Satellite intelligent orbit transfer decision system and method

Abstract

The application belongs to the technical field of spacecraft orbit intelligent control, and discloses a satellite intelligent orbit transfer decision system and method, which comprises a perception layer, a service layer, a decision layer and an execution layer; by encapsulating multiple traditional orbit transfer algorithms into standardized MCP services, combining collision warning services, situation awareness services and satellite proximity monitoring services, using large model reasoning decision, adopting a LangChain framework to realize multi-service collaborative calling, and combining MCP control optimization, the satellite can autonomously generate an optimal avoidance strategy when approaching a target satellite. The application can effectively avoid the target satellite, avoid collision with other space objects, and minimize fuel consumption. The application significantly improves the autonomous survival ability and task execution efficiency of the satellite in a complex space environment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for spacecraft orbits, and relates to a satellite intelligent orbit change decision system and method. Background Technology

[0002] With the development of aerospace technology, the number of low-Earth orbit satellites has increased dramatically, the space environment has become increasingly complex, and the risk of collisions between satellites has continued to rise. At the same time, the space security threats faced by satellites are becoming increasingly diverse, including malicious reconnaissance and close-in observation.

[0003] Traditional satellite orbit change technology mainly relies on manual decision-making at ground control centers, which suffers from problems such as large response delays and limited ability to handle complex situations. Existing orbit change algorithms, such as the Lambert algorithm and the Hohmann orbit change algorithm, can achieve orbit transfer between two points, but lack comprehensive consideration of complex environments and multi-objective constraints.

[0004] In recent years, the application of artificial intelligence technology in the aerospace field has gradually increased, but existing technologies are mostly focused on single-task optimization and have failed to achieve intelligent collaboration of multiple tasks. Summary of the Invention

[0005] The technical problem solved by this invention is to provide a satellite intelligent orbit change decision system and method, which is based on the fusion of large model and multiple orbit change algorithms, and can integrate multiple constraints to autonomously generate the optimal orbit change strategy.

[0006] This invention is achieved through the following technical solution: A satellite intelligent orbit change decision-making system, comprising: The perception layer includes a satellite status monitoring module, a space environment perception module, and a target satellite monitoring module, which respectively acquire the satellite's own status, surrounding space environment information, and target satellite motion parameters; and send the acquired information to the service layer and the decision layer respectively. The service layer includes a trajectory change algorithm service module, a collision warning service module, and a situation assessment service module encapsulated based on the MCP protocol. When called by the decision-making layer as a toolset, these modules provide trajectory change scheme services, collision risk assessment services, and current global situation assessment services, respectively. The decision-making layer comprises a large-scale model inference module, a large-scale model decision-making module, a LangChain service scheduling module, and an MCP controller. The large-scale model inference module receives information from the perception layer, combines it with the satellite's status and the space environment to initiate a decision-making process, determining whether a trajectory change operation is necessary. If a trajectory change is required, it generates service layer tool invocation instructions and uses the LangChain service scheduling module to call various services within the service layer to generate multiple candidate trajectory change schemes. The large-scale model decision-making module comprehensively considers avoidance effectiveness, collision risk, and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision, calls various services within the service layer, generates executable control instructions, and sends them to the execution layer. The execution layer, including the orbit control module and the propulsion system control module, executes the orbit change control commands generated by the decision-making layer.

[0007] Furthermore, the satellite status monitoring module collects satellite orbital parameters and satellite health status parameters; the satellite orbital parameters include semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, and true anomaly; the satellite health status parameters include fuel remaining, thruster operating status, payload status, and the satellite's real-time position and velocity parameters. The space environment perception module collects space debris early warning data and obtains the probability of debris distribution near the current orbital region; The target satellite monitoring module collects the relative distance and relative velocity between the local satellite and the target satellite, and compares the deviation rate between the target satellite and the predicted trajectory in real time, and records whether the target satellite has any abnormal movement.

[0008] Furthermore, the orbit change algorithm service module encapsulates various orbit change algorithms, including the Lambert algorithm, the multipulse orbit change algorithm, and the Hohmann transfer orbit change algorithm, and defines the names and functional descriptions of the orbit change algorithms through the MCP standard protocol; If the orbit change algorithm service module encapsulates the multi-pulse orbit change algorithm, it will also integrate a multi-pulse optimizer subunit, which will convert the theoretical orbit transfer velocity increment output by the multi-pulse orbit change algorithm into a discrete pulse execution sequence that meets the hardware constraints of the satellite propulsion system and optimizes fuel consumption. The collision warning service module calculates the approach distance, approach time, and collision probability between the two objects based on the current state of the local satellite and the target space object, combined with the HPOP forecast of 24-72 hour ephemeris, in order to assess the collision risk level. The situation assessment service module receives information from the satellite status monitoring module and the space environment perception module, and receives the preliminary estimated fuel consumption for orbit change provided by the orbit change algorithm service module. It then assesses the current global situation and determines the threat level.

[0009] Furthermore, the large model reasoning module and the large model decision module are based on the same open-source base model with reasoning capabilities, and are trained by sharing and utilizing the professional knowledge base and simulation dataset in the aerospace field for instruction fine-tuning and reinforcement learning. The instruction fine-tuning is a supervised fine-tuning of task training based on the instruction-output integration of space real-world scenarios and reasoning and decision-making. The input includes satellite situation information and task requirements, and the output includes reasoning conclusions on whether to change orbit, tool call sequences, multi-scheme scoring logic and optimal decision results. The loss function is modeled using standard causal language, so that the training model can simultaneously master situation reasoning logic and multi-objective decision preferences. The reinforcement learning described herein is based on a reward model and uses the PPO algorithm to perform reinforcement learning on the trained model. The reward function considers both the accuracy of reasoning and the optimality of decision-making, so that the decision preference evolves from correct to optimal, maximizing the cumulative reward obtained from the reward model.

[0010] Furthermore, during system initialization, the LangChain service scheduling module traverses all services encapsulated based on the MCP protocol in the service layer and packages them one by one into tools in the LangChain environment. When the large model inference module determines the service layer tool call instruction, it sends a tool call request to the LangChain service scheduling module, described in unstructured natural language or semi-structured logic. The LangChain service scheduling module breaks down the requirement and identifies its intent. It then arranges the calling order and parameter passing path of the orchestration tools according to the logic of the requirement. Next, it calls the corresponding service layer tools and organizes their returned results into structured text or JSON data, generates context information, and feeds it back to the large model inference module for the next step of inference.

[0011] Furthermore, the MCP controller receives the optimal trajectory change decision intention described in natural language or structured JSON format from the large model decision module. The MCP controller parses the intention and calls the corresponding service layer module or execution interface. The MCP controller converts the numerical results of the service layer module or execution interface into standardized control instructions that the execution layer can recognize and process.

[0012] Furthermore, the track control module receives instruction packets from the decision-making layer and performs coordinate transformation and instruction sequence generation; The propulsion system control module receives the instruction sequence sent by the track control module and converts it into underlying drive signals for open-loop control and status monitoring.

[0013] Furthermore, it also includes a system state manager module, which synchronously updates the satellite orbital dynamics model and the propulsion system model, including thruster efficiency and fuel consumption characteristics, based on the real-time data received from the sensing layer. This enables a virtual mapping between the satellite's own state and the space environment, providing accurate model support for large-scale model decision-making. It also stores the satellite's core static parameters, including the satellite's factory-fixed parameters and thruster configuration parameters; It also has a built-in orbit simulation engine that provides a predictive simulation environment, which can load candidate orbit change schemes generated by large models to simulate orbit changes, fuel consumption and collision risks in the next 24 to 72 hours.

[0014] This invention also provides a satellite intelligent orbit change decision-making method, including the following operations: The perception layer collects real-time information on the satellite's own status, the surrounding space environment, and the target satellite's motion parameters; and sends the acquired information to the service layer and the decision layer to provide real-time and comprehensive situational awareness for intelligent decision-making. The service layer provides orbit change scheme services, collision risk assessment services, and current global situation assessment services based on the MCP protocol, providing a large set of tools that can be called upon for intelligent decision-making; The decision-making layer includes a large model inference module, a large model decision module, a LangChain service scheduling module, and an MCP controller. The large model inference module receives information sent by the perception layer, combines the satellite's status and the space environment to initiate a decision-making process. When a threat target is detected approaching or an orbital anomaly is detected, it constructs a decision prompt containing the current status, threat information, and environmental constraints. It calls the situation assessment service of the service layer to complete the global situation analysis, calls the collision warning service to obtain the target satellite's approach distance and collision probability, and combines the two results to determine whether there is a threat of the target satellite approaching and whether an orbital change operation is needed. When a trajectory change is required, a service layer tool call command is generated, and the LangChain service scheduling module calls various services in the service layer to generate multiple candidate trajectory change schemes. The large model decision module comprehensively considers avoidance effect, collision risk and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision and calls various services in the service layer to perform pulse sequence optimization, attitude coordination and fuel budget verification, and generates executable control commands to send to the execution layer. The execution layer executes the orbit change command, and the orbit adjustment is completed through the orbit control module and the propulsion system control module.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: The satellite intelligent orbit change decision system and method provided by this invention combines a large model with a variety of traditional orbit change algorithms, and encapsulates multiple orbit change algorithms, allowing the selection of the most suitable algorithm or combination of them according to specific circumstances; it utilizes the reasoning ability of the large model, and its processing power far exceeds that of traditional algorithms, realizing intelligent comprehensive decision-making in complex environments; moreover, it achieves real-time response through the decision layer, without relying on the ground control center, significantly shortening the decision cycle; The satellite intelligent orbit change decision system and method provided by this invention adopts a service-oriented architecture and the LangChain framework, which realizes flexible collaboration of multiple algorithms and services. The decision layer uses the MCP framework to encapsulate traditional orbit algorithms such as Lambert transfer, Hohmann orbit change and multipulse orbit change as standard services, while providing special services such as collision warning and situation assessment. Moreover, through service-oriented encapsulation, it is easy to add new orbit change algorithms and sensing capabilities. LangChain enables the service encapsulation and flexible invocation of various functional modules: The LangChain framework traverses all services encapsulated through the MCP protocol in the service layer and packages them one by one into "tools" in the LangChain environment. When the large model inference module needs to solve a complex problem, it autonomously analyzes the tool matching and decides which tools to call to complete the task. The LangChain service scheduling module is responsible for executing these call instructions, organizing the tool's return results into a new context, and feeding it back to the large model for the next step of inference, thus forming a closed-loop intelligent decision-making process of perception-planning-action.

[0016] The intelligent orbit change decision system and method for satellites provided by this invention achieves multi-objective optimization. While avoiding specific targets, it also takes into account collision avoidance and fuel consumption minimization. It can avoid both specific target satellites and other space objects. It minimizes fuel consumption while ensuring safety, thereby extending the satellite's on-orbit lifespan. It realizes intelligent orbit change decision-making for satellites in complex space environments, significantly improving the satellite's autonomous survivability and mission execution efficiency.

[0017] The satellite intelligent orbit change decision system and method provided by this invention introduces the MCP control architecture to solve the connection problem between decision-making and execution, ensuring the executability of the strategy. The execution layer transforms the decision scheme into executable instructions through the MCP controller, performs pulse sequence optimization, attitude coordination and fuel budget verification to ensure that the strategy scheme meets the engineering implementation constraints, and finally drives the orbit control module and propulsion system control module to complete precise orbit maneuvers.

[0018] The intelligent satellite orbit change decision system and method provided by this invention enable satellites to autonomously generate optimal avoidance strategies when facing reconnaissance or approach from other satellites. This effectively avoids target satellites while preventing collisions with other space objects and minimizing fuel consumption. This invention significantly improves the autonomous survivability and mission execution efficiency of satellites in complex space environments. It can be widely applied to various types of satellites that require autonomous orbit control, and has significant application value, especially in the fields of space security and satellite protection. Attached Figure Description

[0019] Figure 1 is an architecture diagram of the satellite intelligent orbit change decision system of the present invention; Figure 2 is a flowchart of the satellite intelligent orbit change decision method of the present invention; Figure 3 is a schematic diagram of the satellite avoidance scenario of the present invention; Figure 4 is a schematic diagram of the service invocation and decision-making process of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to embodiments, which are intended to explain the invention and not to limit it.

[0021] See Figure 1 A satellite intelligent orbit change decision-making system, comprising: The perception layer includes a satellite status monitoring module, a space environment perception module, and a target satellite monitoring module, which respectively acquire the satellite's own status, surrounding space environment information, and target satellite motion parameters; and send the acquired information to the service layer and the decision layer respectively. The service layer includes a trajectory change algorithm service module, a collision warning service module, and a situation assessment service module encapsulated based on the MCP protocol. When called by the decision-making layer as a toolset, these modules provide trajectory change scheme services, collision risk assessment services, and current global situation assessment services, respectively. The decision-making layer comprises a large-scale model inference module, a large-scale model decision-making module, a LangChain service scheduling module, and an MCP controller. The large-scale model inference module receives information from the perception layer, combines it with the satellite's status and the space environment to initiate a decision-making process, determining whether a trajectory change operation is necessary. If a trajectory change is required, it generates service layer tool invocation instructions and uses the LangChain service scheduling module to call various services within the service layer to generate multiple candidate trajectory change schemes. The large-scale model decision-making module comprehensively considers avoidance effectiveness, collision risk, and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision, calls various services within the service layer, generates executable control instructions, and sends them to the execution layer. The execution layer, including the orbit control module and the propulsion system control module, executes the orbit change control commands generated by the decision-making layer.

[0022] The following is a detailed explanation of each functional layer and module.

[0023] 1. Perception layer The perception layer acquires information about the satellite's own status, the surrounding space environment, and the motion parameters of the target satellite; it includes a satellite status monitoring module, a space environment perception module, and a target satellite monitoring module. Specific data or parameters of the perception layer include, for example: Orbital parameters: semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, true anomaly, and other instantaneous orbital elements.

[0024] Attitude parameters: three-axis attitude angles and angular velocities, pointing accuracy.

[0025] Health status parameters: remaining fuel on the satellite, operational status of each thruster, payload status, etc.

[0026] Position and velocity vectors: Position vector (x, y, z) and velocity vector (vx, vy, vz) under J2000.

[0027] Furthermore, the satellite status monitoring module collects satellite orbital parameters and satellite health status parameters. The satellite orbital parameters include instantaneous orbital elements such as semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, and true anomaly. The satellite health status parameters include fuel remaining, thruster operating status, payload status, and real-time position and velocity parameters of the satellite. The space environment perception module collects space debris early warning data and obtains the probability of debris distribution near the current orbital area; The target satellite monitoring module mainly collects the relative distance and relative velocity between the two satellites, as well as compares the deviation rate between the target satellite and the predicted trajectory in real time, and records whether the target satellite moves abnormally. All three types of data are sent to the service layer and decision-making layer in real time. After receiving the information, the large model inference module will combine the satellite's status and space environment to quickly initiate a new round of decision-making process to determine whether and how to perform the corresponding orbit change operation.

[0028] 2. Service Layer The service layer encapsulates various professional algorithms and capabilities into standard "tools" that can be understood and invoked by large models through the MCP framework; the service layer receives full data from the perception layer; it includes a trajectory change algorithm service module, a collision warning service module, and a situation assessment service module.

[0029] The orbit change algorithm service module encapsulates various orbit change algorithms, including the Lambert algorithm, multipulse orbit change algorithm, and Hohmann transfer orbit change algorithm. It defines clear names and functional descriptions for the orbit change algorithms through the MCP standard protocol, and the strict input and output parameter list allows large model (LLM) to understand and call them. If the orbit change algorithm service module encapsulates the multi-pulse orbit change algorithm, it will also integrate a multi-pulse optimizer subunit, which will convert the theoretical orbit transfer velocity increment output by the multi-pulse orbit change algorithm into a discrete pulse execution sequence that meets the hardware constraints of the satellite propulsion system and optimizes fuel consumption. For example, the multi-pulse optimizer subunit receives the theoretical trajectory parameters (total ΔV vector, initial value of pulse count, theoretical time of each pulse) output by the multi-pulse orbit change algorithm, as well as engineering constraint parameters (rated thrust of the thruster, minimum ignition time, fuel remaining, attitude maneuver path constraints, solar panel illumination window). Through pulse sequence discretization optimization and pulse time and attitude co-optimization, it generates a standardized pulse execution scheme containing pulse number, thruster ID, ignition start time, ignition duration, thrust direction vector, single-pulse fuel consumption, and total fuel consumption.

[0030] The collision warning service module assesses the risk level, such as "low", "medium", "high" or "urgent", based on the current status of the satellite and the target space object (such as debris or other satellites), combined with the HPOP forecast of 24-72 hour ephemeris, and calculates the minimum distance (approach distance), approach time and collision probability between the two objects using a distance formula. The situation assessment service module receives results from the satellite status monitoring module (fuel, health), the space environment perception module (environmental constraints), and the orbit change algorithm service module (preliminary estimated fuel consumption). When the inference module of the Large Model (LLM) receives a complex task (such as "performing an on-orbit rendezvous") or a sudden event (such as "collision warning"), it will proactively invoke this service to assess the current global situation and determine the threat level.

[0031] 3. Decision-making level The decision-making layer includes a large model inference module, a large model decision-making module, a LangChain service scheduling module, and an MCP controller. It receives data from the perception layer, calls various services of the service layer, obtains external expert advice, generates optimal course-changing decisions, and controls whether to call the execution layer to upload instructions.

[0032] The large-scale model inference module and the large-scale model decision-making module are built on the same open-source base model with complex reasoning capabilities. They share a professional knowledge base and simulation dataset in the aerospace field and are trained through instruction fine-tuning and reinforcement learning, eliminating the need for separate training processes for the two modules. Each module performs its own function, working together to complete the intelligent decision-making process. The large model inference module is responsible for situation assessment and mission planning. After receiving information from the perception layer, it combines the satellite's status and space environment to construct decision prompts and initiate a decision process to determine whether an orbit change operation needs to be performed. At the same time, it generates tool call logic and calls various services of the service layer through the LangChain service scheduling module to generate multiple candidate orbit change schemes. The large model decision module is responsible for multiple scheme selection and decision generation. It receives multiple candidate trajectory change schemes and collision risk and fuel consumption assessment data of each scheme from the large model inference module, and generates the optimal trajectory change decision by comprehensively considering the avoidance effect, collision risk and fuel consumption.

[0033] The MCP controller receives the optimal trajectory change decision intent described in natural language or structured JSON format from the large model decision module. The MCP controller parses the intent and calls the corresponding service layer module or execution interface. The MCP controller converts the numerical results of the service layer module or execution interface into standardized control instructions that the execution layer can recognize and process.

[0034] The LangChain service scheduling module traverses all services encapsulated by the MCP protocol in the service layer and packages them one by one into tools in the LangChain environment; When the large model inference module determines the service layer tool call instructions, the LangChain service scheduling module executes these call instructions, calls the corresponding service layer toolsets, organizes the returned results into context, and feeds it back to the large model inference module for the next step of inference, thus forming a closed-loop intelligent decision-making process of perception-planning-action.

[0035] The large model inference module and the large model decision module do not directly use a general large model. Instead, they are based on an advanced open-source base model and are trained by fine-tuning instructions and reinforcement learning using a built aerospace professional knowledge base and simulation data.

[0036] The instruction fine-tuning is a supervised fine-tuning of task training based on the instruction-output integration of space real-world scenarios and reasoning and decision-making. The input includes satellite situation information and task requirements, and the output includes reasoning conclusions on whether to change orbit, tool call sequences, multi-scheme scoring logic and optimal decision results. The loss function is modeled using standard causal language, so that the training model can simultaneously master situation reasoning logic and multi-objective decision preferences. The reinforcement learning described herein is based on a reward model and uses the PPO algorithm to perform reinforcement learning on the trained model. The reward function considers both the accuracy of reasoning and the optimality of decision-making, so that the decision preference evolves from correct to optimal, maximizing the cumulative reward obtained from the reward model.

[0037] The following is an example of training the large model inference module and the large model decision module.

[0038] Base model: An open-source architecture with powerful inference capabilities, such as ChatGLM, is selected as the foundation.

[0039] Training process: Supervisory fine-tuning: First, it is trained using manually written "instruction-output" sequences. Data simulates various real-world space scenarios (such as orbital transfer, collision avoidance, and fault response), and experts write the correct decision-making processes and tool call sequences. For example, given the input "Approaching debris detected, probability 10^-4, current fuel 50kg, please make a decision," the output should be "Call the collision warning service to confirm the risk, call the situation assessment service to analyze avoidance solutions...".

[0040] Reward Model and Reinforcement Learning: Subsequently, reinforcement learning is introduced to further align model behavior. A "reward model" is trained, which scores the model's output based on multiple metrics such as fuel efficiency, task success rate, and safety. Based on this reward model, the large model is optimized using reinforcement learning through the PPO algorithm, causing its decision preferences to evolve from "correct" to "optimal".

[0041] Loss Function: During the supervised fine-tuning phase, a standard causal language is primarily used to model the loss function, aiming to ensure the model accurately predicts the next token. The core of this is the cross-entropy loss. The formula is shown below: ; This is the set of token positions corresponding to the output sequence, excluding the token range of the input prompt words; Given the preceding sequence, the model predicts that the t-th token is the true label. The probability of; These are model parameters; The purpose of this loss function is to enable the model to "generate decision outputs that conform to expert logic based on the input satellite situation and mission instructions", laying the foundation for the subsequent reinforcement learning stage.

[0042] The reinforcement learning phase aims to ensure that the model learns to follow the provided domain instructions and reasoning formats, while the reinforcement learning phase aims to maximize the cumulative rewards obtained from the reward model.

[0043] The following is an example of complex decision-making reasoning: Suppose the sensing layer reports that "the target satellite performed an unexpected orbital lift maneuver".

[0044] Contextual understanding and goal setting: The large model first interprets the information: "The target has moved, and our original rendezvous mission plan has failed. The new goal is to replan the route and achieve the rendezvous under fuel constraints." Tool Invocation and Information Integration: The large model inference module automatically invokes relevant service layers. Use the "Target Satellite Monitoring" function to obtain the target's new orbital parameters.

[0045] Call the "Lambert orbit change algorithm service", input the current positions of the local satellite and the target, and calculate a series of orbit schemes with different transfer times.

[0046] The "Situation Assessment Service" was invoked to assess the fuel consumption of each Lambert program and compare it with the remaining fuel on the local satellite.

[0047] Reasoning and Trade-offs: After receiving the data for each option, the large model decision module performs logical reasoning: "Option A has the shortest transfer time but consumes 50kg of fuel, while Option B takes longer but only consumes 20kg. Currently, there is 25kg of fuel remaining, so Option A is not feasible. Option B has sufficient fuel, but it is necessary to assess whether the task time limit allows it." Final Decision Generation: After integrating all information, the large model decision module generates the final decision instruction: "Adopt the fuel-optimal Lambert transfer scheme B. The estimated ΔV is XV, and the total time is Y hours. The instruction has been sent to the execution layer for digital simulation verification."

[0048] LangChain Service Scheduling Module: As the intelligent collaboration hub between the large model and the service layer toolset, the core function of the LangChain Service Scheduling Module is to bridge the gap between the large model's "cognitive reasoning" and "tool invocation," enabling dynamic orchestration and result integration of multiple services. It receives tool invocation requests from the large model's reasoning module, described in unstructured natural language or semi-structured logic. It then breaks down and identifies the intent of the request, and matches it with standardized tools encapsulated based on the MCP protocol in the service layer. After completing the tool matching, the module automatically orchestrates the tool invocation order and parameter passing path according to the logical association of the requests.

[0049] For example, first call the situation assessment service to obtain the current global threat level, then use the threat level as a constraint to pass the trajectory change algorithm service to generate candidate solutions that meet the safety threshold, and finally call the collision warning service to perform risk verification on all candidate solutions.

[0050] Meanwhile, the LangChain service scheduling module normalizes the return results from each service: it converts heterogeneous data from different algorithms (such as orbital roots from the Lambert algorithm and pulse sequences from the multipulse algorithm) into structured text or JSON data that the large model can directly parse, and integrates these results into new contextual information, feeding it back to the large model inference module to support its next steps in scheme selection and decision-making inference. Ultimately, this module automates the entire process of "requirement decomposition - tool matching - collaborative invocation - result integration," allowing large models to complete complex multi-tool collaborative tasks without needing to focus on the implementation details of the underlying algorithms, simply through natural language descriptions. This enables the construction of an intelligent decision-making process of "perception - reasoning - invocation - re-reasoning."

[0051] MCP Controller: As the core hub connecting model inference and control execution, the MCP controller operates as follows: It receives high-level decision intentions from the large model inference module, described in natural language or structured JSON format (e.g., "execute a Hohmann maneuver with minimal fuel consumption"). The MCP controller then parses this intention and calls the corresponding underlying algorithm or execution interface, which is encapsulated as an MCP service (e.g., calling the "Hohmann maneuver algorithm service" to calculate the specific ΔV and ignition timing). Finally, the controller converts the precise numerical results returned by the algorithm into standardized control commands that the execution layer can recognize and process, thus completing the closed loop from intelligent decision-making to specific execution.

[0052] 4. Execution layer The execution layer is responsible for translating decisions into concrete actions and controlling the satellite propulsion system to perform orbit change operations. The execution layer includes an orbit control module and a propulsion system control module, which are used to execute the orbit change commands generated by the decision layer.

[0053] The track control module receives instruction packets from the decision-making layer, performs coordinate transformation and generates instruction sequences; the propulsion system control module receives instruction sequences sent by the track control module, converts them into underlying drive signals, and performs open-loop control and status monitoring.

[0054] The orbit control module receives an instruction packet from the decision layer, and the data packet contains key parameters such as: the target velocity increment vector ΔV = [ΔV_x, ΔV_y, ΔV_z] (in the body coordinate system), the maneuver start time t_0, and the pulse width sequence. The core task of this module is coordinate conversion and instruction sequence generation. Specifically, according to the current attitude matrix A_{body→inertial}, it decomposes the target ΔV into the requirements for the thrusters of each axis, and calculates the required ignition duration Δt of each relevant thruster based on the thruster installation matrix B and the rated thrust F_{rated}. This calculation follows the formula: m * |ΔV| ≈ F_{rated} * ΣΔt (where m is the current mass of the satellite), and finally generates an accurate time-sequence control instruction table containing the thruster ID, ignition start time and duration.

[0055] Subsequently, the propulsion system control module receives the instruction table and converts it into underlying drive signals. Its core tasks are open-loop control and state monitoring. An instruction queue is maintained inside the module. When the absolute time t = t_0, a high-voltage level signal with a duration of Δt is sent to the specified thruster solenoid valve, so as to precisely control the injection duration of the propellant. During the entire execution process, the module monitors key parameters in real time through the sensor network; for example, fuel pressure P, pre-valve temperature T of the thruster, and remaining fuel mass m_fuel. These measured values will be compared with the expected instruction values. If any parameter (such as P < P_{threshold}) exceeds the preset safety threshold, the safety mechanism will be triggered immediately, the current instruction stream will be aborted, and an exception code will be reported to the service layer, so as to ensure that the entire orbit transfer operation completes the orbit transfer task within a controlled range.

[0056] Further, the satellite intelligent orbit transfer decision system further comprises a system state manager module, which constructs a satellite digital twin, stores core static parameters, and provides a predictive simulation environment; The construction of the satellite digital twin is based on the real-time data of the perception layer, synchronously updating the satellite orbit dynamics model and the propulsion system model including thruster efficiency and fuel consumption characteristics, so as to realize the virtual mapping of the satellite's own state and the space environment, and provide accurate model support for decision-making; The storing of core static parameters refers to centrally storing basic data that does not change dynamically with time, such as solidified satellite factory parameters (mass, inertia matrix) and thruster configuration parameters (thruster ID, installation position, rated thrust, minimum ignition time), so as to avoid repeated storage by each module and improve system resource utilization; The aforementioned predictive simulation environment utilizes a built-in orbit simulation engine to load candidate orbit change schemes generated from large models. It simulates orbital changes, fuel consumption, and collision risks over the next 24 to 72 hours, pre-screening invalid schemes and improving the decision-making efficiency and accuracy of the decision-making level. This provides a pre-verification guarantee for the execution of the final orbit change command.

[0057] See Figures 2-4 The satellite intelligent orbit change decision method of the present invention includes the following operations: The perception layer collects real-time information on the satellite's own status, the surrounding space environment, and the target satellite's motion parameters; and sends the acquired information to the service layer and the decision layer to provide real-time and comprehensive situational awareness for intelligent decision-making. The service layer provides orbit change scheme services, collision risk assessment services, and current global situation assessment services based on the MCP protocol, providing a large set of tools that can be called upon for intelligent decision-making; The decision-making layer includes a large-scale model inference module, a large-scale model decision-making module, a LangChain service scheduling module, and an MCP controller. The large-scale model inference module receives information from the perception layer and initiates a decision-making process based on the satellite's status and the space environment. When a threat target is detected approaching or an orbital anomaly is detected, it constructs a decision prompt that includes the current status, threat information, and environmental constraints. It calls the situation assessment service of the service layer to complete global situation analysis and calls the collision warning service to obtain the target satellite's approach distance and collision probability. It combines the two results to determine whether there is a threat of a target satellite approaching and whether an orbital change operation is needed. When a trajectory change is required, a service layer tool call command is generated, and the LangChain service scheduling module calls various services in the service layer to generate multiple candidate trajectory change schemes. The large model decision module comprehensively considers avoidance effect, collision risk and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision and calls various services in the service layer to perform pulse sequence optimization, attitude coordination and fuel budget verification, and generates executable control commands to send to the execution layer. The execution layer executes the orbit change command, and the orbit adjustment is completed through the orbit control module and the propulsion system control module.

[0058] The following is a specific example. Example

[0059] A satellite intelligent orbit change decision-making method includes the following operations: 1. During system initialization, the system state manager module creates an internal model containing satellite state and space environment data, providing a predictive simulation environment; Load pre-trained large model weights and configure the service layer toolset (including MCP services such as trajectory change algorithm and collision warning).

[0060] 2. The sensing layer continuously updates the satellite's own status and space environment data (refreshing the status at a fixed period of 1Hz), including orbital information of other satellites and space debris.

[0061] 3. When the perception layer detects a threatening satellite (such as satellite A) approaching, it sends the information to the service layer and the decision layer; the collision warning module calculates the relative distance d between the two satellites and the collision probability P (corresponding to...). Figure 2 (Threat assessment in the system) The system initiates an intelligent decision-making process (such as...) Figure 2 (as shown) 4. Build decision prompts that include current status, threat information, and environmental data; Decision suggestions are as follows: Satellite status: Mass {mass} kg, Position {r} km, Velocity {v} km / s, Fuel {fuel} kg Threat information: Target A, collision probability {Pc}, closest time {t_ca} s, estimated minimum distance {d_min} km Environmental parameters: atmospheric density {ρ} kg / m³, J2 perturbation {J2} Task requirement: Generate an avoidance scheme under the constraint ΔV < {max_dv} m / s. The large model inference module calls the situation assessment service of the service layer to complete the global situation analysis, and calls the collision warning service to obtain the target satellite's approach distance and collision probability. It combines the two results to determine whether there is a threat of the target satellite approaching and whether it is necessary to perform an orbit change operation. 5. The large model inference module uses the LangChain framework to call various tools (orbit change algorithms, collision warnings, etc.) to generate multiple candidate orbit change schemes.

[0062] The specific toolchain used is: LangChain proxy → Lambert solver (generate transfer paths); LangChain agent → Through the orbit change algorithm service, it generates 3-5 different strategy schemes, including: immediate avoidance, delayed avoidance, orbital plane change and other different types; LangChain agent → Multi-pulse optimizer, generates pulse sequences; LangChain Agent → Collision risk assessment, verification of solution safety); 6. Conduct a collision risk assessment for each candidate solution: The collision risk assessment considers not only the target satellite, but also other known satellites and space debris.

[0063] The evaluation employs a three-dimensional judgment logic based on time window, minimum distance, and collision probability. It uses the orbital parameters of candidate satellites and employs the HPOP (High-Precision Orbit Prediction) model to predict the orbital trajectories of the local satellite and the target satellite over the next 24-72 hours, calculating the minimum approach distance between the two satellites within the predicted time window. And determine the corresponding proximity time. By combining the orbital error covariance matrices of the two stars, the collision probability during their rendezvous is calculated. .

[0064] The collision risk assessment results are organized into structured data by the LangChain service scheduling module and fed back to the large model decision module of the decision layer. As one of the core bases for multi-objective weighted scoring, it is ultimately used to select the optimal trajectory change scheme.

[0065] 7. The large model decision module comprehensively considers avoidance effect, collision risk and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller; The calculation methods for each scoring item are as follows: Evasion effect The evaluation score is: ; safe distance threshold The orbital spacing is set according to the size and orbital maneuverability of the target satellite, for example, 3 to 5 times the size of the satellite itself, or the orbital safety spacing commonly used in the aerospace field. Invalid distance threshold When the minimum proximity distance is greater than At that time, the evasion effect is determined to be optimal; Based on the orbital prediction data of candidate scenarios, the minimum approach distance between the local satellite and the target satellite within the next 24-72 hours is calculated using the collision warning service module. ; like The plan is ineffective in avoiding the target satellite, and scores 0. like The solution achieves complete avoidance, scoring 1 point. Middle interval: Score follows Increase linear improvement; Fuel consumption The evaluation score is: ; Base threshold The maximum fuel consumption allowed by the system is typically 90% of the satellite's remaining fuel, with a 10% safety margin. Optimal threshold The theoretical minimum fuel consumption is determined by the optimal solution of the orbit change algorithm, such as the lower limit of fuel consumption for the Hohmann orbit change. The multi-pulse optimizer subunit of the trajectory change algorithm service module calculates the total fuel consumption of candidate routes. ; like Fuel consumption exceeds the safety threshold, the solution is not feasible, score is 0; like Fuel consumption has reached the theoretical optimum, score is 1; The collision risk assessment score is based on the assessment results of the collision warning service module, assigning a risk level coefficient to each type of object. i=1,2,3 correspond to the target satellite, other satellites, and space debris, respectively, representing high risk. =0, medium risk =0.5, low risk =1; The overall assessment is as follows: ; w1, w2, and w3 are the weights for evasion effectiveness, fuel consumption, and collision risk, respectively. For example, the weights [w1,w2,w3] = [0.5, 0.3, 0.2] can be configured according to specific needs.

[0066] 8. The MCP controller analyzes the optimal trajectory change decision and calls various services in the service layer to perform pulse sequence optimization, attitude coordination and fuel budget verification, generate executable control instructions to make them more suitable for actual execution, and convert them into propulsion system instructions to be sent to the execution layer; The main optimizations of the MCP controller include: 1) Pulse sequence discretization: Decompose the theoretical ΔV into a pulse sequence that can be realized by a thruster; 2) Consider the minimum ignition time constraint of the thruster: ensure that the duration of each pulse Δt ≥ 20ms (typical value); 3) Plan the attitude maneuver path to ensure attitude alignment is completed before pulse execution; 4) Review the fuel budget and reserve 10% of the fuel as a safety margin; 5) Optimize pulse timing to avoid periods of solar panel shading and control the angle of sunlight; 6) Consider thermal control constraints to ensure that the thruster's operating time is within a safe range; The following provides a specific optimization example.

[0067] For example, the content that needs optimization in MCP: a) Pulse sequence smoothing, taking into account the thruster's minimum ignition time Δt_min; b) Coordinate attitude maneuvers to ensure accurate thrust direction; c) Review the fuel budget and retain a safety margin; Corresponding instruction conversion: A) Send the following to the track control module: {thruster_id, Δt, ignition_time, thrust_vector} B) Send the following to the attitude control module: {target_attitude, maneuver_time} C) Parameter example: THRUSTER_1, 2.5s, T+300s, [0,0.707,0.707] 9. The execution layer executes the track change command to complete the track adjustment.

[0068] This invention combines a large model with various traditional orbit-changing algorithms to achieve intelligent decision-making in complex environments; it achieves multi-objective optimization, minimizing fuel consumption while ensuring safety; it can be widely applied to various satellites that require autonomous orbit control, and has significant application value, especially in the fields of space security and satellite protection.

[0069] The embodiments given above are preferred examples for implementing the present invention, and the present invention is not limited to the above embodiments. Any non-essential additions or substitutions made by those skilled in the art based on the technical features of the present invention are within the protection scope of the present invention.

Claims

1. A satellite intelligent orbit change decision-making system, characterized in that, include: The perception layer includes a satellite status monitoring module, a space environment perception module, and a target satellite monitoring module, which respectively acquire the satellite's own status, surrounding space environment information, and target satellite motion parameters; and send the acquired information to the service layer and the decision layer respectively. The service layer includes a trajectory change algorithm service module, a collision warning service module, and a situation assessment service module encapsulated based on the MCP protocol. When called by the decision-making layer as a toolset, these modules provide trajectory change scheme services, collision risk assessment services, and current global situation assessment services, respectively. The decision-making layer comprises a large-scale model inference module, a large-scale model decision-making module, a LangChain service scheduling module, and an MCP controller. The large-scale model inference module receives information from the perception layer, combines it with the satellite's status and the space environment to initiate a decision-making process, determining whether a trajectory change operation is necessary. If a trajectory change is required, it generates service layer tool invocation instructions and uses the LangChain service scheduling module to call various services within the service layer to generate multiple candidate trajectory change schemes. The large-scale model decision-making module comprehensively considers avoidance effectiveness, collision risk, and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision, calls various services within the service layer, generates executable control instructions, and sends them to the execution layer. The execution layer, including the orbit control module and the propulsion system control module, executes the orbit change control commands generated by the decision-making layer; The evaluation of the evasion effect is as follows: in, To avoid the effect, D min To determine the minimum approach distance, based on the orbital prediction data of candidate scenarios, the collision warning service module calculates the minimum approach distance between the local satellite and the target satellite within the next 24-72 hours. D min ; This is the safe distance threshold; This is an invalid distance threshold; when the minimum proximity distance is greater than... At that time, the evasion effect is determined to be optimal; like The plan is ineffective in avoiding the target satellite, and scores 0. like The solution achieves complete avoidance, scoring 1 point. Middle interval: Score follows Increase linear improvement; The assessment of fuel consumption is as follows: For fuel consumption; The base threshold is the maximum fuel consumption allowed by the system. The optimal threshold is the theoretical minimum fuel consumption determined by the optimal solution of the orbit change algorithm. The total fuel consumption of the candidate scheme; like ≥ Fuel consumption exceeds the safety threshold, the solution is not feasible, score is 0; like ≤ Fuel consumption has reached the theoretical optimum, score is 1; The collision risk assessment is as follows Based on the evaluation results of the collision warning service module, a risk level coefficient is assigned to each type of object. i=1,2,3 correspond to the target satellite, other satellites, and space debris, respectively, with high risk being the most significant. =0, medium risk =0.5, low risk =1; The comprehensive assessment is as follows: Total score = w1 × + w2× + w3× w1, w2, and w3 are the weights for the avoidance effect assessment, fuel consumption assessment, and collision risk assessment, respectively.

2. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, The satellite status monitoring module collects satellite orbital parameters and satellite health status parameters; the satellite orbital parameters include semi-major axis, eccentricity, orbital inclination, right ascension of the ascending node, argument of perigee, and true anomaly; the satellite health status parameters include fuel remaining, thruster operating status, payload status, and real-time position and velocity parameters of the satellite. The space environment perception module collects space debris early warning data and obtains the probability of debris distribution near the current orbital region; The target satellite monitoring module collects the relative distance and relative velocity between the local satellite and the target satellite, and compares the deviation rate between the target satellite and the predicted trajectory in real time, and records whether the target satellite has any abnormal movement.

3. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, The orbit change algorithm service module encapsulates a variety of orbit change algorithms, including the Lambert algorithm, multipulse orbit change algorithm, and Hohmann transfer orbit change algorithm, and defines the names and functional descriptions of the orbit change algorithms through the MCP standard protocol; If the orbit change algorithm service module encapsulates the multi-pulse orbit change algorithm, it will also integrate a multi-pulse optimizer subunit, which will convert the theoretical orbit transfer velocity increment output by the multi-pulse orbit change algorithm into a discrete pulse execution sequence that meets the hardware constraints of the satellite propulsion system and optimizes fuel consumption. The collision warning service module calculates the approach distance, approach time, and collision probability between the two objects based on the current state of the local satellite and the target space object, combined with the HPOP forecast of 24-72 hour ephemeris, in order to assess the collision risk level. The situation assessment service module receives information from the satellite status monitoring module and the space environment perception module, and receives the preliminary estimated fuel consumption for orbit change provided by the orbit change algorithm service module. It then assesses the current global situation and determines the threat level.

4. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, The large model reasoning module and the large model decision module are based on a unified open-source base model with reasoning capabilities, and are trained by fine-tuning instructions and reinforcement learning using a shared aerospace professional knowledge base and simulation dataset. The instruction fine-tuning is a supervised fine-tuning of task training based on the instruction-output integration of space real-world scenarios and reasoning and decision-making. The input includes satellite situation information and task requirements, and the output includes reasoning conclusions on whether to change orbit, tool call sequences, multi-scheme scoring logic and optimal decision results. The loss function is modeled using standard causal language, so that the training model can simultaneously master situation reasoning logic and multi-objective decision preferences. The reinforcement learning described herein is based on a reward model and uses the PPO algorithm to perform reinforcement learning on the trained model. The reward function considers both the accuracy of reasoning and the optimality of decision-making, so that the decision preference evolves from correct to optimal, maximizing the cumulative reward obtained from the reward model.

5. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, During system initialization, the LangChain service scheduling module traverses all services encapsulated based on the MCP protocol in the service layer and packages them one by one into tools in the LangChain environment. When the large model inference module determines the service layer tool call instruction, it sends a tool call request to the LangChain service scheduling module, described in unstructured natural language or semi-structured logic. The LangChain service scheduling module breaks down the requirement and identifies its intent. It then arranges the calling order and parameter passing path of the orchestration tools according to the logic of the requirement. Next, it calls the corresponding service layer tools and organizes their returned results into structured text or JSON data, generates context information, and feeds it back to the large model inference module for the next step of inference.

6. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, The MCP controller receives the optimal trajectory change decision intent described in natural language or structured JSON format from the large model decision module. The MCP controller parses the intent and calls the corresponding service layer module or execution interface. The MCP controller converts the numerical results of the service layer module or execution interface into standardized control instructions that the execution layer can recognize and process.

7. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, The track control module receives instruction packets from the decision-making layer and performs coordinate transformation and instruction sequence generation. The propulsion system control module receives the instruction sequence sent by the track control module and converts it into underlying drive signals for open-loop control and status monitoring.

8. The satellite intelligent orbit change decision system as described in claim 1, characterized in that, It also includes a system state manager module, which synchronously updates the satellite orbital dynamics model and the propulsion system model, including thruster efficiency and fuel consumption characteristics, based on the real-time data received from the sensing layer. This enables a virtual mapping between the satellite's own state and the space environment, providing accurate model support for large-scale model decision-making. It also stores the satellite's core static parameters, including the satellite's factory-fixed parameters and thruster configuration parameters; It also has a built-in orbit simulation engine that provides a predictive simulation environment, which can load candidate orbit change schemes generated by large models to simulate orbit changes, fuel consumption and collision risks in the next 24 to 72 hours.

9. A satellite intelligent orbit change decision-making method, characterized in that, Includes the following operations: The perception layer collects real-time information on the satellite's own status, the surrounding space environment, and the target satellite's motion parameters; and sends the acquired information to the service layer and the decision layer to provide real-time and comprehensive situational awareness for intelligent decision-making. The service layer provides orbit change scheme services, collision risk assessment services, and current global situation assessment services based on the MCP protocol, providing a large set of tools that can be called upon for intelligent decision-making; The decision-making layer includes a large model inference module, a large model decision module, a LangChain service scheduling module, and an MCP controller. The large model inference module receives information sent by the perception layer, combines the satellite's status and the space environment to initiate the decision-making process, and when it senses an approaching threat target or orbital anomaly, it constructs a decision prompt that includes the current status, threat information, and environmental constraints. The system calls the situation assessment service of the service layer to complete the global situation analysis, calls the collision warning service to obtain the target satellite's approach distance and collision probability, and combines the two results to determine whether there is a threat of the target satellite approaching and whether it is necessary to perform an orbit change operation. When a trajectory change is required, a service layer tool call command is generated, and the LangChain service scheduling module calls various services in the service layer to generate multiple candidate trajectory change schemes. The large model decision module comprehensively considers avoidance effect, collision risk and fuel consumption to generate the optimal trajectory change decision and sends it to the MCP controller. The MCP controller parses the optimal trajectory change decision and calls various services in the service layer to perform pulse sequence optimization, attitude coordination and fuel budget verification, and generates executable control commands to send to the execution layer. The execution layer executes the orbit change command, and completes the orbit adjustment through the orbit control module and the propulsion system control module; The evaluation of the evasion effect is as follows: in, To avoid the effect, D min To determine the minimum approach distance, based on the orbital prediction data of candidate scenarios, the collision warning service module calculates the minimum approach distance between the local satellite and the target satellite within the next 24-72 hours. D min ; This is the safe distance threshold; This is an invalid distance threshold; when the minimum proximity distance is greater than... At that time, the evasion effect is determined to be optimal; like The plan is ineffective in avoiding the target satellite, and scores 0. like The solution achieves complete avoidance, scoring 1 point. Middle interval: Score follows Increase linear improvement; The assessment of fuel consumption is as follows: For fuel consumption; The base threshold is the maximum fuel consumption allowed by the system. The optimal threshold is the theoretical minimum fuel consumption determined by the optimal solution of the orbit change algorithm. The total fuel consumption of the candidate scheme; like ≥ Fuel consumption exceeds the safety threshold, the solution is not feasible, score is 0; like ≤ Fuel consumption has reached the theoretical optimum, score is 1; The collision risk assessment is as follows Based on the evaluation results of the collision warning service module, a risk level coefficient is assigned to each type of object. i=1,2,3 correspond to the target satellite, other satellites, and space debris, respectively, with high risk being the most significant. =0, medium risk =0.5, low risk =1; The comprehensive assessment is as follows: Total score = w1 × + w2× + w3× w1, w2, and w3 are the weights for the avoidance effect assessment, fuel consumption assessment, and collision risk assessment, respectively.

10. The satellite intelligent orbit change decision-making method as described in claim 9, characterized in that, The decision prompts include at least: Satellite status: including mass, position, velocity, and fuel; Threat information includes: target satellite, probability of collision, closest approach time, and estimated minimum distance; Environmental parameters: atmospheric density, J2 perturbation; Task requirement: Generate an evasion plan under constraints; The MCP controller generates control commands through the following optimization: 1) Pulse sequence discretization: Decompose the theoretical ΔV into a pulse sequence that can be realized by a thruster; 2) Considering the minimum ignition time constraint of the thruster: ensure that the duration Δt of each pulse is ≥ 20ms; 3) Plan the attitude maneuver path to ensure attitude alignment is completed before pulse execution; 4) Review the fuel budget and reserve 10% of the fuel as a safety margin; 5) Optimize pulse timing to avoid periods of solar panel shading and control the angle of sunlight; 6) Consider thermal control constraints to ensure that the thruster's operating time is within a safe range.

Images