Distributed task planning method and system based on low-orbit heterogeneous star cluster and sensing algorithm fusion

Abstract

The application discloses a low-orbit heterogeneous star cluster sensing algorithm fusion distributed task planning method and system, and belongs to the technical field of low-orbit satellite constellation and multi-agent collaborative decision-making. The application aims at the problems of single-point failure of the centralized architecture of the existing low-orbit star cluster task planning, non-uniformity of resource and task representation, weak collaborative conflict resolution capability, and insufficient sensing algorithm fusion scheduling, constructs a unified representation model of sensing algorithm resources and dynamic tasks, realizes a centerless distributed decision-making through a multi-agent collaborative game and a hierarchical conflict resolution mechanism, combines a reinforcement learning optimization decision-making strategy with a Lyapunov convergence constraint, and realizes dynamic closed-loop optimization through multi-cycle iteration. The application does not need a global center node, can significantly improve the star cluster task completion rate, resource utilization efficiency and system resilience, and can be used in complex scenes such as disaster emergency and environmental monitoring.

Description

Technical Field

[0001] This invention belongs to the fields of low-Earth orbit satellite constellation technology, multi-agent collaborative decision-making and synergy computing technology, specifically involving a distributed mission planning method and system for low-Earth orbit heterogeneous constellation synergy computing. Background Technology

[0002] With the rapid development of low-Earth orbit (LEO) satellite constellation technology, single-function satellites are no longer sufficient to meet the demands of time-sensitive tasks such as disaster emergency response, global environmental monitoring, and emergency event handling. The integration of communication, remote sensing, and computing functions in heterogeneous LEO satellite constellations—hereinafter referred to as "sensor-computing fusion constellations"—has become a development trend. Sensor-computing fusion constellations achieve closed-loop collaboration through the coordination of communication, sensing, and computing resources among satellites, enabling sensing to trigger communication, communication to empower computing, and computing to guide new sensing, significantly improving the constellation's mission processing capabilities and service range.

[0003] Mission planning is the core of efficient operation of synergistic computing and communication fusion in low-Earth orbit (LEO) heterogeneous satellite constellations. Its core objective is to achieve optimal matching between synergistic computing resources and mission requirements under constraints such as heterogeneous satellite resources, intermittent inter-satellite communication, sudden mission dynamics, and frequent system disturbances like satellite failures and link interruptions, thereby ensuring the overall mission execution efficiency and stability of the constellation. However, existing LEO constellation mission planning methods still have the following key shortcomings:

[0004] (1) Significant architectural bottlenecks: Most methods adopt a centralized planning architecture, which relies on the ground center or core satellite nodes for overall scheduling. This poses a risk of single-point failure and is difficult to adapt to the expansion of the satellite constellation and the constraints of inter-satellite communication delays. When some satellites or links fail, the entire planning system is easily paralyzed.

[0005] (2) Inconsistent representation models: The description methods of synesthetic computing resources and dynamic tasks are not standardized and lack a unified quantitative representation model, which leads to ambiguity in information interaction between satellites and is prone to planning conflicts.

[0006] (3) Weak collaborative decision-making and conflict resolution capabilities: Existing distributed planning methods mostly adopt simple task allocation strategies, which do not fully consider the autonomous decision-making and interactive negotiation capabilities of satellites. They are difficult to resolve resource competition conflicts among multiple satellites and multiple tasks, and the decision sequence lacks convergence guarantee, which easily leads to decision oscillations.

[0007] (4) Insufficient depth of fusion scheduling: The closed-loop collaborative needs of synergistic computing resources are not fully considered. A "divide and conquer" strategy is often used to schedule certain types of resources, resulting in low resource utilization efficiency and difficulty in leveraging the collaborative advantages of synergistic computing fusion of constellations.

[0008] To address the aforementioned issues, there is an urgent need to design a decentralized distributed collaborative task planning method adapted to the scenario of low-Earth orbit heterogeneous constellations with synergistic computing integration, to break through existing technical bottlenecks and improve the autonomous collaborative capability, disturbance adaptability, and resource utilization efficiency of constellations. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing low-Earth orbit (LEO) constellation mission planning methods and to provide a distributed mission planning method and system for LEO heterogeneous constellations that integrates sensing and computing. This method enables satellite autonomous collaborative decision-making, efficient fusion and scheduling of sensing and computing resources, and adaptive response to dynamic disturbances, thereby improving the constellation mission completion rate, resource utilization efficiency, and system resilience. It is also suitable for complex application scenarios of LEO heterogeneous constellations that integrate sensing and computing.

[0010] The technical solution adopted in this invention is as follows:

[0011] A distributed mission planning method for low-Earth orbit heterogeneous satellite constellations integrating communication, remote sensing, and computing capabilities is disclosed. The low-Earth orbit heterogeneous constellation comprises several heterogeneous satellites with autonomous decision-making, inter-satellite information exchange, and resource awareness capabilities. These heterogeneous satellites at least encompass communication, remote sensing, and computing functions, and there is no global central scheduling node within the constellation. The method includes the following steps:

[0012] S1, each satellite autonomously collects its own synesthetic computing resource status and local synesthetic computing task requirements to construct a unified representation model with two dimensions: resources and tasks.

[0013] S2, each satellite, based on a unified representation model, interacts with resource status, task requirements and constraint satisfaction information through local inter-satellite communication, constructs a decentralized decision-making mechanism for multi-agent collaborative game, generates an initial task allocation scheme, and solves the task resource competition problem through a hierarchical conflict resolution strategy, generates a conflict resolution task allocation scheme, and divides the action feasible space.

[0014] S3, each satellite constructs a deep reinforcement learning agent, which takes resources, tasks and conflict states as inputs, and dynamically optimizes task decision-making strategies, conflict resolution and task allocation schemes within the actionable space. At the same time, the optimization process is converged through Lyapunov stability constraints.

[0015] In S4, each satellite executes the optimized task allocation scheme, collects task execution status and resource consumption information in real time, feeds it back to neighboring satellites through inter-satellite local communication, updates its own resource status and constraint satisfaction, and then returns to S1 to enter the next decision cycle, realizing dynamic iterative closed-loop optimization of task planning.

[0016] Furthermore, S1 specifically includes:

[0017] S1.1, Multi-dimensional representation of synesthetic computing resources: The synesthetic computing resources of each satellite i at time t are abstracted into a resource state vector.

[0018] ;

[0019] The components are defined as follows:

[0020] The communication resource vector represents the communication capability of satellite i at time t. , For available bandwidth between satellite and ground, For available inter-satellite bandwidth, Due to satellite-to-ground transmission delay, This is due to inter-satellite transmission delay;

[0021] The sensing resource vector represents the sensing capability of satellite i at time t. , To sense the load type, To sense the load orientation along the track deflection angle, To sense the load-side sway angle, To perceive the field of view, To perceive resolution. For data acquisition rate, For panchromatic load, For multispectral loads, For hyperspectral loading, For radar payload;

[0022] The resource vector represents the computational capability of satellite i at time t. , For floating-point computing power, For fixed-point computing power, For the computing power of the DSP unit, For the computing power of the FPGA unit, The computing power of the GPU unit, Available memory, Available storage;

[0023] The time attribute vector represents the time constraints of synesthetic computing resources. , These are time constraints for communication, sensing, and computing resources, with each type of resource's time constraint specifically consisting of... , , and Let m be the start and end time of resource availability. Let m be the idle time of resource m at time t;

[0024] Spatial attribute vector: representing the satellite's spatial position. , , and These represent the longitude, latitude, and altitude of satellite i at time t.

[0025] S1.2, Dynamic Task Representation: Each synesthetic task k is represented as a task requirement tensor. This enables a formalized and standardized description of dynamic tasks, with each dimension defined as follows:

[0026] The task type includes communication subtasks. Perception subtask and computational subtasks The requirement vector for each subtask is , , The resource requirements of subtask st. The minimum resource requirements for subtask st. and These represent the start and end time windows for the execution of subtask st;

[0027] The time for users to post tasks;

[0028] To define the task value, a task value function is defined to quantify task priority. Combining time decay and the impact of subtask quality, the expression is: ,in The fundamental value of synesthesia calculation task k, The time decay coefficient, Let the quality weight of subtask st satisfy the following condition: , The quality completion rate of subtask st;

[0029] S1.3, Unified Expression of Resource-Task Constraints: Defining the Constraint Satisfaction Function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is quantified, and a task conflict matrix is ​​defined. Mark tasks with resource contention; constraint satisfaction function and task conflict matrix These are two elements that represent the unified expression of resource-task constraints, where k1 and k2 represent two synesthetic computation tasks.

[0030] Furthermore, S1.3 specifically includes:

[0031] S1.3.1, Define the constraint satisfaction function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is expressed as:

[0032] ;

[0033] in, , and These are the weighting coefficients for communication, sensing, and computing resources, respectively, satisfying... ; For a single type of resource m, the satisfaction degree The interval mapping method is used for calculation, with a value range of [0,1]. The single-type resource satisfaction is divided into four types: positive resource satisfaction, negative resource satisfaction, interval resource satisfaction, and Boolean resource satisfaction. The calculation method for each type is as follows:

[0034] Positive resource satisfaction: ;

[0035] Negative resource satisfaction: ;

[0036] Interval resource satisfaction: ;

[0037] Boolean resource satisfaction: ;

[0038] S1.3.2, Define the task conflict matrix When synesthetic computing tasks k1 and k2 compete for the same type of resources on the same satellite and their time windows overlap, Otherwise, it is 0.

[0039] Furthermore, S2 specifically includes:

[0040] S2.1, Construction of Multi-Agent Cooperative Game Model: Treat each satellite as an agent, construct a non-cooperative game model based on the inter-satellite local communication interaction resource status, task requirements and constraint satisfaction information, and solve the Nash equilibrium to generate the initial task allocation scheme;

[0041] S2.2, Layered conflict resolution based on negotiation and arbitration mechanism: For task conflicts in the initial allocation scheme, a two-stage resolution strategy of "local negotiation-priority arbitration" is adopted to obtain a conflict-free task allocation scheme;

[0042] S2.3, Generation of reinforcement learning action feasible space: Based on the conflict-free task allocation scheme, a reinforcement learning action feasible space is constructed for each satellite, which is divided into a baseline reserved action set and a negotiable adjustable action set.

[0043] Furthermore, S2.1 includes:

[0044] Treat each satellite i as an agent and construct a non-cooperative game model. ,in:

[0045] N is the set of participants. where n is the total number of satellites;

[0046] A represents the action space. The action space of each agent i ; 1≤i≤n;

[0047] For information set, , , Let be the set of neighboring satellites connected to satellite i at time t. Information on the constraint satisfaction of m resources for neighboring satellite j;

[0048] U is the utility function, and the utility function for each agent i is... Taking into account the task value, resource consumption, and conflict costs, its expression is:

[0049] ;

[0050] in, For the action combination of agent i, For the action combination of agents other than agent i;

[0051] For the benefit of performing the action, For satellite i, this is the combination of actions to perform the synaptic computing task k.

[0052] The resource consumption cost of performing the action. This represents the resource consumption of m when task k is executed. To avoid the minimum value where the denominator is zero;

[0053] The cost of resource conflict, The maximum value of all tasks at time t;

[0054] and These are the weighting coefficients;

[0055] The satellite determines its initial mission decision by solving the following Nash equilibrium function, i.e., finding the optimal combination of actions. , such that for any agent i, For all Established, This represents the optimal combination of actions for agent i. Let represent the optimal action combination for all agents except agent i. Based on the variational inequality method, this problem is transformed into the following optimization model:

[0056] The initial satellite mission allocation scheme is obtained by solving the inequality using the projective gradient descent algorithm. .

[0057] Furthermore, S2.2 specifically includes:

[0058] S2.2.1, Partial Negotiation Phase: Regarding the initial task allocation scheme In the context of task conflicts, where two or more different satellites perform the same synesthetic computing task k, a negotiation matrix is ​​constructed based on the source surplus of the satellites involved in the synesthetic computing task k. ,in Let be the set of remaining resources of type m for satellite i at time t, and let be a one-to-one correspondence between type m resources and subtask st of synesthesia task k;

[0059] S2.2.2, Priority Arbitration Phase: Calculating Task Priorities Based on the Negotiation Matrix The expression is: The satellite with higher priority retains the right to execute the mission, while the other satellite adds the mission to the negotiated mission pool. And broadcast to neighboring satellites, so that other constraints can be satisfied. Satellite competition execution, forming a conflict resolution task allocation scheme .

[0060] Furthermore, S2.3 specifically includes:

[0061] S2.3.1, based on the task allocation scheme Based on this, the full set of actions of satellite i is divided into: the baseline retained action set. With negotiable adjustment of action sets These two categories together constitute the feasible space of actions. ; Baseline retained action set Task allocation scheme The "accept task" actions already assigned to satellite i, those without resource conflicts, and those with a constraint satisfaction degree greater than or equal to the threshold corresponding to high-matching tasks are included in the mandatory reservation interval of the feasible space; the action set can be negotiated and adjusted. Task allocation scheme Unassigned negotiation task pool Actions corresponding to tasks within the task, low-matching assigned tasks with constraint satisfaction less than the threshold, and non-core subtasks without strong temporal dependencies are included in the action exploration space.

[0062] S2.3.2, the feasible space for all satellite actions is When the resource status, mission conflict matrix, and neighboring satellite topology of satellite i change, the feasible region is updated synchronously. , 1≤i≤n, so that the action feasible space always remains consistent with the current conflict-free task allocation benchmark.

[0063] Furthermore, S3 specifically includes:

[0064] S3.1 defines the state space, action space, and reward function as follows:

[0065] State space S: ,in The resource status of satellite i after normalization. These are communication resource vectors. Perceived resource vectors and compute resource vector The normalized information is the value obtained by dividing each type of resource by its maximum value. The statistical characteristics of the task value distribution include the mean, variance, and maximum value, and are 3-dimensional. The conflict intensity of tasks corresponding to the three types of resources: communication, sensing, and computing. K represents the total number of tasks, and the task size is 3-dimensional.

[0066] Action space A: Action space of the game model Consistent, the action space is the action-feasible space generated by S2.3. ;

[0067] reward function Based on the game utility function, long-term cumulative reward and penalty are introduced, and its expression is:

[0068] ,

[0069] in, It is a utility function; This is a long-term collaborative reward item, guiding agents to consider long-term global collaborative effects. Let τ be the global total utility within the neighborhood. It represents the theoretically optimal global utility, where T is the planning time domain. Spatial penalties are imposed for resource overload and inaction. This is a resource overload penalty item, among which These represent the used amount and total amount of resources m in satellite i, respectively; As a spatial penalty item for non-actionable actions, These are the weighting coefficients;

[0070] S3.2, Agent Construction and Training: The agent network is trained using the asynchronous actor-critic algorithm. The specific steps are as follows:

[0071] An asynchronous actor-critic algorithm is used to train the agent network, employing a dual-network collaborative architecture where the two networks share underlying features but output decisions and evaluations separately. The network structure includes an input layer, hidden layers, and an output layer. The input layer has a dimension of 17; the hidden layers are two fully connected layers with 256 nodes and the activation function is ReLU; the output layer includes an actor network branch and a critic network branch. The actor network branch outputs the probability distribution of each legal action based on the current input state. The system employs the Softmax activation function, with the output dimension equal to the number of tasks to be assigned. The training objective is to maximize the long-term cumulative reward. The critic network branch outputs the state value based on the current input state, using a linear activation function to measure the decision-making performance of the actor network branch. The training objective is to make the satellite's prediction of future rewards more and more accurate, providing the actor network branch with an accurate evaluation basis.

[0072] Output of the actor network branch In the middle, only in the space of action. Perform random exploration within the scope;

[0073] S3.3, Convergence Verification: During the training process of the agent, Lyapunov stability constraints are added, and the decision sequence is defined. Based on Lyapunov stability theory, Lyapunov functions are constructed. ,in The utility of satellite i under globally optimal decision-making; each time the network parameters are updated, it is checked whether the condition is met. If satisfied, then the decision sequence is... If the convergence reaches a globally coordinated state, update the current parameters; otherwise, if the convergence condition is not met, do not update the current parameters.

[0074] A distributed mission planning system integrating communication, remote sensing, and computing capabilities for a low-Earth orbit heterogeneous satellite constellation comprises several heterogeneous satellites with autonomous decision-making, local inter-satellite information exchange, and resource awareness capabilities. These heterogeneous satellites at least encompass communication, remote sensing, and computing functions, and there is no global central scheduling node within the constellation. Each satellite includes:

[0075] The unified representation module is used to execute S1 as described above;

[0076] The collaborative decision-making and conflict resolution module is used to execute S2 as described above;

[0077] The reinforcement learning decision-making module is used to execute S3 as described above;

[0078] The closed-loop optimization module is used to execute S4 as described above.

[0079] Compared with the prior art, the present invention has the following significant advantages:

[0080] (1) This invention breaks through the single-point bottleneck of traditional centralized mission planning. It adopts a decentralized distributed architecture, which does not require ground center or core satellite node coordination. Each satellite autonomously and collaboratively completes mission planning, significantly improving system resilience. Even if some satellites fail or links are interrupted, the mission planning system can still operate normally, avoiding overall paralysis.

[0081] (2) This invention constructs a unified representation model of synesthetic computing resources and dynamic tasks, which solves the problems of information interaction ambiguity and planning conflict caused by resource heterogeneity and non-standard task description in existing distributed planning. It realizes the standardization and quantitative description of resources and tasks, provides a consistent information basis for decentralized collaborative decision-making, and improves the accuracy and efficiency of decision-making.

[0082] (3) This invention innovates a decentralized decision-making mechanism that integrates multi-agent collaborative game and online reinforcement learning. Combined with a hierarchical conflict resolution strategy, it not only solves the problem of task resource competition, but also ensures the convergence of the decision sequence through Lyapunov stability theory, realizing a planning result that emerges from local interactions and is close to the global optimum, and adapts to the constraints of inter-satellite communication discontinuity and local information lag.

[0083] (4) This invention realizes the deep integration and scheduling of synergistic computing resources. It considers the closed-loop collaborative needs of "sensing triggers communication, communication empowers computing, and computing guides new sensing", optimizes resource utilization efficiency, improves the completion rate and execution quality of constellation tasks, adapts to the complex application scenarios of synergistic computing integration of low-orbit heterogeneous constellations, and can be widely used in high-time-efficiency tasks such as disaster emergency response and global environmental monitoring.

[0084] (5) The technical solution of the present invention is highly feasible. Based on existing low-orbit satellite technology and artificial intelligence algorithms, it does not require a large number of new hardware devices. It can be deployed in existing low-orbit heterogeneous constellations through software upgrades. It has good scalability. When the number of satellites increases, it can achieve efficient scheduling through dynamic domain collaboration, which can meet the needs of large-scale constellation development. Attached Figure Description

[0085] Figure 1 This is an overall flowchart of the method of the present invention.

[0086] Figure 2 This is a diagram illustrating the effect of implementing the present invention. Detailed Implementation

[0087] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes the emergency response to floods in the Yangtze River Basin as an application scenario to fully describe the method of the present invention.

[0088] 1. Simulation Environment and Platform Configuration

[0089] The simulation platform used in this embodiment is a commercial / open-source platform commonly used in the field of low-Earth orbit satellites, as detailed below:

[0090] Satellite orbit and scene simulation: STK 12.0, used to construct low-Earth orbit constellation orbit configurations, transit windows, inter-satellite link topologies, and target observation scenarios;

[0091] Satellite communication simulation: OPNET Modeler is used to simulate the bandwidth, delay, and intermittent connectivity characteristics of low-Earth orbit inter-satellite and satellite-to-ground communication, reproducing the real inter-satellite communication environment.

[0092] Decentralized task planning: Python 3.10 + PyTorch 2.0, used to implement synesthetic resource representation, multi-agent non-cooperative game solving, hierarchical conflict resolution, training and inference of asynchronous actor-critic reinforcement learning agents, etc.

[0093] 2. Configuration of basic simulation parameters

[0094] (1) Low-Earth orbit heterogeneous star cluster configuration parameters

[0095] This embodiment constructs a centerless constellation of 12 heterogeneous low-Earth orbit satellites. The satellites are in a 550km altitude sun-synchronous orbit with an inclination of 97.4°. The right ascension of the ascending node is evenly distributed. The constellation covers the entire territory of my country and surrounding areas. Specific node configuration:

[0096] Four high-resolution optical remote sensing satellites: Panchromatic PAN camera resolution The swing angle of the sensing load side is 0.5m. The angle is 25°, and the sensing load is oriented along the track deflection angle. =0, perceived field of view The angle is 1.5°, and the data acquisition rate is... The inter-satellite laser communication bandwidth is 1200 Mbps. 1024Mbps, inter-satellite transmission delay 300ms, satellite-to-ground communication bandwidth 450Mbps, satellite-to-ground transmission delay 150ms, available storage It is 8,000,000 MB;

[0097] Three broadband communication relay satellites: inter-satellite laser communication bandwidth 5120Mbps, inter-satellite transmission latency of 300ms, satellite-to-ground communication bandwidth The satellite-to-ground transmission delay is 800 seconds. 150ms, onboard computing power 2 TFLOPS of available memory 4096MB available storage It is 100,000 MB;

[0098] Computing satellites on 3 stars: Onboard floating-point computing power 256 TFLOPS, fixed-point computing power 512 TOPS, computing power of the DSP unit 200 GFLOPS, computing power of an FPGA unit 512 TOPS, computing power of a GPU unit 256 TFLOPS, inter-satellite communication bandwidth 2048 Mbps, available memory 40960MB available storage It is 100,000 MB;

[0099] Two syn-sensing and computing integrated satellites: Panchromatic PAN camera resolution The swing angle of the sensing load side is 1m. The angle is 15°, and the angle of the sensing load along the track is deflected. =0, perceived field of view The angle is 1.5°, and the data acquisition rate is... The inter-satellite laser communication bandwidth is 300 Mbps. 1024Mbps, inter-satellite transmission delay 300ms, satellite-to-ground communication bandwidth 300Mbps, satellite-to-ground transmission delay The processing time is 150ms, and the computing power of the onboard GPU is [missing information]. 2 TFLOPS available memory 4096MB available storage It has a capacity of 4,000,000 MB and is capable of executing all tasks.

[0100] All satellite nodes are peers, with no global central node or master satellite. They only exchange information locally through inter-satellite links between neighboring nodes, which meets the requirements of the decentralized distributed architecture of this invention.

[0101] (2) Task scenario parameters

[0102] This embodiment adopts a typical task scenario of emergency response to floods in the Yangtze River Basin, which is completely consistent with the application scenario of this invention. The core task requirements are:

[0103] Perception subtask: To perform high-resolution imaging of a 120km×90km disaster area, with a spatial resolution better than 1m, a revisit period of ≤2h, and a data acquisition rate of ≥200Mbps;

[0104] Communication subtask: Real-time transmission of remote sensing raw data and disaster interpretation results back to the ground emergency command center, with a satellite-to-ground transmission bandwidth of ≥100Mbps and an end-to-end transmission delay of ≤500ms;

[0105] Computational subtasks: Onboard tasks include extracting floodwater bodies, interpreting the affected area, and estimating the affected population. The processing latency for a single scene image is ≤30s, the interpretation accuracy is ≥95%, and the minimum required floating-point computing power is ≥1TFLOPS.

[0106] Task value: Base value 1000 points, time decay coefficient 0.05 / hour, value will be zero if the task is not completed within 24 hours after it is released.

[0107] (3) Comparison with the benchmark scheme

[0108] To verify the technical effect of the present invention, three mainstream existing technical solutions are set as comparison benchmarks:

[0109] Benchmark 1: Traditional centralized task planning method;

[0110] Benchmark 2: Distributed Contract Network Task Planning Method;

[0111] Benchmark 3: Divide-and-conquer task planning method based on synesthesia.

[0112] 3. Simulation Implementation Process

[0113] This embodiment presents a distributed task planning method for low-Earth orbit heterogeneous star clusters that integrates sensory computation and data fusion. The decision cycle is 600 seconds, and the entire process follows steps S1-S4. Figure 1 As shown, the details are as follows:

[0114] S1, each satellite autonomously collects its own synesthetic computing resource status and local synesthetic computing task requirements to construct a unified representation model with two dimensions: resources and tasks.

[0115] In this embodiment, at the start time t0 of the decision cycle, all satellite nodes autonomously complete the standardized representation of local resources and tasks.

[0116] S1 specifically includes:

[0117] S1.1, Multi-dimensional representation of synesthetic computing resources: The synesthetic computing resources of each satellite i at time t are abstracted into a resource state vector.

[0118] ;

[0119] The components are defined as follows:

[0120] The communication resource vector represents the communication capability of satellite i at time t. , For available bandwidth between satellite and ground, For available inter-satellite bandwidth, Due to satellite-to-ground transmission delay, This is due to inter-satellite transmission delay;

[0121] The sensing resource vector represents the sensing capability of satellite i at time t. , To sense the load type, To sense the load orientation along the track deflection angle, To sense the load-side sway angle, To perceive the field of view, To perceive resolution. For data acquisition rate, For panchromatic load, For multispectral loads, For hyperspectral loading, For radar payload;

[0122] The resource vector represents the computational capability of satellite i at time t. , For floating-point computing power, For fixed-point computing power, For the computing power of the DSP unit, For the computing power of the FPGA unit, The computing power of the GPU unit, Available memory, Available storage;

[0123] The time attribute vector represents the time constraints of synesthetic computing resources. , These are time constraints for communication, sensing, and computing resources, with each type of resource's time constraint specifically consisting of... , , and Let m be the start and end time of resource availability. Let m be the idle time of resource m at time t;

[0124] Spatial attribute vector: representing the satellite's spatial position. , , and These represent the longitude, latitude, and altitude of satellite i at time t.

[0125] In this embodiment, each satellite autonomously collects its own resource status and constructs a resource status vector. Taking remote sensing satellite No. 1 as an example, the resource vector at time t0 is:

[0126] Communication resource vector: ;

[0127] Perceived resource vector: ;

[0128] Compute resource vectors: ;

[0129] Time attribute vector: The imageable time window for the disaster-stricken area is [t0+120s, t0+180s], with an idle time of 60s;

[0130] Spatial attribute vector: Longitude 116.2°E, Latitude 30.5°N, Altitude 550.2km.

[0131] The remaining 11 satellites simultaneously completed the construction of their own resource state vectors, realizing the standardized representation of heterogeneous resources.

[0132] S1.2, Dynamic Task Representation: Each synesthetic task k is represented as a task requirement tensor. This enables a formalized and standardized description of dynamic tasks, with each dimension defined as follows:

[0133] The task type includes communication subtasks. Perception subtask and computational subtasks The requirement vector for each subtask is , , The resource requirements of subtask st. The minimum resource requirements for subtask st. and These represent the start and end time windows for the execution of subtask st;

[0134] The time for users to post tasks;

[0135] To define the task value, a task value function is defined to quantify task priority. Combining time decay and the impact of subtask quality, the expression is: ,in The fundamental value of synesthesia calculation task k, The time decay coefficient, Let the quality weight of subtask st satisfy the following condition: , The quality completion rate of subtask st;

[0136] In this embodiment, the emergency response to the flood is abstracted into a task requirement tensor, with the core dimension being:

[0137] Task type dimension: Clearly define the resource requirements, minimum resource thresholds, and execution time windows for the three types of sub-tasks: perception, communication, and computing;

[0138] Task time: [t0, t0 + 24 * 3600 s];

[0139] Value dimension: Substituting into the value function ,in The quality completion rates are categorized into three types of sub-tasks.

[0140] S1.3, Unified Expression of Resource-Task Constraints: Defining the Constraint Satisfaction Function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is quantified, and a task conflict matrix is ​​defined. Mark tasks with resource contention; constraint satisfaction function and task conflict matrix These are two elements that represent the unified expression of resource-task constraints, where k1 and k2 represent two synesthetic computation tasks.

[0141] In this embodiment, the weight settings are as follows: The communication weight, perception weight, and computation weight are 0.3, 0.4, and 0.3, respectively.

[0142] Constraint satisfaction calculation: Taking satellite 1 as an example, the satisfaction of sensing resources is 0.95, the satisfaction of communication resources is 0.8, and the satisfaction of computing resources is 0. The overall constraint satisfaction is 0.3×0.8+0.4×0.95+0.3×0=0.62;

[0143] Conflict matrix construction: The time window of the flood perception mission of Satellite 1 overlaps with that of the regular meteorological observation mission. The corresponding conflict matrix position is set to 1, and the other non-conflict positions are set to 0.

[0144] Specifically, S1.3 includes:

[0145] S1.3.1, Define the constraint satisfaction function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is expressed as:

[0146] ;

[0147] in, , and These are the weighting coefficients for communication, sensing, and computing resources, respectively, satisfying... ; For a single type of resource m, the satisfaction degree The interval mapping method is used for calculation, with a value range of [0,1]. The single-type resource satisfaction is divided into four types: positive resource satisfaction, negative resource satisfaction, interval resource satisfaction, and Boolean resource satisfaction. The calculation method for each type is as follows:

[0148] Positive resource satisfaction: ;

[0149] Negative resource satisfaction: ;

[0150] Interval resource satisfaction: ;

[0151] Boolean resource satisfaction: ;

[0152] S1.3.2, Define the task conflict matrix When synesthetic computing tasks k1 and k2 compete for the same type of resources on the same satellite and their time windows overlap, Otherwise, it is 0.

[0153] S2. Based on a unified representation model, each satellite interacts with resource status, task requirements, and constraint satisfaction information through local inter-satellite communication. This constructs a decentralized decision-making mechanism for multi-agent collaborative game, generates an initial task allocation scheme, and resolves task resource competition through a hierarchical conflict resolution strategy. This generates a conflict resolution task allocation scheme and delineates the action feasible space.

[0154] S2 specifically includes:

[0155] S2.1, Construction of Multi-Agent Cooperative Game Model: Treat each satellite as an agent, construct a non-cooperative game model based on the inter-satellite local communication interaction resource status, task requirements and constraint satisfaction information, and solve the Nash equilibrium to generate the initial task allocation scheme;

[0156] Treat each satellite i as an agent and construct a non-cooperative game model. ,in:

[0157] N is the set of participants. where n is the total number of satellites;

[0158] A represents the action space. The action space of each agent i ; 1≤i≤n;

[0159] For information set, , , Let be the set of neighboring satellites connected to satellite i at time t. Information on the constraint satisfaction of m resources for neighboring satellite j;

[0160] U is the utility function, and the utility function for each agent i is... Taking into account the task value, resource consumption, and conflict costs, its expression is:

[0161] ;

[0162] in, For the action combination of agent i, For the action combination of agents other than agent i;

[0163] For the benefit of performing the action, For satellite i, this is the combination of actions to perform the synaptic computing task k.

[0164] The resource consumption cost of performing the action. This represents the resource consumption of m when task k is executed. To avoid the minimum value where the denominator is zero;

[0165] The cost of resource conflict, The maximum value of all tasks at time t;

[0166] and These are the weighting coefficients;

[0167] The satellite determines its initial mission decision by solving the following Nash equilibrium function, i.e., finding the optimal combination of actions. , such that for any agent i, For all Established, This represents the optimal combination of actions for agent i. Let represent the optimal action combination for all agents except agent i. Based on the variational inequality method, this problem is transformed into the following optimization model:

[0168] The initial satellite mission allocation scheme is obtained by solving the inequality using the projective gradient descent algorithm. .

[0169] In this embodiment, the participant set is: a set of intelligent agents consisting of 12 satellites, N={1,2,...,12};

[0170] Action space: The optional actions for each satellite include {accepting the mission, rejecting the mission, negotiating the mission};

[0171] Information set: Each satellite only interacts with its currently connected neighbor nodes. For example, at time t0, the neighbors of satellite 1 are satellites 2, 3, and 5. It only obtains the constraint satisfaction and resource status of these 3 nodes, without global information interaction.

[0172] Utility function: weighting coefficients λ1=0.5, λ2=0.1. Taking into account the task benefits, resource consumption costs and conflict costs, the utility value of satellite 1 accepting the sensing sub-task is 420 points, and the utility value of rejection is 0 points.

[0173] Nash equilibrium solution: The projective gradient descent algorithm is used with 50 iterations and a convergence threshold of 1e-6 to obtain the initial task allocation scheme: Satellites 1 and 4 undertake the sensing sub-task, satellites 6 and 7 undertake the communication sub-task, and satellites 9 and 10 undertake the computation sub-task. After sensing, satellite 1 transmits the data to satellite 7, and satellite 7 transmits it to satellite 9 for computation; after sensing, satellite 4 transmits the data to satellite 6, and satellite 6 transmits it to satellite 10 for computation.

[0174] S2.2, Layered Conflict Resolution Based on Negotiation and Arbitration Mechanism: For task conflicts in the initial allocation scheme, a two-stage resolution strategy of "partial negotiation - priority arbitration" is adopted to obtain a conflict-free task allocation scheme; specifically including:

[0175] S2.2.1, Partial Negotiation Phase: Regarding the initial task allocation scheme In the context of task conflicts, where two or more different satellites perform the same synesthetic computing task k, a negotiation matrix is ​​constructed based on the source surplus of the satellites involved in the synesthetic computing task k. ,in Let be the set of remaining resources of type m for satellite i at time t, and let be a one-to-one correspondence between type m resources and subtask st of synesthesia task k;

[0176] S2.2.2, Priority Arbitration Phase: Calculating Task Priorities Based on the Negotiation Matrix The expression is: The satellite with higher priority retains the right to execute the mission, while the other satellite adds the mission to the negotiated mission pool. And broadcast to neighboring satellites, so that other constraints can be satisfied. Satellite competition execution, forming a conflict resolution task allocation scheme .

[0177] In this embodiment, a two-stage resolution process is performed to address task conflicts in the initial allocation scheme:

[0178] Partial negotiation phase: Satellite 1 and Satellite 22 involved in the conflict mission exchange mission value and remaining resource information through inter-satellite communication to construct a negotiation matrix. Among them, Satellite 1 has a flood mission value of 850 points and 60 seconds of remaining imaging time; Satellite 2 has a meteorological mission value of 300 points and 120 seconds of remaining imaging time.

[0179] Priority arbitration phase: Calculate task priorities. The priority of flood control task is 0.68, and the priority of meteorological task is 0.18. The flood control task has a higher priority. Satellite 1 retains the right to execute. The meteorological task is put into the negotiation pool and broadcast to neighboring nodes. Finally, Satellite 3 competes to take over the task, thus completing the conflict resolution and obtaining the conflict resolution task allocation scheme.

[0180] S2.3, Generation of reinforcement learning action feasibility space: Based on a conflict-free task allocation scheme, a reinforcement learning action feasibility space is constructed for each satellite, divided into a quasi-reserved action set and a negotiable action set; specifically including:

[0181] S2.3.1, based on the task allocation scheme Based on this, the full set of actions of satellite i is divided into: the baseline retained action set. With negotiable adjustment of action sets These two categories together constitute the feasible space of actions. ; Baseline retained action set Task allocation scheme The "accept task" actions already assigned to satellite i, those without resource conflicts, and those with a constraint satisfaction degree greater than or equal to the threshold corresponding to high-matching tasks are included in the mandatory reservation interval of the feasible space; the action set can be negotiated and adjusted. Task allocation scheme Unassigned negotiation task pool Actions corresponding to tasks within the task, low-matching assigned tasks with constraint satisfaction less than the threshold, and non-core subtasks without strong temporal dependencies are included in the action exploration space.

[0182] S2.3.2, the feasible space for all satellite actions is When the resource status, mission conflict matrix, and neighboring satellite topology of satellite i change, the feasible region is updated synchronously. , 1≤i≤n, so that the action feasible space always remains consistent with the current conflict-free task allocation benchmark.

[0183] S3: Each satellite constructs a deep reinforcement learning agent, which takes resources, tasks and conflict states as inputs, and dynamically optimizes task decision-making strategies, conflict resolution and task allocation schemes within the actionable space. At the same time, Lyapunov stability constraints are used to make the optimization process converge.

[0184] S3 specifically includes:

[0185] S3.1 defines the state space, action space, and reward function as follows:

[0186] State space S: ,in The resource status of satellite i after normalization. These are communication resource vectors. Perceived resource vectors and compute resource vector The normalized information is the value obtained by dividing each type of resource by its maximum value. The statistical characteristics of the task value distribution include the mean, variance, and maximum value, and are 3-dimensional. The conflict intensity of tasks corresponding to the three types of resources: communication, sensing, and computing. K represents the total number of tasks, and the task size is 3-dimensional.

[0187] Action space A: Action space of the game model Consistent, the action space is the action-feasible space generated by S2.3. ;

[0188] reward function Based on the game utility function, long-term cumulative reward and penalty are introduced, and its expression is:

[0189] ,

[0190] in, It is a utility function; This is a long-term collaborative reward item, guiding agents to consider long-term global collaborative effects. Let τ be the global total utility within the neighborhood. It represents the theoretically optimal global utility, where T is the planning time domain. Spatial penalties are imposed for resource overload and inaction. This is a resource overload penalty item, among which These represent the used amount and total amount of resources m in satellite i, respectively; As a spatial penalty item for non-actionable actions, These are the weighting coefficients;

[0191] In this implementation, the state space S has a 23-dimensional input, which is fully linked with the outputs of S1 and S2.

[0192] Action Space A: Completely consistent with the action space of the game model. The communication and computing task constraints of satellites 6 and 9 are 0.86 and 0.91 respectively. Therefore, the action space for these two satellites is to accept tasks.

[0193] reward function Weighting coefficient The values ​​are 0.6, 0.4, 0.7, and 0.5 respectively.

[0194] Planning time domain T: set to 100 decision cycles.

[0195] S3.2, Agent Construction and Training: The agent network is trained using the asynchronous actor-critic algorithm. The specific steps are as follows:

[0196] An asynchronous actor-critic algorithm is used to train the agent network, employing a dual-network collaborative architecture where the two networks share underlying features but output decisions and evaluations separately. The network structure includes an input layer, hidden layers, and an output layer. The input layer has a dimension of 17; the hidden layers are two fully connected layers with 256 nodes and the activation function is ReLU; the output layer includes an actor network branch and a critic network branch. The actor network branch outputs the probability distribution of each legal action based on the current input state. The system employs the Softmax activation function, with the output dimension equal to the number of tasks to be assigned. The training objective is to maximize the long-term cumulative reward. The critic network branch outputs the state value based on the current input state, using a linear activation function to measure the decision-making performance of the actor network branch. The training objective is to make the satellite's prediction of future rewards more and more accurate, providing the actor network branch with an accurate evaluation basis.

[0197] Output of the actor network branch In the middle, only in the space of action. Perform random exploration within the scope;

[0198] In this embodiment, the network deployment is as follows: the reinforcement learning network model is 1.2MB in size, and 12 instances are started by the computer to run separately, simulating the intelligent agents of 12 satellites. The single-step inference takes 80ms, which meets the real-time decision-making requirements.

[0199] Training method: The asynchronous actor-critic algorithm is adopted. Each satellite trains its local agent independently. After each decision cycle, the local resources, tasks, conflict states and other samples are shared with neighboring nodes to achieve asynchronous parameter updates. The learning rate is 1e-4 and the discount factor η=0.95.

[0200] Training objectives: Maximize the long-term cumulative reward for the actor network, and minimize the value prediction error for the critic network.

[0201] S3.3, Convergence Verification: During the training process of the agent, Lyapunov stability constraints are added, and the decision sequence is defined. Based on Lyapunov stability theory, Lyapunov functions are constructed. ,in The utility of satellite i under globally optimal decision-making; each time the network parameters are updated, it is checked whether the condition is met. If satisfied, then the decision sequence is... If the convergence is successful, update the current parameters; otherwise, if the convergence condition is not met, do not update the current parameters.

[0202] In this embodiment, when the network iterates to the 8th decision cycle... If the value is consistently less than 0, the decision sequence converges stably to a globally coordinated state without oscillations.

[0203] In S4, each satellite executes the optimized task allocation scheme, collects task execution status and resource consumption information in real time, feeds it back to neighboring satellites through inter-satellite local communication, updates its own resource status and constraint satisfaction, and then returns to S1 to enter the next decision cycle, realizing dynamic iterative closed-loop optimization of task planning.

[0204] In this implementation:

[0205] (1) Each satellite performs its mission according to the final allocation plan: Satellites 1 and 4 complete the imaging of the disaster area, and the raw data are sent to satellites 9 and 10 via satellites 7 and 6 respectively for on-board disaster interpretation. The interpretation results are returned to the ground first.

[0206] (2) During the mission execution, each satellite collects resource consumption and mission completion information in real time and updates its own resource state vector and constraint satisfaction function;

[0207] (3) Dynamic disturbance adaptation: simulate a temporary failure of the onboard computer of satellite 9, with a 50% decrease in computing power, making it unable to complete the computing sub-task. Satellite 9 immediately puts the task into the negotiation pool for broadcast. Neighboring nodes recalculate the constraint satisfaction through the S1-S3 process, and finally satellite 11 takes over the task to complete the dynamic replanning.

[0208] Every 600 seconds, the next decision cycle begins, repeating the S1-S3 process to achieve dynamic iterative optimization of task planning until all tasks are completed.

[0209] 4. Implementation effect verification

[0210] Test data such as Figure 2 As shown;

[0211] (1) Task completion efficiency: All sub-tasks of this flood emergency task were completed within about 3 hours after the release. The task completion rate was 98.2% based on the completion rate of the observed area. Compared with the traditional centralized task planning method, the task completion rate increased by 36% and the task completion efficiency increased by 2.5 times.

[0212] (2) System resilience: In the extreme scenario of one computing satellite failing, this method still completes the mission, while the traditional centralized planning method is completely paralyzed;

[0213] (3) Resource utilization efficiency: The comprehensive utilization rate of synesthesia computing resources reached 82.7%, which is 35% higher than the divide-and-conquer task planning method of synesthesia computing;

[0214] (4) Convergence: The decision sequence converges stably within 8 cycles without decision oscillation, meeting the real-time decision-making requirements of the satellite. The distributed contract network task planning method converges the slowest and oscillates severely.

[0215] The above experimental results show that the present invention solves the core defects of the centralized architecture, such as single point of failure, inconsistent resource and task representation, weak collaborative decision-making and conflict resolution capabilities, and insufficient synergy and computing fusion scheduling proposed in the background technology. It achieves the expected technical effect and has strong feasibility and engineering value.

Claims

1. A distributed task planning method for low-Earth orbit heterogeneous constellation synergistic computing fusion, characterized in that, The low-Earth orbit heterogeneous constellation comprises several heterogeneous satellites with autonomous decision-making, inter-satellite information exchange, and resource sensing capabilities. These heterogeneous satellites at least encompass communication, remote sensing, and computing functions, and there is no global central scheduling node within the constellation. The method includes the following steps: S1, each satellite autonomously collects its own synesthetic computing resource status and local synesthetic computing task requirements to construct a unified representation model with two dimensions: resources and tasks. S2, each satellite, based on a unified representation model, interacts with resource status, task requirements and constraint satisfaction information through local inter-satellite communication, constructs a decentralized decision-making mechanism for multi-agent collaborative game, generates an initial task allocation scheme, and solves the task resource competition problem through a hierarchical conflict resolution strategy, generates a conflict resolution task allocation scheme, and divides the action feasible space. S3, each satellite constructs a deep reinforcement learning agent, which takes resources, tasks and conflict states as inputs, and dynamically optimizes task decision-making strategies, conflict resolution and task allocation schemes within the actionable space. At the same time, the optimization process is converged through Lyapunov stability constraints. In S4, each satellite executes the optimized task allocation scheme, collects task execution status and resource consumption information in real time, feeds it back to neighboring satellites through inter-satellite local communication, updates its own resource status and constraint satisfaction, and then returns to S1 to enter the next decision cycle, realizing dynamic iterative closed-loop optimization of task planning.

2. The distributed task planning method for low-orbit heterogeneous star cluster fusion based on sensory computation as described in claim 1, characterized in that, S1 specifically includes: S1.1, Multi-dimensional representation of synesthetic computing resources: The synesthetic computing resources of each satellite i at time t are abstracted into a resource state vector. ； The components are defined as follows: The communication resource vector represents the communication capability of satellite i at time t. , For available bandwidth between satellite and ground, For available inter-satellite bandwidth, Due to satellite-to-ground transmission delay, This is due to inter-satellite transmission delay; The sensing resource vector represents the sensing capability of satellite i at time t. , To sense the load type, To sense the load orientation along the track deflection angle, To sense the load-side sway angle, To perceive the field of view, To perceive resolution. For data acquisition rate, For panchromatic load, For multispectral loads, For hyperspectral loading, For radar payload; The resource vector represents the computational capability of satellite i at time t. , For floating-point computing power, For fixed-point computing power, For the computing power of the DSP unit, For the computing power of the FPGA unit, The computing power of the GPU unit, Available memory, Available storage; The time attribute vector represents the time constraints of synesthetic computing resources. , These are time constraints for communication, sensing, and computing resources, with each type of resource's time constraint specifically consisting of... , , and Let m be the start and end time of resource availability. Let m be the idle time of resource m at time t; Spatial attribute vector: representing the satellite's spatial position. , , and These represent the longitude, latitude, and altitude of satellite i at time t. S1.2, Dynamic Task Representation: Each synesthetic task k is represented as a task requirement tensor. This enables a formalized and standardized description of dynamic tasks, with each dimension defined as follows: The task type includes communication subtasks. Perception subtask and computational subtasks The requirement vector for each subtask is , , The resource requirements of subtask st. The minimum resource requirements for subtask st. and These represent the start and end time windows for the execution of subtask st; The time for users to post tasks; To define the task value, a task value function is defined to quantify task priority. Combining time decay and the impact of subtask quality, the expression is: ,in The fundamental value of synesthesia calculation task k, The time decay coefficient, Let the quality weight of subtask st satisfy the following condition: , The quality completion rate of subtask st; S1.3, Unified Expression of Resource-Task Constraints: Defining the Constraint Satisfaction Function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is quantified, and a task conflict matrix is ​​defined. Mark tasks with resource contention; constraint satisfaction function and task conflict matrix These are two elements that represent the unified expression of resource-task constraints, where k1 and k2 represent two synesthetic computation tasks.

3. The distributed task planning method for low-orbit heterogeneous star cluster synergy computing fusion according to claim 2, characterized in that, S1.3 specifically includes: S1.3.1, Define the constraint satisfaction function The degree to which the synesthetic computing resources of satellite i at time t meet the requirements of synesthetic computing task k is expressed as: ； in, , and These are the weighting coefficients for communication, sensing, and computing resources, respectively, satisfying... ; For a single type of resource m, the satisfaction degree The interval mapping method is used for calculation, with a value range of [0,1]. The single-type resource satisfaction is divided into four types: positive resource satisfaction, negative resource satisfaction, interval resource satisfaction, and Boolean resource satisfaction. The calculation method for each type is as follows: Positive resource satisfaction: ; Negative resource satisfaction: ; Interval resource satisfaction: ; Boolean resource satisfaction: ; S1.3.2, Define the task conflict matrix When synesthetic computing tasks k1 and k2 compete for the same type of resources on the same satellite and their time windows overlap, Otherwise, it is 0.

4. The distributed task planning method for low-orbit heterogeneous star cluster synergy computing fusion according to claim 3, characterized in that, S2 specifically includes: S2.1, Construction of Multi-Agent Cooperative Game Model: Treat each satellite as an agent, construct a non-cooperative game model based on the inter-satellite local communication interaction resource status, task requirements and constraint satisfaction information, and solve the Nash equilibrium to generate the initial task allocation scheme; S2.2, Layered conflict resolution based on negotiation and arbitration mechanism: For task conflicts in the initial allocation scheme, a two-stage resolution strategy of "local negotiation-priority arbitration" is adopted to obtain a conflict-free task allocation scheme; S2.3, Generation of reinforcement learning action feasible space: Based on the conflict-free task allocation scheme, a reinforcement learning action feasible space is constructed for each satellite, which is divided into a baseline reserved action set and a negotiable adjustable action set.

5. The distributed task planning method for low-orbit heterogeneous star cluster fusion based on inductive computing according to claim 4, characterized in that, S2.1 includes: Treat each satellite i as an agent and construct a non-cooperative game model. ,in: N is the set of participants. where n is the total number of satellites; A represents the action space. The action space of each agent i ; 1≤i≤n; For information set, , , Let be the set of neighboring satellites connected to satellite i at time t. Information on the constraint satisfaction of m resources for neighboring satellite j; U is the utility function, and the utility function for each agent i is... Taking into account the task value, resource consumption, and conflict costs, its expression is: ； in, For the action combination of agent i, For the action combination of agents other than agent i; For the benefit of performing the action, For satellite i, this is the combination of actions to perform the synaptic computing task k. The resource consumption cost of performing the action. This represents the resource consumption of m when task k is executed. To avoid the minimum value where the denominator is zero; The cost of resource conflict, The maximum value of all tasks at time t; and These are the weighting coefficients; The satellite determines its initial mission decision by solving the following Nash equilibrium function, i.e., finding the optimal combination of actions. , such that for any agent i, For all Established, This represents the optimal combination of actions for agent i. Let represent the optimal action combination for all agents except agent i. Based on the variational inequality method, this problem is transformed into the following optimization model: The initial satellite mission allocation scheme is obtained by solving the inequality using the projective gradient descent algorithm. .

6. The distributed mission planning method for low-Earth orbit heterogeneous star cluster fusion based on inductive computing, as described in claim 5, is characterized in that... S2.2 specifically includes: S2.2.1, Partial Negotiation Phase: Regarding the initial task allocation scheme In the context of task conflicts, where two or more different satellites perform the same synesthetic computing task k, a negotiation matrix is ​​constructed based on the source surplus of the satellites involved in the synesthetic computing task k. ,in Let be the set of remaining resources of type m for satellite i at time t, and let be a one-to-one correspondence between type m resources and subtask st of synesthesia task k; S2.2.2, Priority Arbitration Phase: Calculating Task Priorities Based on the Negotiation Matrix The expression is: The satellite with higher priority retains the right to execute the mission, while the other satellite adds the mission to the negotiated mission pool. And broadcast to neighboring satellites, so that other constraints can be satisfied. Satellite competition execution, forming a conflict resolution task allocation scheme .

7. The distributed task planning method for low-Earth orbit heterogeneous star cluster synergy computing fusion according to claim 6, characterized in that, S2.3 specifically includes: S2.3.1, based on the task allocation scheme Based on this, the full set of actions of satellite i is divided into: the baseline retained action set. With negotiable adjustment of action sets These two categories together constitute the feasible space of actions. ; Baseline retained action set Task allocation scheme The "accept task" action already assigned to satellite i, the high-matching task with no resource conflicts and a constraint satisfaction degree greater than or equal to the threshold, is included in the mandatory reservation interval of the feasible space; the action set can be negotiated and adjusted. Task allocation scheme Unassigned negotiation task pool Actions corresponding to tasks within the task, low-matching assigned tasks with constraint satisfaction less than the threshold, and non-core subtasks without strong temporal dependencies are included in the action exploration space. S2.3.2, the feasible space for all satellite actions is When the resource status, mission conflict matrix, and neighboring satellite topology of satellite i change, the feasible region is updated synchronously. , 1≤i≤n, so that the action feasible space always remains consistent with the current conflict-free task allocation benchmark.

8. The distributed task planning method for low-orbit heterogeneous star cluster fusion based on sensory computation according to claim 7, characterized in that, S3 specifically includes: S3.1 defines the state space, action space, and reward function as follows: State space S: ,in The resource status of satellite i after normalization. These are communication resource vectors. Perceived resource vectors and compute resource vector The normalized information is the value obtained by dividing each type of resource by its maximum value. The statistical characteristics of the task value distribution include the mean, variance, and maximum value, and are 3-dimensional. The conflict intensity of tasks corresponding to the three types of resources: communication, sensing, and computing. K represents the total number of tasks, and the task size is 3-dimensional. Action space A: Action space of the game model Consistent, the action space is the action-feasible space generated by S2.

3. ; reward function Based on the game utility function, long-term cumulative reward and penalty are introduced, and its expression is: ， in, It is a utility function; This is a long-term collaborative reward item, guiding agents to consider long-term global collaborative effects. Let τ be the global total utility within the neighborhood. It represents the theoretically optimal global utility, where T is the planning time domain. Spatial penalties are imposed for resource overload and inaction. This is a resource overload penalty item, among which These represent the used amount and total amount of resources m in satellite i, respectively; As a spatial penalty item for non-actionable actions, These are the weighting coefficients; S3.2, Agent Construction and Training: The agent network is trained using the asynchronous actor-critic algorithm. The specific steps are as follows: An asynchronous actor-critic algorithm is used to train the agent network, employing a dual-network collaborative architecture where the two networks share underlying features but output decisions and evaluations separately. The network structure includes an input layer, hidden layers, and an output layer. The input layer has a dimension of 17; the hidden layers are two fully connected layers with 256 nodes and the activation function is ReLU; the output layer includes an actor network branch and a critic network branch. The actor network branch outputs the probability distribution of each legal action based on the current input state. The system employs the Softmax activation function, with the output dimension equal to the number of tasks to be assigned. The training objective is to maximize the long-term cumulative reward. The critic network branch outputs the state value based on the current input state, using a linear activation function to measure the decision-making performance of the actor network branch. The training objective is to make the satellite's prediction of future rewards more and more accurate, providing the actor network branch with an accurate evaluation basis. Output of the actor network branch In the middle, only in the space of action. Perform random exploration within the scope; S3.3, Convergence Verification: During the training process of the agent, Lyapunov stability constraints are added, and the decision sequence is defined. Based on Lyapunov stability theory, Lyapunov functions are constructed. ,in The utility of satellite i under globally optimal decision-making; each time the network parameters are updated, it is checked whether the condition is met. If satisfied, then the decision sequence is... If the convergence reaches a globally coordinated state, update the current parameters; otherwise, if the convergence condition is not met, do not update the current parameters.

9. A distributed mission planning system integrating low-Earth orbit heterogeneous constellation sensing and computation, characterized in that, It includes several heterogeneous satellites with autonomous decision-making capabilities and local inter-satellite information exchange and resource awareness capabilities. These heterogeneous satellites at least encompass communication, remote sensing, and computing functions, and there is no global central scheduling node within the constellation. Each satellite includes: A unified representation module is used to perform S1 as described in any one of claims 1-8; A collaborative decision-making and conflict resolution module, used to execute S2 as described in any one of claims 1-8; A reinforcement learning decision module for performing S3 as described in any one of claims 1-8; A closed-loop optimization module is used to execute S4 as described in any one of claims 1-8.

Images