Satellite mission planning methods, systems, electronic equipment and storage media

By constructing a satellite knowledge twin and a multi-agent collaborative architecture, dynamic adaptability and reliability of satellite mission planning are achieved, solving the problems of low efficiency and insufficient robustness in traditional methods, and improving the mission scheduling capability of large-scale satellite constellations.

CN122865116BActive Publication Date: 2026-11-17XIAN ZHONGKE XIGUANG AEROSPACE TECHNOLOGY GROUP CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611363553.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17
Estimated Expiration
2046-09-04

AI Technical Summary

Technical Problem

Traditional satellite mission planning methods are inefficient and error-prone in large-scale, multi-mission, and highly coupled scenarios, and cannot adapt to dynamic changes, resulting in insufficient reliability and robustness.

Method used

A satellite knowledge twin is constructed for unified modeling, a multi-agent collaborative architecture is deployed for distributed decision-making, conflicts are resolved through negotiation protocols, a coupled model is established and converted into a linear mixed-integer programming model, and the planning scheme is updated in real time.

Benefits of technology

It improved the task constraint satisfaction rate and the utilization rate of the satellite-to-ground arc segment, ensuring that the planning scheme is synchronized with dynamic changes, and improving planning efficiency and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122865116B_ABST
    Figure CN122865116B_ABST
Patent Text Reader

Abstract

This invention relates to the field of spacecraft control technology, specifically proposing a satellite mission planning method, system, electronic equipment, and storage medium. The satellite mission planning method, through the coordinated action of knowledge twin unified modeling, multi-agent large model collaborative decision-making, non-convex constraint accurate convex transformation, large model hot start accelerated solution, formal verification security check, and dynamic closed-loop replanning throughout the entire process, ensures a significant improvement in the mission constraint satisfaction rate, a significant improvement in the utilization rate of the satellite-to-ground arc segment, and a significant improvement in planning efficiency during satellite mission planning. It provides a reliable technical solution for complex mission scheduling scenarios of large-scale satellite constellations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of spacecraft control technology, specifically relating to a satellite mission planning method, system, electronic equipment, and storage medium. Background Technology

[0002] With the rapid development of global aerospace technology, large-scale deployment of satellite constellations has become the mainstream trend. The number of satellite launches and the demand for on-orbit testing and observation missions are growing exponentially, and the contradictions of too many satellites and too few stations, scarce arc segments, and mission coupling are becoming increasingly prominent.

[0003] Traditional satellite mission planning relies mainly on manual pre-scheduling, which depends on engineers' experience to manually allocate tasks and adjust timing. This has drawbacks such as long decision-making cycles, low planning efficiency, susceptibility to errors, and low resource utilization, and cannot adapt to complex satellite scheduling scenarios involving large scale, multiple tasks, and high coupling.

[0004] Most existing satellite mission planning methods achieve planning by modeling a single mission type (such as remote sensing imaging or satellite-to-ground testing) separately. Such methods not only lack a unified constraint paradigm and are difficult to be compatible with multiple mission coupling scenarios such as on-orbit testing, Earth observation, and data transmission, but also cannot adjust the plan in real time in the face of dynamic scenarios such as satellite failures, arc changes, and emergency missions, resulting in insufficient reliability and robustness of satellite mission planning. Summary of the Invention

[0005] This invention aims to at least address the technical problems of poor dynamic adaptability in satellite mission planning, leading to insufficient reliability and robustness in related technologies. Therefore, the purpose of this invention is to propose a satellite mission planning method, system, electronic device, and storage medium.

[0006] Firstly, the present invention proposes a satellite mission planning method, comprising the following steps: constructing a satellite knowledge twin, wherein the satellite knowledge twin is used for unified modeling of missions, resources, and constraints, the satellite knowledge twin includes physical data, domain knowledge, and constraint rules, and is dynamically updated according to the real-time collected satellite execution status; deploying a multi-agent collaborative architecture composed of multiple agents, wherein each agent loads a domain-fine-tuned large language model to achieve autonomous decision-making, and performs distributed collaboration through a negotiation protocol to parse the input natural language task requirements, retrieve historical knowledge, and generate an initial execution strategy, wherein the domain fine-tuning is performed using a space mission planning corpus through a low-rank adaptation method, the negotiation protocol is a contract network protocol, and the multiple agents communicate through a publish-subscribe mechanism, the contract network protocol being used to enable the multiple agents to negotiate to resolve task conflicts; Based on the initial execution strategy, a coupled model is established to define the matching relationship and matching conditions between subtask sliders and time windows. According to the matching relationship and matching conditions, a mathematical optimization model containing decision variables, objective functions, and constraints is constructed. Non-convex constraints in the mathematical optimization model are converted into linear constraints to form a linear mixed-integer programming model. The initial execution strategy is transformed into an initial solution of the linear mixed-integer programming model, and the initial solution is input to the solver for hot-start solving. The solution output by the solver is submitted to a formal verifier for physical constraint compliance verification. If the verification passes, the optimal task orchestration scheme is output. During the execution of the optimal task orchestration scheme, the satellite execution status is collected in real time, and the satellite knowledge twin is dynamically updated based on the satellite execution status to trigger the multi-agent collaborative architecture to re-make collaborative decisions, achieving closed-loop optimization of the planning scheme.

[0007] According to the satellite mission planning method of this invention, firstly, a satellite knowledge twin is constructed. This satellite knowledge twin includes physical data, domain knowledge, and constraint rules, used for unified modeling of tasks, resources, and constraints. It is dynamically updated based on real-time satellite execution status data, providing a unified digital model foundation for subsequent planning processes that integrates physical state, domain knowledge, and historical experience and can evolve synchronously. Then, a multi-agent collaborative architecture is deployed, consisting of multiple domain-fine-tuned large language model agents. Each agent collaborates in a distributed manner through a negotiation protocol, parsing the input natural language task requirements, retrieving historical knowledge, and generating initial... The execution strategy automates the transformation from vague human-language requirements to machine-executable initial solutions. Based on this initial execution strategy, a coupled model is established to define the matching relationship and conditions between subtask sliders and time windows. A mathematical optimization model containing decision variables, objective functions, and constraints is then constructed based on these matching relationships and conditions. Non-convex constraints in the mathematical optimization model are converted into linear constraints to form a linear mixed-integer programming model, enabling the efficient handling of previously difficult-to-solve non-convex programming problems by a standard mixed-integer programming solver. Finally, the initial execution strategy is transformed into the initial solution of the linear mixed-integer programming model and input into the solver. The solver performs a hot-start solution, effectively reducing the search space and accelerating the convergence process. The solution output is submitted to a formal verifier for physical constraint compliance verification, ensuring the final solution is fully feasible in the real physical environment. If the verification passes, the optimal task orchestration scheme is output. Finally, during the execution of the optimal task orchestration scheme, the satellite execution status is collected in real time, and the satellite knowledge twin is dynamically updated based on the satellite execution status to trigger the multi-agent collaborative architecture to re-coordinate decision-making, achieving closed-loop optimization of the planning scheme and keeping the planning scheme synchronized with the dynamically changing actual satellite state. This is achieved through unified modeling and multi-agent collaboration using knowledge twins. The collaborative decision-making of large-scale intelligent models, precise convex transformation of non-convex constraints, accelerated solution through hot start of large models, formal verification and security checks, and dynamic closed-loop replanning throughout the entire process effectively solves the technical problems in traditional satellite mission planning, such as insufficient model versatility making it difficult to be compatible with multiple types of missions, poor dynamic adaptability making it impossible to respond to changes in satellite status in real time, weak constraint processing capabilities leading to limited number of missions to be scheduled, and lack of intelligent collaboration capabilities leading to difficulties in resolving large-scale resource conflicts. It ensures a significant improvement in mission constraint satisfaction rate, a significant improvement in satellite-to-ground arc utilization rate, and a significant improvement in planning efficiency, providing a reliable technical solution for complex mission scheduling scenarios of large-scale satellite constellations.

[0008] In addition, the satellite mission planning method according to embodiments of the present invention may also have the following additional technical features: Furthermore, the construction of the satellite knowledge twin includes: constructing the satellite knowledge twin comprising a physical twin layer, a knowledge twin layer, and a constraint twin layer; wherein, the physical twin layer is used to collect and store real-time physical data of the satellite; the knowledge twin layer is used to define core ontology classes and their semantic relationships of satellite, mission, resources, and constraints based on the physical data using a web ontology language and store them in a graph database to construct a semantic knowledge graph, wherein the semantic knowledge graph is used to define the semantic relationships of satellite, mission, resources, and constraints, and implements symbolic reasoning through a rule engine and neural reasoning through a graph neural network; the constraint twin layer is used to define basic constraints and specific constraints according to the physical data and the semantic knowledge graph, and implements real-time verification and conflict detection of the basic constraints and the specific constraints through a rule engine.

[0009] Furthermore, the deployment of the multi-agent collaborative architecture, consisting of multiple agents, includes: deploying the multi-agent collaborative architecture composed of a task parsing agent, a knowledge reasoning agent, a cooperative scheduling agent, and a solution execution agent; wherein, the task parsing agent is used to receive the natural language task requirements and decompose the natural language task requirements into structured atomic sub-tasks; the knowledge reasoning agent is used to retrieve historical similar tasks through a retrieval enhancement generation mechanism and generate the initial execution strategy based on the retrieval results; the cooperative scheduling agent is used to coordinate resources from multiple satellites and multiple ground stations through the negotiation protocol to resolve task conflicts and determine task and resource allocation schemes; and the solution execution agent is used to perform modeling and solving of the mathematical optimization model according to the task and resource allocation scheme.

[0010] Furthermore, the establishment of a coupled model for defining the matching relationship and matching conditions between subtask sliders and time windows includes: establishing a coupled model comprising a subtask slider model, a time window model, and a mapping relationship; wherein, the subtask slider model is used to decompose a task into multiple subtask sliders, each subtask slider defining its execution duration, executable time range, and required resource attributes; the time window model is used to model the visible arc between the satellite and the ground station as a time window with start and end times, tracking mode, and elevation angle information; the mapping relationship is used to establish matching decision variables between subtask sliders and time windows, the matching decision variables being used to characterize the allocation relationship from task to resource.

[0011] Furthermore, the conversion of non-convex constraints in the mathematical optimization model into linear constraints includes at least one of the following methods: employing the Big-M method, by introducing a constant. MThe non-convex constraint containing a product term is converted into a linear inequality; or, the auxiliary variable method is used to convert the non-convex constraint containing the L0 norm into the linear constraint by introducing a binary auxiliary variable; or, the absolute value function transformation method is used to convert the non-convex constraint containing the absolute value function into a system of linear inequalities by introducing auxiliary variables and binary variables.

[0012] Further, the step of transforming the initial execution strategy into an initial solution of the linear mixed integer programming model, inputting the initial solution into the solver for hot-start solving, and submitting the solution output by the solver to the formal verifier for physical constraint compliance verification includes: generating the initial solution by the knowledge reasoning agent in the multi-agent collaborative architecture, and using the initial solution as the search starting point of the solver; if the solver reports no feasible solution, extracting a subset of conflicting constraints and feeding it back to the knowledge reasoning agent to trigger the knowledge reasoning agent to adjust the initial execution strategy and regenerate a new initial solution for iterative solving; or, if the solver outputs the solution result, but the formal verifier fails to perform physical constraint compliance verification, feeding back the subset of conflicting constraints returned by the formal verifier to the knowledge reasoning agent to trigger iterative adjustment.

[0013] Furthermore, after constructing a mathematical optimization model containing decision variables, objective functions, and constraints, and converting the non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed integer programming model, the method further includes: dividing the tasks into small-scale, medium-scale, and large-scale problems according to the number of tasks, and selecting a hierarchical solution strategy according to the number of tasks; wherein, for the small-scale problems, the solver is directly invoked for exact solution; for the medium-scale problems, a column generation and Benders decomposition strategy is used for solution; for the large-scale problems, an adaptive large neighborhood search heuristic method is used to generate feasible solutions, and the feasible solutions are used as the initial solutions input into the solver for hot-start solution.

[0014] Secondly, based on the same inventive concept, this invention also proposes a satellite mission planning system, comprising: a knowledge twin server for constructing a satellite knowledge twin, wherein the satellite knowledge twin is used for unified modeling of missions, resources, and constraints, and the satellite knowledge twin includes physical data, domain knowledge, and constraint rules, and is dynamically updated according to the real-time collected satellite execution status; and a multi-agent collaborative server, which deploys a multi-agent collaborative architecture composed of multiple agents, each of which loads a domain-fine-tuned large language model to achieve autonomous decision-making and performs distributed collaboration through a negotiation protocol, for parsing the input natural language task requirements, retrieving historical knowledge, and generating an initial execution strategy, wherein the domain fine-tuning is performed using a space mission planning corpus through a low-rank adaptation method, the negotiation protocol is a contract network protocol, and the multiple agents communicate through a publish-subscribe mechanism, the contract network protocol being used to enable the multiple agents to negotiate to resolve task... The system includes a task conflict resolution mechanism; a solution server, which receives the initial execution strategy, establishes a coupled model defining the matching relationship and conditions between subtask sliders and time windows, constructs a mathematical optimization model containing decision variables, objective functions, and constraints based on the matching relationship and conditions, converts non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed-integer programming model; it also converts the initial execution strategy into an initial solution of the linear mixed-integer programming model, inputs the initial solution to the solver for hot-start solving, submits the solution output by the solver to a formal verifier for physical constraint compliance verification, and outputs the optimal task orchestration scheme if the verification passes; and a communication terminal, which collects satellite execution status in real time during the execution of the optimal task orchestration scheme and dynamically updates the satellite knowledge twin based on the satellite execution status to trigger the multi-agent collaborative architecture to re-make collaborative decisions and achieve closed-loop optimization of the planning scheme.

[0015] The satellite mission planning system provided in the second aspect has the same technical features as the satellite mission planning method provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0016] Thirdly, based on the same inventive concept, the present invention also proposes 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; the memory is used to store computer programs; and the processor is used to execute the computer programs stored in the memory to implement the steps of the satellite mission planning method as described in the above embodiments of the present invention.

[0017] The electronic equipment provided in the third aspect has the same technical features as the satellite mission planning method provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0018] Fourthly, based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the satellite mission planning method as described in the above embodiments of the present invention.

[0019] The computer-readable storage medium provided in the fourth aspect has the same technical features as the satellite mission planning method provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a satellite mission planning method according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a satellite mission planning system according to an embodiment of the present invention; Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present invention.

[0022] Figure label: 100-Satellite mission planning system; 110-Knowledge twin server; 120-Multi-agent collaborative server; 130-Solver server; 140-Communication terminal. Detailed Implementation

[0023] The embodiments of the present invention are described in detail below, and the embodiments described with reference to the accompanying drawings are exemplary.

[0024] To address the problems of poor dynamic adaptability in satellite mission planning leading to insufficient reliability and robustness in related technologies, this invention provides a satellite mission planning method, system, electronic device, and storage medium, as described below. Figures 1-3 A satellite mission planning method, system, electronic device, and storage medium according to embodiments of the present invention are described.

[0025] Figure 1 This is a flowchart of a satellite mission planning method according to an embodiment of the present invention. Figure 1As shown, a satellite mission planning method according to an embodiment of the present invention includes the following steps: Step S1: Construct a satellite knowledge twin. The satellite knowledge twin is used to uniformly model tasks, resources, and constraints. The satellite knowledge twin contains physical data, domain knowledge, and constraint rules, and is dynamically updated based on the real-time collected satellite execution status.

[0026] In a specific embodiment, a three-layer satellite knowledge twin architecture is constructed. Specifically, First, a physical twin layer is constructed as the data entry point. Using the OPC-UA (Open Platform Communications Unified Architecture) protocol or the RESTful API (Representational State Transfer Application Programming Interface) interface, six types of physical data are collected in real-time at a sampling frequency of no less than 1Hz. These include satellite orbital parameters (semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, and mean perigee), attitude angles (pitch, yaw, and roll), power output (solar array output power, remaining battery capacity, and load power consumption), storage capacity (remaining solid-state storage capacity and data download rate), and ground station visible arc segments (start and end times, maximum elevation angle, and tracking mode, including S-band USB (Unified S-Band) telemetry and control, X-band data transmission, and Ka-band high-throughput communication). This data is then synchronously written to a time-series database (such as InfluxDB or TimescaleDB) via a CDC (Change Data Capture) mechanism. Subsequently, STK (Systems Tool) is invoked. The kit (system toolkit / satellite simulation toolkit) or autonomous orbit prediction algorithm calculates the visible arc between the satellite and ground station within the future planning period based on synchronized orbit parameters, generates a list of time windows, and achieves prediction accuracy better than 1km (position) and 0.01° (angle). Each window output is... Stored in a time-series database, where, For time window, A unique identifier for the time window. The start and end times of the window. The window start time, The end time of the window. For tracking mode, Maximum elevation angle, This indicates that the ground station antenna is in an available state, and ∈{0,1}, where 0 indicates unavailable and 1 indicates available.

[0027] Secondly, a knowledge twin layer is constructed, building a semantic knowledge graph based on physical data. This layer uses OWL (WebOntology Language) to define core ontology classes such as Satellite, Mission, GroundStation, and Constraint, along with their semantic relationships. The core ontology classes include: Satellite (attributes: orbit type, payload type, power status), Mission (attributes: mission type, priority, execution duration), GroundStation (attributes: geographical location, antenna aperture, supported frequency bands), and Constraint (attributes: constraint type, constraint strength, applicable scenario), and are stored in a graph database (such as Neo4j). Simultaneously, this layer combines the Drools symbolic inference engine to handle hard constraints (such as elevation threshold verification and tracking pattern matching), and GNN (Graph Neural Network). The GNN uses a Graph Neural Network (GNN) for neural reasoning, taking satellite nodes, task nodes, and constraint nodes from the knowledge graph as input. After message passing and aggregation, it outputs the correlation strength between tasks to assist in subsequent task decomposition and conflict resolution. In addition, this layer stores knowledge in the satellite mission planning domain, including mission type rules (on-orbit testing mission rules, Earth observation mission rules, and satellite-to-ground communication mission rules), constraints (timing constraints, resource constraints, uniqueness constraints, elevation angle constraints, tracking mode constraints, and arc segment exclusivity constraints), priority strategies (a dynamic evaluation algorithm for mission priority, which comprehensively considers mission urgency, scientific value, and resource consumption), and historical best solutions (stored in vector form in the Milvus vector database, supporting semantic similarity retrieval).

[0028] Finally, a constraint twin layer is constructed, defining basic constraints (temporal constraints, resource constraints, and uniqueness constraints) and specific constraints (elevation angle constraints: minimum elevation angle ≥ 5°, maximum elevation angle ≤ 90°; tracking mode constraints: telemetry and data transmission tasks do not share the same ground station antenna; arc segment exclusivity constraints: the same ground station only supports one satellite at the same time) based on knowledge graphs and physical data. The constraints are verified in real time and conflict detection is achieved through the Drools rule engine, and the rule format adopts DRL (Drools Rule Language).

[0029] The data flow relationship among the three layers is as follows: real-time telemetry data from the physical twin layer serves as the foundational data source for the knowledge twin layer to construct and update the knowledge graph; the semantic knowledge (satellite capabilities, mission rules, historical schemes) output by the knowledge twin layer serves as the basis for defining and verifying constraints in the constraint twin layer; the set of valid constraints formed after the constraint twin layer passes verification is ultimately passed to the model solution layer to construct the constraints for the mathematical programming model. This satellite knowledge twin is not a static model but is dynamically updated in response to the real-time satellite execution status collected in subsequent steps.

[0030] Specifically, this invention first constructs a three-in-one satellite knowledge twin. By building a physical twin layer, it collects six key physical parameters in real time: satellite orbit, attitude, energy, storage, visible arc segment at ground stations, and tracking mode. These parameters are synchronously written to a time-series database via the CDC mechanism. Then, a high-precision time window list is generated through STK orbit prediction, providing a real-time and accurate satellite status data foundation for the entire system. This solves the problem of delayed or missing physical information in traditional methods and serves as the data basis for dynamic closed-loop optimization. Furthermore, by constructing the knowledge twin layer, it uses OWL ontology to define four core ontology classes and their semantic relationships, integrating scattered domain knowledge, historical experience, and physical data into a structured knowledge graph. This achieves semantic representation of knowledge and efficient reasoning. Simultaneously, this layer combines the Drools symbolic reasoning engine to handle hard constraints and graph neural networks to perform neural reasoning to predict task relevance. This enables the system to utilize domain knowledge in a human-like manner, providing a high-quality knowledge base for the subsequent retrieval and enhancement of large models, fundamentally alleviating the "illusion" problem of large models in professional fields. By constructing a constraint twin layer, basic and proprietary constraints are defined based on knowledge graphs and physical data, and the Drools rule engine is used to realize real-time verification and conflict detection of constraints. Complex business rules are transformed into a unified constraint paradigm that is executable and verifiable by machines, ensuring that all subsequent planning schemes are within the feasible domain from the outset, avoiding invalid calculations for infeasible schemes. The three layers follow a data flow of "physical acquisition, knowledge extraction, and constraint definition," realizing unified modeling and dynamic updating of physical states, domain knowledge, and constraint rules. This provides a general, accurate, and evolvable digital foundation for the entire task planning process, helping to solve the fundamental problems of insufficient modeling versatility and poor dynamic adaptability.

[0031] Step S2: Deploy a multi-agent collaborative architecture consisting of multiple agents. Each agent loads a domain-fine-tuned large language model to achieve autonomous decision-making and conducts distributed collaboration through negotiation protocols to parse the input natural language task requirements, retrieve historical knowledge, and generate an initial execution strategy.

[0032] In a specific embodiment, static deployment and basic configuration need to be completed before the multi-agent collaborative system runs: Four types of agent microservices (Docker containerized, Kubernetes orchestrated) are deployed on the multi-agent collaborative server at the ground control center. Each agent internally loads a large language model finely tuned for the aerospace field (the base model uses LLaMA-2 13B, finely tuned using LoRA (Low-Rank Adaptation), rank r=64, trained on an aerospace mission planning corpus). The inference framework can use accelerated frameworks such as vLLM, with a throughput of no less than 1000 tokens / s. A vector database (such as Milvus) is deployed on the knowledge twin server. Historical planning schemes are pre-encoded into 768-dimensional vectors using the Sentence-BERT model, and an IVF_FLAT (Inverted File Flat) index is established to provide semantic search capabilities for subsequent retrieval. A publish-subscribe mechanism is used between agents, with low-latency message delivery based on the MQTT (Message Queuing Telemetry Transport) protocol. The message format uses Protobuf (Protocol). Buffers (protocol buffers) are serialized and registered with corresponding topics on the MQTT Broker. Task parsing agents register the topic "task / parse", knowledge reasoning agents register the topic "knowledge / reasoning", cooperative scheduling agents register the topic "schedule / negotiate", and solver / execute agents register the topic "solver / execute". Each type of agent initializes its state machine to the idle state. The state transition logic is: idle, reasoning (calling the large model), negotiation (contract network interaction), execution (solving / issuing instructions), and feedback (collecting execution status), triggered by events. A JSON (JavaScript Object Notation) format validator based on ANTLR (Another Tool for Language Recognition) and an SMT (Satisfiability Modulo Theories) solver based on Z3 are deployed as formal validators. Syntax rule files and constraint assertion templates are loaded at system startup.The multi-agent collaborative server employs the Kubernetes HPA (Horizontal Pod Autoscaler) strategy, automatically scaling up the corresponding agent's Pod instances when CPU (Central Processing Unit) utilization exceeds 70%. The knowledge twin server uses a master-slave failover architecture (Keepalived + VRRP (Virtual Router Redundancy Protocol), with a failover time of <3 seconds). Data synchronization utilizes Raft (Raft Consensus Algorithm) to ensure RPO (Recovery Point Objective) = 0 (zero data loss).

[0033] The runtime processing flow unfolds sequentially in the order of "task input, parsing and decomposition, knowledge retrieval and reasoning, collaborative scheduling, modeling and solving, and solution execution": The task-resolving agent (4 instances, each with 2 CPU cores / 8GB memory / 1×A100 GPU) receives the user's natural language task requirements (e.g., "Complete the satellite-to-ground link test of low-Earth orbit constellation A and the Earth observation of region B within the next 24 hours"). Its state automatically transitions from Idle to Reasoning. Using a few-shot learning prompt template, it calls the large model to decompose the task into structured JSON-formatted atomic subtasks. The prompt template contains 2-3 pre-set examples (e.g., {"Task Description": "Satellite-to-Ground Link Test", "Subtask": [{"Subtask ID (Identifier): "T1-1", "Type": "Antenna Pointing Calibration", "Duration": 300, "Dependencies": []}, {"Subtask ID": "T1-2", "Type": "Link Establishment", "Duration": 600, "Dependencies": ["T1-1"]}]}) to guide the large model to output in the same format. The format compliance (including JSON) is verified by the ANTLR parser. The validation checks include schema field integrity, field type correctness, and dependency validity. If the validation fails, the large model will retry (up to 3 times, with the validation error message injected into the prompt template each time). If the validation passes, the data will be passed to the knowledge reasoning agent via MQTT (Topic: task / parse→knowledge / reasoning).

[0034] The knowledge reasoning agent (4 instances, each with 2 CPU cores / 8GB memory / 1×A100 GPU) enters the Reasoning state after receiving the subtask list. It then retrieves historical similar solutions using the RAG (Retrieval-Augmented Generation) mechanism: First, it inputs the text description of the current task (composed of subtask types and constraints) into the Sentence-BERT model and encodes it into a 768-dimensional query vector. An approximate nearest neighbor search is then performed in the Milvus vector database to retrieve the 5 most similar historical planning solutions. Finally, the ReAct (Reasoning + ...) mechanism is used... The Acting reasoning paradigm uses historical solutions as context, concatenating them with the current subtask list and inputting them into a large language model. First, it generates a reasoning idea (e.g., "The current task requires S-band USB telemetry and control, so we need to query ground stations that support S-band"). Then, it executes a query action (e.g., querying the Neo4j knowledge graph for ground stations that support S-band and are currently available). It observes the query results (e.g., "Ground stations G1 and G3 support S-band, and G1 has available arcs in the next 3 hours"). It then continues reasoning based on the observation results, repeating this process until a complete initial execution strategy is generated, including a recommended satellite list, a recommended ground station list, recommended candidate time windows, and constraint adjustment suggestions.

[0035] After receiving the initial execution strategy, the collaborative scheduling agent (6 instances, supporting high concurrency) enters the Negotiating state and conducts distributed negotiation via the Contract Net Protocol: It publishes task announcement messages to the corresponding satellite and ground station agents (as child nodes of the collaborative scheduling agent); each child node agent submits a bid (Bid, including executable time window, resource occupancy, and execution confidence) after evaluating its own resource availability; the collaborative scheduling agent collects all bids and runs an evaluation function (task completion rate weight 0.5 + resource utilization rate weight 0.3 + conflict minimization weight 0.2) to rank the bids, selects the best bidder, and publishes the contract; the winning bidder confirms the contract, and unsuccessful bidders release resources; when multiple tasks compete for the same resource, the higher priority task takes precedence, and if priorities are the same, a second-price sealed auction is used to determine the allocation; after the Contract Net Protocol is completed, the collaborative scheduling agent transmits the determined allocation scheme of "task, satellite, ground station, and candidate time window" via MQTT (Topic: The schedule / negotiate → solver / execute function is passed to the solver execution agent, and its own state returns to Idle.

[0036] Specifically, this embodiment of the invention then deploys a multi-agent large-model collaborative architecture. Before system operation, it completes the containerized deployment of four types of agent microservices (task parsing agent, knowledge reasoning agent, collaborative scheduling agent, and solution execution agent), large model loading and LoRA fine-tuning, vector database index establishment, MQTT communication topic registration, state machine initialization, and formal verifier deployment, etc., and ensures high availability of the system through Kubernetes HPA automatic scaling and Keepalived+VRRP master-slave switching. During runtime, this architecture introduces four large-model agents with distinct responsibilities and domain knowledge to decompose the complex satellite mission planning process into four collaborative stages: parsing, reasoning, negotiation, and solution, thereby achieving automation and intelligence of the planning process. The task parsing agent utilizes the semantic understanding capability of the large model and adopts a few-shot approach. Learning prompt templates accurately translate ambiguous natural language requirements into structured machine instructions, and ANTLR syntax validation ensures output standardization, solving the problem of low efficiency in traditional human-computer interaction. The knowledge reasoning agent forces the large model to reason based on real historical cases through the RAG mechanism and uses the ReAct paradigm to interact with the external knowledge graph of the large model in multiple steps. This ensures that the generated initial execution strategy is not merely an experience-based "guess," but a logically verified and evidence-based "pre-solution," greatly improving the rationality and reliability of the generated initial solution and overcoming the limitations of large models in specialized fields from the outset. The domain is prone to illusions; the collaborative scheduling agent, through the contract network protocol, simulates the bidding-tendering-evaluation-winning mechanism in the market economy. With the aid of the evaluation function and the second-price sealed auction strategy, it realizes distributed autonomous negotiation of resources from multiple satellites and ground stations. Compared with traditional centralized scheduling, it has higher flexibility, robustness and solution efficiency, and can effectively handle large-scale, dynamic resource allocation conflicts. Finally, through the efficient collaboration of multiple agents, it can autonomously complete the entire process from user needs to initial feasible solutions without human intervention, providing a high-quality and highly feasible starting point for subsequent accurate optimization solutions.

[0037] Step S3: Based on the initial execution strategy, establish a coupled model to define the matching relationship and matching conditions between the subtask slider and the time window. According to the matching relationship and matching conditions, construct a mathematical optimization model containing decision variables, objective function and constraints. Convert the non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed integer programming model.

[0038] In a specific embodiment, the solution execution agent (two instances, loading the CPLEX hybrid integer programming solver connection configuration, Z3 constraint assertion template (covering four types of constraints: energy conservation, storage capacity, orbital mechanics, and temporal consistency) and the ALNS (Adaptive Large Neighborhood Search) operator library (containing implementations of three destructive operators and three repair operators) upon system startup) first constructs a task slider-time window coupling model based on the allocation scheme; the model construction follows the order of "first defining the slider, then defining the window, and finally establishing the mapping relationship"; first, a slider model is defined for each subtask, and the subtask slider is uniformly represented as a quintuple: In the formula, For subtask sliders, As a unique identifier for the task, Used as a unique identifier for subtasks. Slider length / execution time The sliding range, The earliest start time, The latest end time, For satellite ID, For ground station ID, Antenna ID The minimum elevation angle is determined; secondly, the visible arc segment data corresponding to the ground stations and satellites involved in the allocation scheme are queried and allocated from the time-series database (initialized in step S1), and a time window model is constructed. The time window is uniformly represented as a quintuple: The start and end times of the window and the maximum elevation angle parameters are pre-calculated using STK or autonomous orbit prediction algorithms, with prediction accuracy better than 1 km (position) and 0.01° (angle). Finally, a mapping relationship is established, and decision variables are defined. ,when When =1, it indicates a subtask. Assign to window Execution and matching require three conditions to be met simultaneously: ① Time inclusion constraint: execution range Must be fully included Window interval [ ],Right now and ② Tracking pattern matching constraints: ③ Elevation angle constraint: Elevation angle of the ground station to the satellite during execution ,in, and These are the minimum and maximum elevation angles required by the subtask slider, respectively, and this interval must be within the corresponding time window simultaneously. Within the range of available elevation angles.

[0039] Subsequently, the agent that solves the problem constructs a complete MIP (Mixed Integer Programming) model based on this model; first, five types of decision variables are defined: binary decision variables... Characterization subtask i Assign to window? j Continuous decision variables Characterization subtask i The actual start time, continuous decision variables Characterization subtask i The actual end time, auxiliary binary variable Characterization subtask i With window j Are there timing conflicts? Auxiliary binary variables Representation task k Whether it was successfully orchestrated; then construct the objective function: In the formula, K For the set of all top-level tasks, For the task k Priority weights (assessed by a comprehensive evaluation of task urgency, scientific value, and resource consumption). This is the conflict penalty coefficient. The task span penalty coefficient is defined as follows: the first term maximizes the completion rate of high-priority tasks, the second term minimizes inter-task conflicts, and the third term minimizes the task execution span. Next, basic constraints are added: subtask order constraints (for subtasks with prerequisite relationships). i and j , must meet ,in Minimum switching time (usually ≥60 seconds) and window matching constraints (if...) ,but and Satellite uniqueness constraint (only one sub-task is executed by the same satellite at the same time). Task non-conflict constraint (if two sub-tasks) i and j If the same resource is used, their execution times must not overlap); and proprietary constraints: satellite feasible set constraints (subtasks) i Can only be assigned to the satellite adapter set Satellites and antenna feasible set constraints (sub-tasks) i Only can be assigned to ground station antenna adapter sets A i (antenna in the middle), maximum elevation angle constraint (subtask) i Elevation angle of the ground station to the satellite during execution ), tracking mode constraints (subtasks) i Tracking mode Must be with window j Tracking mode Consistency), arc segment exclusivity constraint (only one satellite is supported by the same ground station at the same time, i.e. ).

[0040] Finally, for the non-convex forms containing product terms, L0 norms, and absolute value functions in the above constraints, three methods are used to perform convex transformation: (1) Use the Big-M method to transform the product non-convex constraint: for constraints of the form The product term constraint (in the formula, For 0-1 variables, For continuous variables, For window j (At the start time), the agent first determines the start and end times of the planning period. and Calculate a sufficiently large constant. Then, the original product constraint is equivalently transformed into the following two linear inequalities: , In the formula, For window j At the end of the transition, the logic is: when =1 (i.e., subtask) i Indeed assigned to a window j ), constraints are equivalent to and (That is, the subtask must be executed within the window); when =0 (i.e., subtask) i Not assigned to a window j ), = If it is large enough, both inequalities will always hold true, and no constraint effect will be produced.

[0041] (2) Transform L0 norm constraints using the auxiliary variable method: For constraints of the form of L0 norm constraint (representing vector) (There is only one non-zero element in the solution, meaning each subtask can only be assigned to one window), introducing a binary auxiliary variable to solve the execution agent. The L0 norm constraints are equivalently transformed into the following set of linear constraints: , s, where, s The first constraint guarantees that there is only one index for the candidate window. =1, the second constraint guarantees that only if =1, the second constraint guarantees that only if When =1, the corresponding Non-zero values ​​are allowed; for example, if the subtask... k If there are three candidate windows, then , , If only one value is 1, the corresponding window is selected.

[0042] (3) Use absolute value functions to transform timing conflict constraints: for constraints of the form of The absolute value constraint (ensuring the time interval between two tasks is not less than the minimum switching time) is used to solve the problem by introducing auxiliary variables for the executing agent. And introduce a binary auxiliary variable The absolute value constraint {0,1} is equivalently transformed into the following system of linear inequalities: , , , The first two inequalities guarantee (When timing conflicts exist, the task interval is not less than the minimum switching time), the latter two inequalities will affect the auxiliary variables. With the original variable , The correct relationship binding.

[0043] After the transformation, the standard MILP (Mixed Integer Linear Programming) problem is obtained.

[0044] Specifically, this embodiment of the invention then constructs a task slider-time window coupled model and completes the convex transformation; through the abstraction of "task slider" and "time window", the diverse actual tasks (testing, observation, communication) and resources (satellites, ground stations, antennas) are unified into quintuples and the mathematical language of quintuples, solving the problem of insufficient generality of traditional modeling methods; the MIP model constructed on this basis includes five types of decision variables, a triple objective function, and nine types of constraints (4 types of basic constraints + 5 types of specific constraints), which can accurately describe various complex logical relationships and constraints in task planning. More importantly, by employing three mathematical transformation techniques—the Big-M method, the auxiliary variable method, and the absolute value function transformation method—all non-convex constraints in the original problem are precisely converted into linear constraints. This transformation enables the originally NP-hard non-convex programming problem to be solved efficiently and accurately by mature commercial solvers, ensuring both solution quality (global optimality) and solution efficiency. It fundamentally solves the inherent problems of traditional heuristic algorithms being prone to getting trapped in local optima and having low solution efficiency.

[0045] Step S4: Transform the initial execution strategy into the initial solution of the linear mixed integer programming model, and input the initial solution into the solver for hot start solving. Submit the solution output by the solver to the formal verifier for physical constraint compliance verification. If the verification passes, the optimal task orchestration scheme is output.

[0046] In a specific embodiment, the solution execution agent transforms the initial task allocation scheme (including task-satellite allocation, task-window allocation, and task start time estimation) generated by the knowledge reasoning agent into an initial solution input format recognizable by CPLEX (i.e., a set of trial solutions that meet the variable type requirements: binary variables). Continuous variables By constraining the slack variables to be non-negative, the solution is injected into the CPLEX solver as the starting point for its branch and bound search. CPLEX uses this initial solution as the starting point during the branch and bound search process, which can significantly reduce the search space. Experiments show that for medium-sized problems (100-300 tasks), hot start can accelerate the solution by 30%-70%.

[0047] Subsequently, the solution-executing agent selects a hierarchical solution strategy based on the problem size. The solution server is equipped with the mixed integer programming solver CPLEX, configured with high-performance computing nodes (AMD EPYC 9654 96 cores × 2 / 1TB memory / 8TB NVMe SSD), and the solution results are cached in a Redis (Remote Dictionary Server) cluster (3 nodes). The solution server uses a task queue (RabbitMQ, RabbitMQ...). A queue (Rabbit message queue) is used for load balancing. MIP problem instances to be solved are entered into the queue. The solver scheduler pulls tasks from the queue and allocates solving resources based on the number of idle cores in the current CPLEX solver. Small-scale problems (≤50 tasks) are solved precisely by directly calling the CPLEX solver (using branch and bound algorithm + cutting plane augmentation), with a maximum solution time of 300 seconds. Medium-scale problems (50 < ≤ 500 tasks) use column generation and Benders decomposition strategies to decompose the original problem into a main problem (task-resource allocation, integer variables, solved using CPLEX) and subproblems (time window scheduling of each resource, continuous variables, solvable quickly). Large-scale problems (>500 tasks) first use an adaptive large neighborhood search (ALNS) heuristic algorithm to quickly generate... For the initial feasible solution, ALNS maintains various destructive operators (random removal, worst-case removal, and associated removal) and repair operators (greedy insertion, regretful insertion, and adaptive insertion). A roulette wheel mechanism dynamically adjusts the selection probability based on the historical performance of each operator, with operators that perform well having a higher probability of being selected. The initial feasible solution generated by ALNS is then used as the initial solution input for CPLEX during a warm-start process, assisting CPLEX in finding a better solution within the time box. When CPLEX times out or experiences a memory overflow, the feasible solution generated by ALNS is directly used as the final output (this solution has been guaranteed to be feasible during the ALNS phase). The ALNS solution simultaneously serves as both a "warm-start seed" and a "timeout backup solution." A time box mechanism is implemented for all scales (online planning ≤ 5 minutes, offline optimization ≤ 2 hours), and the current best feasible solution is returned upon timeout.

[0048] After the CPLEX solver outputs the solution, the solver agent does not directly issue the solution; instead, it first submits the solution to the Z3-based SMT formal verifier (which has been deployed and loaded with constraint assertion templates at system startup). The verifier then performs four checks in sequence: energy conservation check (verification...). The total energy consumed by all sub-tasks does not exceed the satellite's available energy budget, where, For subtasks i Average power consumption during execution (Total energy budget available for the satellite), storage capacity verification (validation) The total amount of data generated by all sub-tasks does not exceed the available storage capacity of the satellite, where, For subtasks i The amount of data generated The process includes: verifying the satellite's total available storage capacity, orbital mechanics (using SGP4 (Simplified General Perturbations Version 4) / SDP4 (Simplified Deep-space Perturbations Version 4) orbital prediction models to verify the satellite's position and velocity at the mission execution time, ensuring correct elevation angle calculation), and timing consistency verification (verifying that the start time of all sub-tasks is greater than or equal to the earliest start time, the end time is less than or equal to the latest end time, and the end time of the preceding task plus the switching time is less than or equal to the start time of the following task). If all verifications pass, the plan enters the deployment phase. The plan includes the assigned satellite ID, assigned ground station ID, assigned antenna ID, and actual execution time window for each sub-task. The tracking mode stores the data in a structured format in a Redis cache (for subsequent solution comparison) and a Neo4j historical solution library (for future RAG retrieval).

[0049] If the CPLEX solver reports no feasible solution, the solution execution agent extracts the IIS conflict set from CPLEX and feeds it back to the knowledge reasoning agent. If CPLEX outputs the optimal solution but the Z3 verifier fails to validate it (indicating that the solution violates implicit constraints, such as elevation angle errors found in orbital mechanics calculations), Z3 returns a subset of conflict constraints, which is also fed back to the knowledge reasoning agent. Both failure paths trigger an iterative feedback loop. The knowledge reasoning agent adjusts its strategy, regenerates the initial solution, and iterates, with an upper limit of 5 iterations. In addition, the system establishes a large model safety assistance mechanism: outputs a confidence score (0-1), below which a score below a threshold is set (≥0.6 for regular tasks, ≥0.8 for high-priority tasks). Decisions involving emergency tasks with a priority of ≥0.9 trigger manual review; the HITL (Human-in-the-Loop) mechanism generates a decision report (including similar historical cases, constraint conflict analysis, and risk assessment) for critical decisions (deleting tasks with a priority of ≥4, changing the execution order of emergency tasks, and migrating tasks across constellations), which is submitted to the ground control center engineer for confirmation. After confirmation timeout (default 30 minutes), the alternative plan is automatically executed; the constraint sandbox mechanism pre-executes all large model outputs in a virtual satellite environment (based on the STK simulation engine, with a simulation step size of 1 second and a simulation duration covering the entire planning cycle). Real commands can only be issued after the simulation execution results are verified.

[0050] Specifically, in this embodiment of the invention, a large-scale model hot start and constraint verification are then performed. First, the initial solution generated by the large model is used to hot start the MIP solver. The initial task allocation scheme is transformed into an experimental solution format containing binary variables, continuous variables, and constraint relaxation variables, which is then injected into CPLEX. Because the high-quality initial solution greatly reduces the solver's search space, it can converge to the global optimum more quickly. This combines the "intuitive" intelligence of the large model with the "logical" precision of the MIP solver, achieving complementary advantages. Second, a hierarchical solution strategy is designed: for small-scale models, CPLEX is directly called for precise solution; for medium-scale models, column generation and Benders decomposition are used for acceleration; and for large-scale models, ALNS heuristics are used to quickly generate initial feasible solutions before hot starting CPLEX. The ALNS solution simultaneously serves as both a "hot start seed" and a "timeout backup solution," working in conjunction with... The timebox mechanism ensures real-time performance, enabling efficient solutions to problems of varying scales. Furthermore, a formal verifier based on the Z3 solver is introduced as a "safety valve," performing four independent physical constraint compliance checks on the solution results. This ensures that the solution delivered to the satellite is not only mathematically optimal but also fully feasible in the real physical world, eliminating the risks associated with model errors or implicit constraint violations from an engineering perspective. Finally, when a solution fails or verification fails, the system automatically extracts the IIS conflict set and feeds it back to the large model for strategy adjustment, with an iteration limit of 5 times. Combined with confidence scoring, STK constraint sandbox, and other multiple safety mechanisms, a safe closed loop of "intelligent generation - accurate solution - safe verification - feedback correction" is formed, greatly improving the system's robustness and solution success rate, ensuring a safe and feasible solution can be output in any complex scenario.

[0051] Step S5: During the execution of the optimal task orchestration scheme, the satellite execution status is collected in real time, and the satellite knowledge twin is dynamically updated according to the satellite execution status to trigger the multi-agent collaborative architecture to re-make collaborative decisions and realize the closed-loop optimization of the planning scheme.

[0052] In a specific embodiment, during the execution of the optimal solution, the system runs a dynamic closed-loop optimization process. First, the system defines four types of dynamic triggering conditions, which automatically trigger the replanning process when any of the following events occur: emergency task injection (user submits a high-priority emergency task), satellite failure (energy system failure, payload failure, communication link interruption), arc change (ground station closure due to weather, satellite orbital maneuver), or execution deviation (actual execution time deviates from the planned time by more than a threshold ±30 seconds). The triggering condition detection adopts a combination of threshold rules and machine learning models (isolation forest for anomaly detection, LSTM (Long Short-Term Memory) for time series anomaly prediction). When an emergency task is injected during the execution of the solution, the system first determines whether the currently executing instruction sequence is interruptible: if the current subtask is a non-critical operation (such as data downlink), it is immediately interrupted and replanning is triggered; if the current subtask is an uninterruptible operation (such as orbital maneuver), it waits for the subtask to be completed before triggering replanning to ensure satellite safety.

[0053] Subsequently, the status monitoring module (using OPC-UA protocol and RESTful API technology) collects satellite telemetry data (orbit, attitude, energy, storage, and mission execution status) in real time at a frequency of 1Hz through the telemetry and control link, and uses the collected data as input to the anomaly detection module. After receiving the telemetry data, the anomaly detection module performs dual detection through a threshold rule engine and a machine learning model (isolated forest to detect outliers in multidimensional telemetry data, and LSTM model to predict time-series anomaly trends). If any detection channel identifies an abnormal state, it passes the abnormal event type and the list of affected missions to the satellite knowledge twin update module.

[0054] Upon receiving an abnormal event, the satellite knowledge twin update module synchronously updates the physical data to the time-series database (e.g., marking the faulty satellite as unavailable) and updates the status of the corresponding node in the knowledge graph (by modifying the energy status attribute of the corresponding Satellite node in Neo4j through a Cypher query). After the update is completed, the list of affected task IDs is passed to the multi-agent collaboration layer to trigger replanning.

[0055] The replanning strategy is tiered based on the scope of the change's impact: A contingency plan mechanism (for common failure scenarios, such as single-satellite power failure, single ground station shutdown, and link interruption) pre-generates backup plans during system initialization and stores them in a contingency plan repository (Redis cluster). When a failure occurs, it retrieves and switches to the backup plan within seconds via key-value matching (Key = failure type + affected resource ID), without needing to re-solve the problem; Incremental replanning (for small-scale changes, such as single task time adjustments), when the percentage of affected tasks is <10%, only performs partial replanning on affected tasks, freezing the allocation plans for unaffected tasks (and freezing their corresponding allocation plans). (Variables are fixed to known values), and only the affected tasks are used as variables in the new MIP problem. The collaborative reasoning and solution process of steps S2 to S4 is re-executed; global replanning (for large-scale changes, such as multi-satellite failures or multi-task injections) is triggered when the number of affected tasks accounts for ≥10%, triggering a complete replanning process of "task parsing, collaborative reasoning, coupled modeling, and MIP solution".

[0056] After the new scheme is generated, the scheme comparison module evaluates the difference between the new scheme and the current execution scheme (number of changed tasks / total number of tasks). If the difference is <10%, an incremental update method is used to gradually switch (new instructions are issued for each task); if the difference is ≥10%, a global switch method is used to replace it all at once. Before issuing the instructions, the solving execution agent encapsulates the execution instructions of each subtask in the optimized scheme into the CCSDS (Consultative Committee for Space Data Systems) standard TC (Telecommand) frame format (containing APID (Application Process Identifier), instruction sequence number, execution timestamp, and instruction parameters), and simultaneously sends them to the corresponding satellite through the primary and backup dual links (S-band USB primary link + X-band data transmission backup link). Both links perform CRC32 (Cyclic Redundancy Check). The satellite employs 32-bit cyclic redundancy check (CRC32) verification plus digital signature (RSA-2048) verification, executing only when the checksums and signatures of the two instructions are completely identical. The communication terminal deploys a CCSDS protocol stack, supporting encoding and decoding of TC, TM, and AOS frames, as well as a security mechanism of CRC32 verification plus digital signature (RSA-2048). When both links are interrupted, the satellite automatically switches to emergency storage mode and executes a pre-stored instruction sequence. Once the closed-loop optimization is complete, the status data (time consumption, success rate, scheme difference) of the replanning process is fed back to the historical scheme database of the satellite's knowledge twin, used to optimize the similarity calculation weights for subsequent RAG retrieval, forming a continuous "monitoring, detection, updating, replanning, switching, monitoring" closed loop.

[0057] Specifically, the embodiments of this invention ultimately achieve dynamic closed-loop optimization; through a dual detection mechanism combining four types of dynamic triggering conditions with threshold rules and an isolated forest / LSTM machine learning model, the system acquires the ability to "perceive" changes in the external environment and internal state deviations. LSTM is used for time-series anomaly prediction, and isolated forest is used for multi-dimensional anomaly detection. When an emergency task is injected, the system can intelligently determine whether the current command is interruptible (uninterruptible operations such as orbital maneuvers require waiting for completion), ensuring satellite safety. When changes occur, the system first synchronously updates physical data to the time-series database and Neo4j knowledge graph through the satellite knowledge twin update module, ensuring that the information model upon which subsequent decisions rely is always synchronized with the physical world. This is the information foundation for achieving dynamic adaptability. Furthermore, a three-level re- Planning Strategy: The contingency plan library mechanism ensures a second-level response to the most common faults. Incremental replanning achieves a balance between solution speed and solution stability by freezing the unchanged parts and only re-optimizing the affected tasks. Global replanning is used to deal with major changes, ensuring the system's survivability in extreme situations. After difference assessment and primary / backup dual-link verification, data is distributed in the CCSDS standard TC frame format, automatically switching to emergency storage mode when both links are interrupted. Replanning data is fed back to the historical solution library to optimize RAG retrieval weights. This hierarchical strategy perfectly resolves the contradiction between "rapid response" and "global optimization" in dynamic environments, enabling the system to quickly, stably, and efficiently generate corrective solutions when facing sudden situations such as satellite failures and emergency tasks, fundamentally enhancing the robustness and practicality of the mission planning system.

[0058] Therefore, according to the satellite mission planning method provided in the embodiments of the present invention, a satellite knowledge twin is first constructed. This satellite knowledge twin includes physical data, domain knowledge, and constraint rules, used for unified modeling of tasks, resources, and constraints. It is dynamically updated based on real-time collected satellite execution status, providing a unified digital model foundation for subsequent planning processes that can integrate physical state, domain knowledge, and historical experience and evolve synchronously. Then, a multi-agent collaborative architecture is deployed, consisting of multiple domain-fine-tuned large language model agents. Each agent collaborates in a distributed manner through a negotiation protocol, parsing the input natural language task requirements, retrieving historical knowledge, and... The initial execution strategy is generated, achieving a fully automated transformation from vague human-language requirements to a machine-executable initial solution. Then, based on the initial execution strategy, a coupled model is established to define the matching relationship and conditions between subtask sliders and time windows. A mathematical optimization model containing decision variables, objective functions, and constraints is constructed based on the matching relationship and conditions. The non-convex constraints in the mathematical optimization model are converted into linear constraints to form a linear mixed-integer programming model, enabling the originally difficult-to-solve non-convex programming problem to be efficiently handled by a standard mixed-integer programming solver. Finally, the initial execution strategy is transformed into the initial solution of the linear mixed-integer programming model and output. The solution is fed into the solver for a warm-start solution, which effectively reduces the solver's search space and accelerates the convergence process. The solution output is submitted to a formal verifier for physical constraint compliance verification, ensuring that the final solution is fully feasible in the real physical environment. If the verification passes, the optimal task orchestration scheme is output. Finally, during the execution of the optimal task orchestration scheme, the satellite execution status is collected in real time, and the satellite knowledge twin is dynamically updated based on the satellite execution status to trigger the multi-agent collaborative architecture to re-coordinate decision-making, realizing closed-loop optimization of the planning scheme and keeping the planning scheme synchronized with the dynamically changing actual satellite state. This is achieved through unified modeling using knowledge twins. The system effectively addresses the technical challenges of traditional satellite mission planning, including insufficient model versatility for multi-agent large-scale model collaborative decision-making, precise convex transformation of non-convex constraints, accelerated solution through large-scale model hot start, formal verification for security checks, and dynamic closed-loop replanning. These challenges include poor dynamic adaptability, inability to respond to satellite state changes in real time, limited task scheduling capacity due to weak constraint handling capabilities, and difficulty in resolving large-scale resource conflicts due to a lack of intelligent collaboration capabilities. The system ensures a significant improvement in task constraint satisfaction, a substantial increase in satellite-to-ground arc utilization, and a marked improvement in planning efficiency, providing a reliable technical solution for complex mission scheduling scenarios in large-scale satellite constellations.

[0059] In one embodiment of the present invention, constructing a satellite knowledge twin includes: constructing a satellite knowledge twin comprising a physical twin layer, a knowledge twin layer, and a constraint twin layer; wherein, the physical twin layer is used to collect and store real-time physical data of the satellite; the knowledge twin layer is used to construct a semantic knowledge graph based on the physical data, wherein the semantic knowledge graph is used to define semantic relationships between satellites, missions, resources, and constraints, and supports symbolic reasoning and neural reasoning; the constraint twin layer is used to define basic constraints and proprietary constraints according to the physical data and the semantic knowledge graph, and to implement real-time verification and conflict detection of basic constraints and proprietary constraints through a rule engine.

[0060] In a specific embodiment, the physical twin layer collects six types of physical data in real time at a sampling frequency of no less than 1Hz, including satellite orbital parameters, attitude angles, power output, storage capacity, visible arc segment of the ground station, and tracking mode. This data is written to a time-series database via a CDC mechanism and combined with STK orbit predictions to generate a list of time windows. The prediction accuracy is better than 1km and 0.01°. The server employs Keepalived+VRRP master-slave failover (<3 seconds) and the Raft consensus algorithm (RPO=0). The knowledge twin layer uses the OWL ontology to define Satellite, Mission, GroundStation, and Constraint. Four core ontology classes and their semantic relationships are stored in the Neo4j graph database. The symbolic reasoning with hard constraints is handled by the Drools rule engine, and the neural reasoning with GNN outputs the task association strength after message passing and aggregation. Historical solutions are encoded into 768-dimensional vectors by Sentence-BERT and stored in Milvus. The constraint twin layer defines basic constraints (temporal, resource, uniqueness) and specific constraints (elevation angle ≥5° and ≤90°, tracking modes do not share antennas, arc segment exclusive). The constraints are verified and conflict detected in real time by the Drools rule engine (DRL format, such as elevation angle constraint verification rules). After verification, the constraints are passed to the model solution layer.

[0061] Specifically, this invention implements a physical twin layer to achieve high-frequency acquisition and accurate forecasting of six types of physical data, providing a real-time and accurate satellite status data foundation for the entire system; a knowledge twin layer uses OWL ontology to achieve semantic representation of knowledge, and a hybrid architecture of Drools symbolic reasoning and GNN neural reasoning to achieve complementarity between rule rigor and deep association mining, enabling the system to have human-like decision-making capabilities; a constraint twin layer formalizes six types of complex constraints into a unified paradigm that can be automatically executed by machines in DRL format, moving constraint verification from post-event to in-event, ensuring the physical feasibility and mission safety of the final solution; the three layers follow a progressive data flow of "physical acquisition, knowledge extraction, and constraint definition," fundamentally solving the problems of insufficient modeling versatility and poor dynamic adaptability of traditional methods.

[0062] In one embodiment of the present invention, a multi-agent collaborative architecture consisting of multiple agents is deployed, including: deploying a multi-agent collaborative architecture consisting of a task parsing agent, a knowledge reasoning agent, a cooperative scheduling agent, and a solution execution agent; wherein, the task parsing agent is used to receive natural language task requirements and decompose the natural language task requirements into structured atomic sub-tasks; the knowledge reasoning agent is used to retrieve historical similar tasks through a retrieval enhancement generation mechanism and generate an initial execution strategy based on the retrieval results; the cooperative scheduling agent is used to coordinate resources from multiple satellites and multiple ground stations through a negotiation protocol to resolve task conflicts and determine task and resource allocation schemes; and the solution execution agent is used to perform modeling and solving of a mathematical optimization model according to the task and resource allocation scheme.

[0063] In a specific implementation, the task parsing agent (4 instances) uses a Few-shot Learning prompt template (with 2-3 pre-set examples) to call LLaMA-2 13B (LoRA fine-tuning, rank r=64) to decompose the requirements into atomic subtasks. The JSON Schema fields are verified for completeness, type correctness, and dependency validity using ANTLR, with a maximum of 3 retries for failures. The knowledge reasoning agent (4 instances) encodes the task description into a 768-dimensional vector using Sentence-BERT, performs an approximate nearest neighbor search (Top-K=5) in Milvus to retrieve historical solutions, and generates an initial strategy containing recommendations for satellites / ground stations / time windows using the ReAct paradigm (Thought→Action→Observation loop). The collaborative scheduling agent (6 instances) publishes tenders, receives bids (including time windows, resource usage, and execution confidence) through the Contract Network protocol, and runs the evaluation letter. The system selects the winner and publishes the contract based on the following criteria: task completion rate 0.5 + resource utilization rate 0.3 + conflict minimization 0.2. When multiple tasks compete for the same resource, the one with higher priority takes precedence; if they are the same, a sealed auction with the second-highest price is used. The system solves the execution agent (2 instances) by loading the CPLEX connection configuration, Z3 four-type constraint assertion templates, and the ALNS operator library (3 types of destruction + 3 types of repair operators), and performs modeling, convex transformation, hierarchical solving, and formal verification. The agents communicate with each other via the MQTT protocol publish-subscribe, and the state machine (Idle→Reasoning→Negotiating→Executing→Feedback) drives the collaboration.

[0064] Specifically, this invention addresses the inefficiency of traditional human-computer interaction by employing a task parsing agent with Few-shot Learning prompt templates and ANTLR format verification to achieve precise conversion from fuzzy natural language to standardized instructions. A knowledge reasoning agent uses the RAG mechanism to force large models to reason based on historical cases to suppress illusions and the ReAct paradigm to achieve multi-step interaction with the knowledge graph, ensuring that the generated initial strategy is a well-founded "pre-solution." A collaborative scheduling agent uses the contract network protocol in conjunction with the evaluation function and a second-price sealed auction to achieve fair and efficient distributed resource conflict resolution. A solution execution agent transforms intelligent decision-making into rigorous mathematical optimization, serving as a crucial bridge connecting "agent soft decision-making" and "mathematical hard constraints." The four types of agents collaborate in an orderly manner driven by a state machine, achieving fully automated conversion from fuzzy requirements to feasible solutions.

[0065] In one embodiment of the present invention, a coupled model for defining the matching relationship and matching conditions between subtask sliders and time windows is established, including: establishing a coupled model comprising a subtask slider model, a time window model, and a mapping relationship; wherein, the subtask slider model is used to decompose a task into multiple subtask sliders, each subtask slider defining its execution duration, executable time range, and required resource attributes; the time window model is used to model the visible arc between the satellite and the ground station as a time window with start and end times, tracking mode, and elevation angle information; the mapping relationship is used to establish matching decision variables between subtask sliders and time windows, and the matching decision variables are used to characterize the allocation relationship from task to resource.

[0066] In a specific embodiment, the slider model is defined as a quintuple: The attributes include satellite ID, ground station ID, antenna ID, elevation range, and tracking mode; the window model is defined as a quintuple: The start and end times of the window and the maximum elevation angle are pre-calculated using STK or autonomous orbit prediction; the mapping relationship defines the decision variables. Matching requires meeting three conditions simultaneously: time inclusion, tracking pattern matching, and elevation angle range.

[0067] Specifically, this invention, through a unified mathematical representation of a quintuple slider and a quintuple window, abstracts diverse types of tasks and resources into a "two-sided matching" problem, using decision variables... The three matching conditions precisely mathematically represent the matching relationship, solving the problem of insufficient generality in traditional modeling methods and laying a complete variable foundation for subsequent accurate MIP solutions.

[0068] In one embodiment of the present invention, converting non-convex constraints in a mathematical optimization model into linear constraints includes at least one of the following methods: employing the Big-M method, by introducing a constant.M The non-convex constraints containing product terms can be converted into linear inequalities; or, the auxiliary variable method can be used to convert non-convex constraints containing the L0 norm into linear constraints by introducing a binary auxiliary variable; or, the absolute value function transformation method can be used to convert non-convex constraints containing absolute value functions into a system of linear inequalities by introducing auxiliary variables and binary variables.

[0069] In a specific embodiment, the Big-M method is used for... make , converted and Auxiliary variable method Introduction , converted , Absolute value function transformation method for Introduction and {0,1} can be transformed into a system of four linear inequalities.

[0070] Specifically, the embodiments of the present invention employ the Big-M method to precisely linearize the non-convex product terms coupled with integer and continuous variables, transforming the complex logic of "if assigned, the window must be satisfied" into a simple linear constraint; the auxiliary variable method is used to precisely linearize the L0 norm constraint of "selecting the only one from multiple candidates", avoiding the accuracy loss caused by heuristics or relaxation approximations; the absolute value function transformation method is used to precisely linearize the non-smooth temporal interval constraint, eliminating the solution difficulty caused by non-convexity; all three transformations are mathematically equivalent, enabling NP-hard non-convex programming to be solved efficiently and accurately by the standard MIP solver.

[0071] In one embodiment of the present invention, the initial execution strategy is transformed into an initial solution of a linear mixed integer programming model, and the initial solution is input to the solver for hot-start solving. The solution output by the solver is submitted to the formal verifier for physical constraint compliance verification. This includes: generating an initial solution by the knowledge reasoning agent in the multi-agent collaborative architecture, and using the initial solution as the search starting point of the solver; if the solver reports no feasible solution, extracting a subset of conflicting constraints and feeding it back to the knowledge reasoning agent to trigger the knowledge reasoning agent to adjust the initial execution strategy and regenerate a new initial solution for iterative solving; or, if the solver outputs a solution result, but the formal verifier fails to verify the physical constraint compliance, feeding back the subset of conflicting constraints returned by the formal verifier to the knowledge reasoning agent to trigger iterative adjustment.

[0072] In a specific embodiment, the solution-executing agent transforms the initial allocation scheme into an initial solution format (binary variables). Continuous variables Inject non-negative (constraint relaxation variables) into CPLEX; if CPLEX reports no feasible solution, extract the IIS feedback knowledge reasoning agent; if the Z3 validator fails to verify, return the conflict constraint subset feedback knowledge reasoning agent; both failed paths trigger iterations, up to a maximum of 5 times.

[0073] Specifically, this invention constructs a three-way iterative feedback mechanism of "agent-solver-verifier", opening up a reverse channel from "precise solution / verification" to "intelligent strategy". When precise solution or physical verification fails, the system uses IIS to accurately locate the root cause of the conflict and provide structured feedback to the large model, giving it the ability to "trial and error correction". This allows the model to adjust its strategy based on the conflict information and regenerate the initial solution. The iteration limit is 5 times, realizing the deep coupling between the agent's "soft decision-making" and the mathematical "hard constraints", which greatly improves the robustness and success rate of the solution.

[0074] In one embodiment of the present invention, after constructing a mathematical optimization model containing decision variables, objective function, and constraints, and converting the non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed integer programming model, the method further includes: dividing the tasks into small-scale problems, medium-scale problems, and large-scale problems according to the number of tasks, and selecting a hierarchical solution strategy according to the number of tasks; wherein, for small-scale problems, the solver is directly called for exact solution; for medium-scale problems, a column generation and Benders decomposition strategy is used for solution; for large-scale problems, an adaptive large neighborhood search heuristic method is used to generate feasible solutions, and the feasible solutions are used as initial solutions to input the solver for hot-start solution.

[0075] In specific implementations, when the number of tasks is ≤50, CPLEX (branch and bound + cut plane enhancement) is directly invoked, with a maximum time limit of 300 seconds; when the number of tasks is 50 < ≤500, column generation decomposition is used to decompose the main problem (integer variables, solved by CPLEX) and subproblems (continuous variables, solved quickly), combined with Benders decomposition for acceleration; when the number of tasks is >500, ALNS maintains 3 types of destruction operators (random removal, worst-case removal, and associated removal) and 3 types of repair operators (greedy insertion, regretful insertion, and adaptive insertion), dynamically adjusts the selection probability through a roulette wheel mechanism, generates feasible solutions as the initial solutions for CPLEX hot start, and directly outputs ALNS feasible solutions when timeout or memory overflow occurs; time boxes are set for all scales (online ≤5 minutes, offline ≤2 hours).

[0076] Specifically, the embodiments of the present invention employ a hierarchical solution strategy. For small-scale problems, direct and accurate solutions ensure optimality. For medium-scale problems, column generation and Benders decomposition overcome memory and time bottlenecks. For large-scale problems, a combination of ALNS multi-operator combinations and roulette wheel dynamic selection finds high-quality feasible solutions in a very short time. It also serves as a hot-start seed and a timeout backup solution. Combined with a timebox mechanism, it achieves the optimal balance between solution quality and real-time performance for problems of different scales.

[0077] Further embodiments of the present invention disclose a satellite mission planning system. Figure 2 This is a structural block diagram of a satellite mission planning system according to an embodiment of the present invention. Figure 2 As shown, in one embodiment of the present invention, the satellite mission planning system 100 includes: a knowledge twin server 110, a multi-agent collaborative server 120, a solver server 130, and a communication terminal 140.

[0078] Specifically, the knowledge twin server 110 is used to construct a satellite knowledge twin, which is used to uniformly model tasks, resources, and constraints. The satellite knowledge twin contains physical data, domain knowledge, and constraint rules, and is dynamically updated based on the real-time collected satellite execution status.

[0079] The multi-agent collaborative server 120 is deployed with a multi-agent collaborative architecture consisting of multiple agents. Each agent loads a domain-fine-tuned large language model to achieve autonomous decision-making and conducts distributed collaboration through negotiation protocols. It is used to parse the input natural language task requirements, retrieve historical knowledge, and generate initial execution strategies.

[0080] Solver 130 is used to receive the initial execution strategy, establish a coupled model for defining the matching relationship and matching conditions between subtask sliders and time windows, construct a mathematical optimization model containing decision variables, objective functions and constraints based on the matching relationship and matching conditions, convert non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed integer programming model; it is also used to convert the initial execution strategy into the initial solution of the linear mixed integer programming model, input the initial solution into the solver for hot-start solving, and submit the solution output by the solver to the formal verifier for physical constraint compliance verification. If the verification passes, the optimal task orchestration scheme is output.

[0081] The communication terminal 140 is used to collect the satellite execution status in real time during the execution of the optimal task scheduling scheme, and dynamically update the satellite knowledge twin based on the satellite execution status, so as to trigger the multi-agent collaborative architecture to re-make collaborative decisions and realize the closed-loop optimization of the planning scheme.

[0082] It should be noted that the specific implementation of the satellite mission planning system 100 in this embodiment of the invention is similar to the specific implementation of the satellite mission planning method described in the above embodiments of the invention, and therefore has similar technical effects. For details, please refer to the description in the satellite mission planning method section. To reduce redundancy, it will not be repeated here.

[0083] Further embodiments of the present invention also disclose an electronic device, Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 3 As shown, in one embodiment of the present invention, the electronic device 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; the memory is used to store computer programs; when the processor executes the computer programs stored in the memory, it implements the satellite mission planning method as described in any of the above embodiments of the present invention.

[0084] 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 illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0085] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0086] 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.

[0087] 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.

[0088] 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.

[0089] It should be noted that the specific implementation of the electronic device in the embodiments of the present invention is similar to the specific implementation of the satellite mission planning method described in the above embodiments of the present invention, and therefore has similar technical effects. For details, please refer to the description in the satellite mission planning method section. In order to reduce redundancy, it will not be repeated here.

[0090] Further embodiments of the present invention disclose a computer-readable storage medium storing a computer program that, when executed by a processor, implements the satellite mission planning method as described in any of the above embodiments of the present invention.

[0091] It should be noted that the specific implementation of the computer-readable storage medium in the embodiments of the present invention is similar to the specific implementation of the satellite mission planning method described in the above embodiments of the present invention, and therefore has similar technical effects. For details, please refer to the description in the satellite mission planning method section. To reduce redundancy, it will not be repeated here.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are 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.

[0093] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A satellite mission planning method, characterized in that, Includes the following steps: A satellite knowledge twin is constructed, which is used to uniformly model tasks, resources and constraints. The satellite knowledge twin includes physical data, domain knowledge and constraint rules, and is dynamically updated according to the real-time collected satellite execution status. A multi-agent collaborative architecture consisting of multiple agents is deployed. Each agent loads a domain-fine-tuned large language model to achieve autonomous decision-making and conducts distributed collaboration through a negotiation protocol to parse the input natural language task requirements, retrieve historical knowledge, and generate an initial execution strategy. The domain fine-tuning is performed using a space mission planning corpus through low-rank adaptation. The negotiation protocol is a contract network protocol. The multiple agents communicate with each other through a publish-subscribe mechanism. The contract network protocol is used to enable the multiple agents to negotiate and resolve task conflicts. Based on the initial execution strategy, a coupled model is established to define the matching relationship and matching conditions between the subtask slider and the time window. According to the matching relationship and the matching conditions, a mathematical optimization model containing decision variables, objective functions and constraints is constructed. The non-convex constraints in the mathematical optimization model are converted into linear constraints to form a linear mixed integer programming model. The initial execution strategy is transformed into the initial solution of the linear mixed integer programming model, and the initial solution is input into the solver for hot-start solution. The solution output by the solver is submitted to the formal verifier for physical constraint compliance verification. If the verification passes, the optimal task orchestration scheme is output. During the execution of the optimal task scheduling scheme, the satellite execution status is collected in real time, and the satellite knowledge twin is dynamically updated according to the satellite execution status to trigger the multi-agent collaborative architecture to re-make collaborative decisions and realize closed-loop optimization of the planning scheme. The construction of the satellite knowledge twin includes: Construct the satellite knowledge twin, comprising a physical twin layer, a knowledge twin layer, and a constraint twin layer; wherein, The physical twin layer is used to collect and store real-time physical data of the satellite; The knowledge twin layer is used to define core ontology classes and their semantic relationships of satellites, missions, resources and constraints based on the physical data using a network ontology language and store them in a graph database to construct a semantic knowledge graph. The semantic knowledge graph is used to define the semantic relationships of satellites, missions, resources and constraints, and implements symbolic reasoning through a rule engine and neural reasoning through a graph neural network. The constraint twin layer is used to define basic constraints and proprietary constraints based on the physical data and the semantic knowledge graph, and to realize real-time verification and conflict detection of the basic constraints and proprietary constraints through a rule engine. The process of transforming the initial execution strategy into an initial solution of the linear mixed-integer programming model, inputting the initial solution into the solver for warm-start solving, and submitting the solution output by the solver to the formal verifier for physical constraint compliance verification includes: The initial solution is generated by the knowledge reasoning agent in the multi-agent collaborative architecture, and the initial solution is used as the search starting point of the solver; If the solver reports no feasible solution, a subset of conflicting constraints is extracted and fed back to the knowledge reasoning agent to trigger the agent to adjust its initial execution strategy and regenerate a new initial solution for iterative solving; or, If the solver outputs the solution result, but the formal verifier fails to verify the compliance of physical constraints, the subset of conflicting constraints returned by the formal verifier is fed back to the knowledge reasoning agent to trigger iterative adjustment.

2. The satellite mission planning method according to claim 1, characterized in that, The deployment of a multi-agent collaborative architecture consisting of multiple agents includes: Deploy the multi-agent collaborative architecture consisting of a task parsing agent, a knowledge reasoning agent, a cooperative scheduling agent, and a solution execution agent; wherein... The task parsing agent is used to receive the natural language task requirements and decompose the natural language task requirements into structured atomic subtasks; The knowledge reasoning agent is used to retrieve historical similar tasks through a retrieval-enhanced generation mechanism and generate the initial execution strategy based on the retrieval results; The collaborative scheduling agent is used to coordinate resources from multiple satellites and ground stations through the negotiation protocol to resolve task conflicts and determine task and resource allocation schemes. The solution execution agent is used to perform modeling and solving of the mathematical optimization model according to the task and resource allocation scheme.

3. The satellite mission planning method according to claim 1, characterized in that, The establishment of the coupled model for defining the matching relationship and matching conditions between the subtask slider and the time window includes: A coupled model is established, comprising a subtask slider model, a time window model, and mapping relationships; wherein, The subtask slider model is used to decompose a task into multiple subtask sliders, and each subtask slider defines its execution duration, executable time range, and required resource attributes. The time window model is used to model the visible arc between the satellite and the ground station as a time window with start and end times, tracking mode and elevation angle information; The mapping relationship is used to establish matching decision variables between the subtask slider and the time window, and the matching decision variables are used to characterize the allocation relationship between tasks and resources.

4. The satellite mission planning method according to claim 1, characterized in that, The conversion of non-convex constraints in the mathematical optimization model into linear constraints includes at least one of the following methods: The Big-M method is adopted, by introducing a constant. M Transform the non-convex constraint containing the product term into a linear inequality; or, The auxiliary variable method is used to transform the non-convex constraint containing the L0 norm into the linear constraint by introducing a binary auxiliary variable; or, The absolute value function transformation method is adopted, which introduces auxiliary variables and binary variables to transform the non-convex constraint containing the absolute value function into a system of linear inequalities.

5. The satellite mission planning method according to claim 1, characterized in that, After constructing a mathematical optimization model that includes decision variables, an objective function, and constraints, and converting the non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed-integer programming model, the method further includes: Based on the number of tasks, the tasks are divided into small-scale, medium-scale, and large-scale problems, and a hierarchical solution strategy is selected according to the number of tasks; wherein, For the small-scale problem, the solver can be directly invoked for an accurate solution. For the medium-scale problem, a column generation and Benders decomposition strategy is used to solve it; For the large-scale problem, an adaptive large neighborhood search heuristic is used to generate a feasible solution, and the feasible solution is used as the initial solution to input into the solver for hot-start solution.

6. A satellite mission planning system, characterized in that, include: A knowledge twin server is used to construct a satellite knowledge twin, which is used to uniformly model tasks, resources and constraints. The satellite knowledge twin includes physical data, domain knowledge and constraint rules, and is dynamically updated according to the real-time collected satellite execution status. A multi-agent collaborative server is deployed with a multi-agent collaborative architecture consisting of multiple agents. Each agent loads a domain-fine-tuned large language model to achieve autonomous decision-making and conducts distributed collaboration through a negotiation protocol. It is used to parse the input natural language task requirements, retrieve historical knowledge, and generate an initial execution strategy. The domain fine-tuning is performed using a space mission planning corpus through low-rank adaptation. The negotiation protocol is a contract network protocol. The multiple agents communicate with each other through a publish-subscribe mechanism. The contract network protocol is used to enable the multiple agents to negotiate and resolve task conflicts. The solver server is configured to receive the initial execution strategy, establish a coupled model defining the matching relationship and conditions between subtask sliders and time windows, construct a mathematical optimization model including decision variables, objective functions, and constraints based on the matching relationship and conditions, convert non-convex constraints in the mathematical optimization model into linear constraints to form a linear mixed integer programming model, and convert the initial execution strategy into an initial solution of the linear mixed integer programming model. The initial solution is then input into the solver for hot-start solving, and the solution output by the solver is submitted to a formal verifier for physical constraint compliance verification. If the verification passes, the optimal task orchestration scheme is output. The communication terminal is used to collect the satellite execution status in real time during the execution of the optimal task scheduling scheme, and dynamically update the satellite knowledge twin according to the satellite execution status, so as to trigger the multi-agent collaborative architecture to re-make collaborative decisions and realize the closed-loop optimization of the planning scheme. The knowledge twin server constructs a satellite knowledge twin, including: Construct the satellite knowledge twin, comprising a physical twin layer, a knowledge twin layer, and a constraint twin layer; wherein, The physical twin layer is used to collect and store real-time physical data of the satellite; The knowledge twin layer is used to define core ontology classes and their semantic relationships of satellites, missions, resources and constraints based on the physical data using a network ontology language and store them in a graph database to construct a semantic knowledge graph. The semantic knowledge graph is used to define the semantic relationships of satellites, missions, resources and constraints, and implements symbolic reasoning through a rule engine and neural reasoning through a graph neural network. The constraint twin layer is used to define basic constraints and proprietary constraints based on the physical data and the semantic knowledge graph, and to realize real-time verification and conflict detection of the basic constraints and proprietary constraints through a rule engine. The solver server transforms the initial execution strategy into an initial solution for the linear mixed-integer programming model, inputs the initial solution into the solver for warm-start solving, and submits the solution output by the solver to the formal verifier for physical constraint compliance verification, including: The initial solution is generated by the knowledge reasoning agent in the multi-agent collaborative architecture, and the initial solution is used as the search starting point of the solver; If the solver reports no feasible solution, a subset of conflicting constraints is extracted and fed back to the knowledge reasoning agent to trigger the agent to adjust its initial execution strategy and regenerate a new initial solution for iterative solving; or, If the solver outputs the solution result, but the formal verifier fails to verify the compliance of physical constraints, the subset of conflicting constraints returned by the formal verifier is fed back to the knowledge reasoning agent to trigger iterative adjustment.

7. 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; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the satellite mission planning method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the satellite mission planning method according to any one of claims 1-5.

Citation Information

Patent Citations

  • Satellite constellation in-orbit cooperative task planning system based on digital twinning

    CN121822864A

  • Satellite task global optimization scheduling method and device

    CN122198416A