy output and execution, ensuring the executability of scheduling decisions and the stable improvement of overall system performance.

[0103] In one implementation, the multi-star collaborative computing and task scheduling system applicable to constellation-scale evolution further includes a model training and evolution module. This module further comprises a curriculum learning control submodule and a global value network. Before obtaining a task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot, the model training and evolution module, when the number of iterated time slots meets a preset training condition, uses global value estimation technology to update the action vectors corresponding to each task to be decomposed in each time slot and the system-level reward value, obtaining the updated action vectors corresponding to each task to be decomposed in each time slot. The model training and evolution module is used for the step of dynamic evolution of scene complexity.

[0104] In this embodiment of the invention, the dynamic state perception module is responsible for the real-time acquisition of multi-source heterogeneous data, including a task request information acquisition unit and a network link information acquisition unit. It is responsible for receiving and preprocessing satellite network status and task requests that change over time. Its design supports dynamic changes in the number of satellites, providing flexibility for subsequent processing.

[0105] The unified feature encoding module uses the Transformer architecture as its backbone network. Through dynamic embedding layers, multi-head self-attention layers, and feature aggregation layers, it encodes a variable number of input entities into a unified, fixed-dimensional feature space by means of its internal dynamic embedding, self-attention computation, and feature aggregation mechanisms. This solves the problem of input dimension explosion caused by network scaling.

[0106] The joint strategy decision-making module, as the core decision-making unit of the system, includes a task decomposition decision-making submodule and a task allocation decision-making submodule. The former further encompasses functions such as task decomposition frequency decision, decomposition location identification, and subtask size calculation, while the latter realizes the rational allocation of subtasks based on resource status and task decomposition status. By synchronously outputting task decomposition schemes and resource allocation schemes during a forward propagation process, deep coupling and joint optimization of the two key decision-making steps are achieved.

[0107] The task decomposition and allocation joint action generation module serves as a key link between strategy output and execution, ensuring the executability of scheduling decisions and the stable improvement of overall system performance.

[0108] The environment simulation and execution module is responsible for building a virtual satellite network environment, including sub-modules such as environment status management, resource status update, task execution and reward function generation. It can simulate the execution effect of decisions in a real satellite network and output various evaluation data, including reward signals.

[0109] The model training and evolution module integrates a curriculum learning control mechanism and a global value network, supporting the system in continuously optimizing its policies in dynamic and complex environments. Utilizing environmental feedback, the parameters of the policy and value networks are updated through the MAPPO (Multi-Agent Proximal Policy Optimization) algorithm. Simultaneously, its built-in curriculum learning controller systematically increases the complexity of the simulation environment based on the learning progress, guiding the adjustment of neural network hyperparameters to ensure the system can stably and efficiently learn and evolve from simple scenarios to complex ones, ultimately achieving excellent scalability.

[0110] In this embodiment of the invention, a multi-stage course learning and training system oriented towards scale evolution is adopted. The system dynamically adjusts environmental complexity through a course learning controller and achieves intelligent parameter migration across course levels in conjunction with a model parameter manager. This system designs course sequences from simple to complex, employs a selective parameter reset strategy to preserve general decision-making patterns, and utilizes centralized training guidance from a global value network to achieve a smooth evolution of the system from basic, simple scenarios to large-scale, complex scenarios.

[0111] See Figure 6 , Figure 6This is a schematic diagram of a space-ground integrated satellite network scenario applied in an embodiment of the present invention. This scenario provides a specific physical deployment environment for each functional module of the system. In this hierarchical heterogeneous network, the LEO layer is interconnected in a planar manner through a "one-satellite-four-chain" laser link, while the LEO and MEO layers are connected in three dimensions through directional laser links, constructing a complete space information transmission infrastructure. Each LEO satellite in the system is equipped with a dynamic state perception module, responsible for collecting its own state data and task requests in real time. The control satellite in each LEO orbital plane acts as the "intelligent agent and computing power scheduling center" of that orbital plane, integrating a unified feature encoding module, a joint strategy decision-making module, and a task decomposition and allocation joint action generation module. It is responsible for aggregating and processing the state information of all satellites in the orbital plane, performing deep feature extraction and fusion, and finally generating joint optimization decisions for task decomposition and allocation. In terms of computing power architecture, LEO and MEO satellites together constitute a "cloud-edge" collaborative system. The MEO satellite, as a cloud computing node, together with the LEO edge nodes, supports the operation of the environment simulation and execution module, forming a high-fidelity simulation verification environment through inter-layer link interconnection. The ground station deploys a model training and evolution module, which receives observation and decision-making data from various control satellites through the space-to-ground link, implements centralized training and course learning scheduling, and generates global signals to guide distributed strategy optimization, thereby achieving deep integration of space-based distributed perception and decision-making with ground-based centralized training and optimization.

[0112] This invention provides a multi-satellite collaborative computing and task scheduling system applicable to constellation-scale evolution. It constructs a unified and scalable decision-making framework, achieving continuous adaptation to the dynamic expansion of the satellite network by integrating the dynamic encoding capabilities of the Transformer architecture with a multi-agent near-end policy optimization algorithm. Specifically, when the number of satellites expands from dozens to thousands, the system intelligently encodes variable-length satellite state inputs using the Transformer's self-attention mechanism, mapping heterogeneous satellite resource states to a unified feature representation space. Global environmental awareness is then obtained through feature aggregation, overcoming the limitations of traditional neural networks on fixed input dimensions. Furthermore, the system employs an end-to-end joint optimization strategy, synchronously completing task decomposition and resource allocation decisions during the forward propagation of a single-policy network. The shared feature representation space ensures deep coupling between task decomposition granularity and satellite computing power and communication resources, effectively solving the suboptimal decision-making problem caused by the traditional "decompose first, then allocate" serial mode. To further improve the system's learning efficiency during scaling, this embodiment of the invention introduces a course learning mechanism. By designing training course sequences ranging from simple to complex, the policy network can progressively adapt to satellite network environments of different scales. During course switching, an intelligent parameter migration strategy is employed to retain learned general decision-making patterns, significantly accelerating the convergence process. The entire system adopts a centralized training and distributed execution architecture, guiding distributed policy optimization through a global value network. This ensures autonomous decision-making for each satellite while achieving optimal system-level performance, ultimately constructing an intelligent task scheduling system that can continuously evolve with the scale of the satellite network, possessing both high efficiency and robustness.

[0113] This invention also provides an electronic device, such as... Figure 7 As shown, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704.

[0114] Memory 703 is used to store computer programs;

[0115] When the processor 701 executes the program stored in the memory 703, it implements the method steps of any of the above-mentioned applicable constellation-scale evolution multi-star collaborative computing and task scheduling methods.

[0116] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0117] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0118] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0119] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0120] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program implements the method steps of any of the above-described methods for multi-star collaborative computing and task scheduling applicable to constellation-scale evolution.

[0121] Optionally, the computer-readable storage medium may be non-volatile memory (NVM), such as at least one disk storage device.

[0122] Optionally, the aforementioned computer-readable storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0123] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the steps of the method described in any of the above-described applicable constellation-scale evolution multi-star collaborative computing and task scheduling methods.

[0124] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0125] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0126] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0127] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0128] For system / electronic device / storage medium embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0129] It should be noted that the system, electronic device and storage medium in the embodiments of the present invention are respectively the system, electronic device and storage medium applying the above-mentioned method for multi-star collaborative computing and task scheduling applicable to constellation-scale evolution. Therefore, all embodiments of the above-mentioned method for multi-star collaborative computing and task scheduling applicable to constellation-scale evolution are applicable to the system, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0130] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-star collaborative computing and task scheduling method applicable to constellation-scale evolution, characterized in that, The multi-satellite collaborative computing and task scheduling method includes: Collect network link information and mission request information of each LEO satellite to construct an orbital plane status view in the current time slot; After obtaining heterogeneous state data of each satellite based on the orbital plane state view, the heterogeneous state data of each satellite is uniformly feature-encoded to obtain the feature representation of each satellite; the feature representations of each satellite are aggregated into global state features at the orbital plane level; the satellite is a LEO satellite or a MEO satellite. Based on the global state features, each task to be decomposed is segmented and encoded to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme for each task to be decomposed. The system-level reward value is determined by calculating the task completion rate of each priority group based on the action vector corresponding to each task to be decomposed and the preset task priority. Update the satellite status according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane status view of the next time slot based on the updated satellite status. Then return to execute the step of obtaining the heterogeneous status data of each satellite based on the orbital plane status view. After iterating through all time slots, obtain the action vector and system-level reward value corresponding to each task to be decomposed in each time slot. The task scheduling scheme is obtained based on the action vectors corresponding to each task to be decomposed in each time slot. The process of uniformly encoding the heterogeneous state data of each satellite to obtain the feature representation of each satellite includes: Using the Transformer architecture as the backbone network, the heterogeneous state data of each satellite are processed through a dynamic embedding layer, a multi-head self-attention layer, and a feature aggregation layer. Through the internal dynamic embedding, self-attention computation, and feature aggregation mechanisms, the feature representations of each satellite are obtained.

2. The multi-satellite collaborative computing and task scheduling method according to claim 1, characterized in that, Based on the global state features, each task to be decomposed is segmented and encoded to obtain the action vector corresponding to each task, including: The number of decompositions for each task to be decomposed is determined based on the global state characteristics and each task to be decomposed. By utilizing global state features and the number of decompositions for each task to be decomposed, each task to be decomposed is divided into multiple subtasks, thus obtaining the task decomposition scheme corresponding to each task to be decomposed. Based on the task decomposition scheme of each task to be decomposed, the execution satellite allocation scheme corresponding to each task to be decomposed is determined.

3. The multi-satellite collaborative computing and task scheduling method according to claim 1, characterized in that, The system-level reward value is determined by calculating the task completion rate of each priority group based on the action vector corresponding to each task to be decomposed and the preset task priority, including: Calculate the available computation time for each subtask based on the action vector corresponding to each task to be decomposed and multiple subtasks; Based on preset task priorities, the subtasks in each task to be decomposed are divided into different priority groups; The task completion rate of each priority group is calculated based on the subtasks in each priority group and the available computation time for each subtask. The system-level reward value is determined by weighted summation of the task completion rates of each priority group.

4. The multi-satellite collaborative computing and task scheduling method according to claim 3, characterized in that, The system-level reward value is determined by weighted summation of the task completion rates for each priority group, including: ； in, , Indicates the total number of priority groups; Indicates the first Task completion rate for each priority group; Indicates the first The weights corresponding to each priority group; This represents the system-level reward value.

5. The multi-satellite collaborative computing and task scheduling method according to claim 3, characterized in that, The calculation methods for the available computation time of each subtask include: ； ； in, This represents the transmission delay incurred by each subtask during transmission; This represents the propagation delay incurred by each subtask during transmission; This represents the queuing delay incurred by each subtask during transmission; This represents the total communication delay incurred by each subtask during transmission; This indicates the maximum allowed end-to-end latency for each subtask; This indicates the available computation time for each subtask.

6. The multi-satellite collaborative computing and task scheduling method according to claim 1, characterized in that, Before obtaining the task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot, the multi-satellite collaborative computing and task scheduling method further includes: When the number of time slots in the iteration meets the preset training conditions, the global value estimation technique is used to update the action vectors and system-level reward values ​​corresponding to each task to be decomposed in each time slot, so as to obtain the updated action vectors corresponding to each task to be decomposed in each time slot. The global value estimation technique includes estimating the global value based on the action vectors and system-level reward values ​​corresponding to each task to be decomposed in each time slot, and calculating the generalized advantage estimate and λ-reward based on the obtained global value estimate for updating.

7. A multi-star collaborative computing and task scheduling system suitable for constellation-scale evolution, characterized in that, The multi-satellite collaborative computing and task scheduling system includes: The dynamic state perception module is used to collect network link information and mission request information of each LEO satellite in order to construct an orbital plane state view in the current time slot; A unified feature encoding module is used to obtain heterogeneous state data of each satellite based on the orbital plane state view, perform unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite; and aggregate the feature representations of each satellite into a global state feature at the orbital plane level; wherein the satellite is a LEO satellite or a MEO satellite. The joint strategy decision module is used to segment and encode each task to be decomposed based on the global state characteristics to obtain the action vector corresponding to each task to be decomposed; the action vector includes the number of subtask decompositions, the task decomposition scheme and the execution satellite allocation scheme corresponding to each task to be decomposed. The environment simulation and execution module is used to calculate the task completion rate of each priority group based on the action vectors corresponding to each task to be decomposed and the preset task priorities to determine the system-level reward value. The environment simulation and execution module is also used to update the state of each satellite according to the computing resources used by each task to be decomposed in the current time slot, and generate the orbital plane state view of the next time slot based on the updated satellite state, and then return to execute the step of obtaining the heterogeneous state data of each satellite based on the orbital plane state view, until all time slots are iterated, and the action vector and system-level reward value corresponding to each task to be decomposed in each time slot are obtained. The environment simulation and execution module is also used to obtain a task scheduling scheme based on the action vectors corresponding to each task to be decomposed in each time slot. The unified feature encoding module performs unified feature encoding on the heterogeneous state data of each satellite to obtain the feature representation of each satellite, including: Using the Transformer architecture as the backbone network, the heterogeneous state data of each satellite are processed through a dynamic embedding layer, a multi-head self-attention layer, and a feature aggregation layer. Through the internal dynamic embedding, self-attention computation, and feature aggregation mechanisms, the feature representations of each satellite are obtained.

8. The multi-satellite collaborative computing and task scheduling system according to claim 7, characterized in that, The joint strategy decision module is specifically used to determine the number of decompositions for each task based on the global state characteristics and each task to be decomposed; to divide each task to be decomposed into multiple sub-tasks using the global state characteristics and the number of decompositions for each task to be decomposed, thereby obtaining a task decomposition scheme for each task to be decomposed; and to determine the execution satellite allocation scheme for each task to be decomposed based on the task decomposition scheme for each task to be decomposed.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a computer program stored in memory, it implements the multi-star collaborative computing and task scheduling method for constellation-scale evolution as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the multi-star collaborative computing and task scheduling method for constellation-scale evolution as described in any one of claims 1-6.

Images