Satellite-ground collaborative decision-making system and method fusing large language model and expert knowledge

Abstract

The application discloses a satellite-ground collaborative decision-making system and method fusing a large language model and expert knowledge, and relates to the technical field of intelligent management and control of aerospace satellites. The application comprises the following steps: deploying a gravity gradient sensor array and a star sensor combination on a satellite to form an attitude reference self-calibration mechanism; judging the real-time condition of the satellite body by the attitude reference self-calibration mechanism to generate uplink and downlink data on the satellite; constructing an expert Skill semantic space; generating a corresponding task semantic vector according to the uplink and downlink data on the satellite; judging whether the task semantic vector is a new type of task; generating a decision experience model according to the judgment result and the task semantic vector by the expert Skill semantic space; generating a preliminary decision scheme according to the decision experience model; generating a satellite-ground collaborative decision-making result by comparing the preliminary decision scheme with a preset expert Skill standard output; obtaining the execution result data of the satellite-ground collaborative decision-making; and modifying the parameters of the expert Skill semantic space according to the execution result data.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for aerospace satellites, specifically to a space-ground collaborative decision-making system and method that integrates large language models and expert knowledge. Background Technology

[0002] During their on-orbit operation, satellites need to continuously make decisions regarding attitude maintenance, orbit maintenance, collision avoidance, and mission adjustments based on their own status and mission requirements. These decisions are strictly constrained by remaining fuel and response time, demanding high real-time performance, reliability, and safety. Current satellite decision-making mainly relies on ground-based telemetry, tracking, and command (TT&C), where telemetry data is sent from the satellite, analyzed by a ground team, and then back up to control commands. This approach involves many links and long response times, making it difficult to complete timely decisions in time-sensitive scenarios such as orbital maneuvers and collision warnings. While some satellites possess a certain degree of onboard autonomy, their autonomous logic is often based on pre-defined rules or single models, covering only limited mission scenarios foreseen during the design phase. They lack generalization and reasoning capabilities when faced with new, undefined mission types.

[0003] In recent years, large language models have demonstrated strong generalization capabilities in semantic understanding and mission planning. However, directly applying them to satellite decision-making results in uncertain outputs, insufficient interpretability, and potential deviations from aerospace safety constraints, making it difficult to guarantee the standardization and reliability of decisions. Furthermore, existing methods struggle to organically combine historical mission experience, expert knowledge, and the reasoning capabilities of large language models, and lack a mechanism for continuously revising decision knowledge based on actual execution results, potentially leading to repeated occurrences of the same biases. Therefore, how to integrate large language models and expert knowledge to construct a space-ground collaborative decision-making mechanism while ensuring safety constraints, and how to ensure the continuous evolution of decision knowledge throughout mission execution, has become a pressing technical problem in this field. This paper presents a space-ground collaborative decision-making system and method that integrates large language models and expert knowledge. Summary of the Invention

[0004] The purpose of this invention is to provide a space-ground collaborative decision-making system and method that integrates large language models and expert knowledge, so as to solve the problems in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A space-ground collaborative decision-making method integrating large language models and expert knowledge includes the following steps: Step S1: Deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. Step S2: Construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector in the expert skill semantic space, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. Step S3: Obtain the execution result data of the satellite-ground collaborative decision-making, and modify the parameters of the expert Skill semantic space based on the execution result data.

[0006] Furthermore, the process of deploying a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism includes: An attitude reference self-calibration mechanism is formed by deploying a gravity gradient sensor array on the satellite body. The attitude reference self-calibration mechanism continuously measures the attitude deviation angle and angular velocity of the satellite body relative to the local vertical direction, calculates the attitude of the satellite body relative to the local vertical direction by gravity gradient tensor, and performs self-calibration of the attitude with reference to the inertial attitude reference provided by the star sensor, thereby obtaining the real-time attitude and attitude change trend of the satellite body without relying on continuous ground intervention, and then judging the real-time status of the satellite body. The attitude reference self-calibration mechanism is equipped with criteria for triggering the decision reporting process. When the attitude deviation angle exceeds a preset threshold, it is determined that the satellite body has entered a state that requires ground collaborative decision-making, thereby triggering the decision reporting process. The gravity gradient sensor array adopts a redundant configuration. When an individual gravity gradient sensor exhibits a measurement anomaly, the attitude reference self-calibration mechanism identifies the abnormal channel by comparing the consistency of measurement results between different gravity gradient sensors. It then performs a fusion calculation with the inertial attitude reference of the star sensor and the remaining normal channels to ensure continuous output of attitude deviation angle and angular velocity. The fusion calculation uses the inertial attitude reference provided by the star sensor as a long-term reference and the continuous measurement of the gravity gradient sensor array as a high-frequency input. It performs time alignment and error complementation between the two, so that even when the star sensor temporarily fails due to illumination or occlusion, the gravity gradient tensor can still maintain the short-term continuous operation of the attitude reference self-calibration mechanism.

[0007] Furthermore, the process of generating uplink and downlink data by determining the real-time status of the satellite through the attitude reference self-calibration mechanism includes: Once the decision reporting process is triggered, the key telemetry data related to this collaborative decision will be packaged. The key telemetry data includes the current orbital six elements, attitude quaternion, gravity gradient tensor, remaining fuel, collision probability, and mission requirement summary. The packaged key telemetry data is marked as satellite uplink and downlink data and transmitted to the ground station via the downlink. The downlink adopts a directional transmission method to reduce satellite-to-ground link overhead and improve the relevance of key telemetry data transmission.

[0008] Furthermore, the process of constructing an expert skill semantic space, generating corresponding task semantic vectors based on satellite uplink and downlink data, and determining whether a task semantic vector represents a novel task type includes: After receiving uplink and downlink data from the satellite, the ground station uses a large language model to perform semantic parsing on the key telemetry data and mission requirement summary, and maps the structured telemetry data and the mission requirements in natural language form into a unified mission semantic vector. The mission semantic vector is the unified input for subsequent expert skill matching and decision generation. An expert skill semantic space is pre-constructed. The expert skill semantic space is constructed with a multi-gradient field based on the dimensions of task type, fuel constraint, and time constraint. Each expert skill corresponds to the standard operation process and parameter template for a specific task scenario. The expert Skill semantic space determines whether the task semantic vector is a new type of task. Based on the determination result, a preliminary decision plan is generated, or the expert Skill semantic space generates a decision experience model based on the task semantic vector, and then generates a preliminary decision plan based on the decision experience model.

[0009] Furthermore, the process of conducting multi-gradient parallel search and generating space-ground collaborative decisions for novel mission types includes: The initial decision scheme is compared with the preset expert Skill standard output by the deviation rate. Based on the comparison results, the initial decision scheme is weighted and corrected according to expert knowledge. The weighted and corrected initial decision scheme is then sent to a multi-objective optimization program. The multi-objective optimization program takes minimum fuel consumption and shortest response time as optimization objectives. Minimum fuel consumption is used to constrain the remaining fuel consumption of attitude maneuvers or orbital maneuvers, and shortest response time is used to constrain the time cost from decision to on-board execution. Then, Pareto front search is used to find a non-dominated compromise solution among the various optimization objectives. This compromise solution is used as the space-ground collaborative decision.

[0010] Furthermore, the process of generating decision-making experience models for non-novel task types and leveraging historical experience to accelerate decision-making includes: When the task semantic vector is determined not to be a new type of task, the expert Skill semantic space generates a decision experience model based on the task semantic vector. Based on the decision experience model, historical tasks with the same task type, similar orbital position and similar space weather conditions are retrieved according to the task semantic vector, and the similarity between the current task and each historical task is calculated. If a match is successful, the historical best Skill parameter of the historical task is used as a pre-constraint input into the large language model. The large language model directly generates an accelerated decision-making scheme under the pre-constraint. The accelerated decision-making scheme only needs minor modifications before it can be sent to the satellite for execution. If the similarity between the current task and the historical task is less than or equal to the preset similarity threshold, then the historical experience is used as the initial solution to guide the optimization direction of the multi-gradient parallel search algorithm, so that the multi-gradient parallel search algorithm can expand the candidate skill matching path in the region pointed to by the historical experience, thereby reducing the number of iterations and accelerating the convergence speed.

[0011] Furthermore, the process of acquiring the execution result data of satellite-ground collaborative decision-making includes: After the satellite-ground collaborative decision-making system ascends to the satellite, it completes the corresponding attitude maneuver or orbital maneuver. After the maneuver is completed, the actual execution results are packaged and transmitted back to the ground. The actual execution results include attitude deviation residuals, fuel consumption, and maneuver completion time. The large language model compares the actual execution results with the preliminary decision-making scheme and the preset expert Skill standard output in multiple dimensions, and calculates the decision deviation rate, fuel consumption deviation and response time deviation for each item. Various deviations, along with their corresponding task conditions, are organized into execution deviation memory. This execution deviation memory records the deviation of each space-ground collaborative decision after it is actually executed on the satellite, serving as the basis for parameter modification of the expert Skill semantic space.

[0012] To ensure comparability of deviations across different dimensions, decision deviation rate, fuel consumption deviation, and response time deviation are all normalized according to their respective preset benchmarks and organized into an execution deviation memory sequence based on the execution time sequence of satellite-ground collaborative decision-making.

[0013] Furthermore, the process of modifying the parameters of the expert Skill semantic space based on the execution result data includes: The deviation data in the execution deviation memory is used as a stress signal to identify the source of the deviation and extract new rules. While modifying the parameters according to the identification results, the orbital position, season and space weather conditions when the deviation occurs are recorded, and the above environmental conditions are bound with the corresponding parameter adjustment results and new rules.

[0014] The space-ground collaborative decision-making system, which integrates large language models and expert knowledge, includes a space-ground data interaction module, a space-ground collaborative decision-making construction module, and a data self-calibration module. The satellite-to-ground data interaction module is used to deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. The satellite-ground collaborative decision-making construction module is used to construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. The data self-calibration module is used to acquire the execution result data of satellite-ground collaborative decision-making and modify the parameters of the expert Skill semantic space based on the execution result data.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention uses an on-board gravity gradient sensor array and a star sensor to form an attitude reference self-calibration mechanism, and downlink telemetry data is used as needed; a large language model generates task semantic vectors, and the optimal skill combination is matched through multi-gradient parallel search in the expert skill semantic space to form a preliminary decision scheme. The standard output of the expert skill is compared with the Pareto front search to generate a satellite-ground collaborative decision with the minimum fuel consumption and the shortest response time, taking into account both generalization and security.

[0016] 2. This invention constructs a decision-making experience model by using non-novel mission types, retrieves historical missions with similar mission types, orbital positions, and space weather conditions to generate accelerated decision-making schemes and shorten response time; and transmits on-board execution results back to form execution deviation memory, writes new rules or parameters into the stress signal increment to form parameter modification, and records environmental conditions with epigenetic markers, so that the expert skill semantic space can continuously evolve, thereby improving the accuracy and adaptability of decision-making in complex space environments to a certain extent. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the satellite-ground collaborative decision-making method that integrates large language models and expert knowledge as described in this invention.

[0019] Figure 2 This is a system block diagram of the space-ground collaborative decision-making system that integrates a large language model and expert knowledge as described in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, the space-ground collaborative decision-making method integrating large language models and expert knowledge includes the following steps: Step S1: Deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. Step S2: Construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector in the expert skill semantic space, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. Step S3: Obtain the execution result data of the satellite-ground collaborative decision-making, and modify the parameters of the expert Skill semantic space based on the execution result data.

[0022] Furthermore, step S1 is implemented through the following process: Step S101: Deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The specific process includes: A gravity gradient sensor array is deployed on the satellite body. The gravity gradient sensor array consists of multiple gravity gradient sensors arranged in a set spatial configuration to measure the gravity gradient tensor of the satellite body's location. At the same time, a star sensor is configured on the satellite body. The star sensor provides an inertial attitude reference by observing stars. The gravity gradient sensor array and the star sensor are combined to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism continuously measures the attitude deviation angle and angular velocity of the satellite body relative to the local vertical direction, calculates the attitude of the satellite body relative to the local vertical direction by gravity gradient tensor, and performs self-calibration of the attitude with reference to the inertial attitude reference provided by the star sensor, thereby obtaining the real-time attitude and attitude change trend of the satellite body without relying on continuous ground intervention, and then judging the real-time status of the satellite body. The attitude reference self-calibration mechanism is equipped with criteria for triggering the decision reporting process. When the attitude deviation angle exceeds a preset threshold, it is determined that the satellite body has entered a state that requires ground collaborative decision-making, thereby triggering the decision reporting process. The gravity gradient sensor array adopts a redundant configuration. When an individual gravity gradient sensor exhibits a measurement anomaly, the attitude reference self-calibration mechanism identifies the abnormal channel by comparing the consistency of measurement results between different gravity gradient sensors. It then performs a fusion calculation with the inertial attitude reference of the star sensor and the remaining normal channels to ensure continuous output of attitude deviation angle and angular velocity. The fusion calculation uses the inertial attitude reference provided by the star sensor as a long-term reference and the continuous measurement of the gravity gradient sensor array as a high-frequency input. It performs time alignment and error complementation between the two, so that even when the star sensor temporarily fails due to illumination or occlusion, the gravity gradient tensor can still maintain the short-term continuous operation of the attitude reference self-calibration mechanism.

[0023] Step S102: The satellite's real-time status is determined through an attitude reference self-calibration mechanism, generating uplink and downlink data. The specific process includes: Once the decision reporting process is triggered, the key telemetry data related to this collaborative decision will be packaged. The key telemetry data includes the current orbital six elements, attitude quaternion, gravity gradient tensor, remaining fuel, collision probability, and mission requirement summary. Among them, the orbital hexagram is used to characterize the current orbital state of the satellite, the attitude quaternion is used to characterize the current attitude of the satellite body, the gravity gradient tensor is used to characterize the gravitational environment in which the satellite is located, the remaining fuel is used to characterize the resource constraints for subsequent attitude maneuvers or orbital maneuvers, the collision probability is used to characterize the on-orbit safety margin, and the mission requirement summary is used to characterize the objectives and constraints of the mission to be performed. The packaged key telemetry data is marked as satellite uplink and downlink data and transmitted to the ground station via the downlink. The downlink adopts a directional transmission method to reduce satellite-to-ground link overhead and improve the targeting of key telemetry data transmission. The key telemetry data are uniformly timestamped and identified by the satellite during packaging. Downlink priorities are set for orbital six-element data, attitude quaternion data, gravity gradient tensor data, remaining fuel data, collision probability data, and mission requirement summary data according to their importance to collaborative decision-making. This ensures that data directly related to satellite safety, such as collision probability and attitude deviation, occupies the downlink first. When the downlink is restricted by the visible arc between the satellite and the ground, the on-board and downlink data are first cached on the satellite and queued in chronological order. Once the ground station enters the visible arc, the data is transmitted in a directional manner according to the downlink priority, ensuring that the ground station obtains key telemetry data that is temporally continuous, field-complete, and corresponds to the trigger time of the decision reporting process.

[0024] Furthermore, step S2 is implemented through the following process: Step S201: Construct the expert Skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, and determine whether the task semantic vector is a new type of task. The specific process includes: After receiving uplink and downlink data from the satellite, the ground station uses a large language model to perform semantic parsing on the key telemetry data and mission requirement summary, and maps the structured telemetry data and the mission requirements in natural language form into a unified mission semantic vector. The mission semantic vector is the unified input for subsequent expert skill matching and decision generation. The large language model includes an input encoding layer, a semantic reasoning layer, a skill scheduling layer, and a decision output layer. The input encoding layer is used to encode structured key telemetry data such as orbital six-element numbers, attitude quaternions, gravity gradient tensors, remaining fuel, and collision probabilities, along with the task requirement summary, into a unified task semantic vector. The semantic reasoning layer is used to identify the task type and understand the constraints of the task semantic vector. The skill scheduling layer is used to retrieve, match, and call the corresponding expert skills in the expert skill semantic space based on the task semantic vector. The decision output layer is used to organize the outputs of each expert skill into a preliminary decision scheme or an accelerated decision scheme, and to ensure that the output format is consistent with the preset expert skill standard output. The large language model on the ground side is responsible for semantic parsing, expert skill scheduling, and decision generation. Its reasoning process is constrained by the expert skill semantic space. That is, the large language model does not directly generate control commands on its own, but calls parameters, organizes processes, and outputs decisions within the scope defined by the standard operation procedures and parameter templates of the expert skills. This combines the semantic understanding and generalization ability of the large language model with the certainty and reliability of expert knowledge, and avoids the generated results from deviating from the safety constraints of the space mission. The ground station, the large language model, and the expert skill semantic space together constitute the ground collaborative decision-making link. An expert skill semantic space is pre-constructed. The expert skill semantic space is constructed with multiple gradient fields based on task type, fuel constraints, and time constraints. Each expert skill corresponds to the standard operation process and parameter template for a specific task scenario. That is, each expert skill encapsulates the handling process, input and output specifications, and parameter value template for a specific task scenario, so that expert knowledge is deposited in the expert skill semantic space in the form of callable and composable expert skills. The task semantic vector has clear coordinate components in four dimensions. The task type dimension corresponds to the semantic affiliation of the task semantic vector with various standard task scenarios. The fuel constraint dimension is determined by the remaining fuel amount and the expected consumption of the task. The time constraint dimension is determined by the time limit requirement in the task requirement summary and the orbit position. The four-dimensional coordinates of the task semantic vector correspond one-to-one with the multi-gradient field coordinates of the expert skill semantic space, so that any task semantic vector can be located at a certain position in the multi-gradient field, and the subsequent candidate skill matching path can be developed from this position. After the large language model inputs the task semantic vector into the expert skill semantic space, it first determines whether the task semantic vector is a novel task type. A novel task type refers to a task type that does not have a directly reusable historical task in the expert skill semantic space and requires re-optimization through expert skill matching. If it is determined to be a novel task type, the task semantic vector is input into the expert skill semantic space for multi-gradient parallel search to generate a preliminary decision scheme. If it is determined not to be a novel task type, the expert skill semantic space generates a decision experience model based on the task semantic vector, and then generates a preliminary decision scheme based on the decision experience model.

[0025] Step S202: Perform multi-gradient parallel search for the new mission type and generate satellite-ground collaborative decision-making. The specific process includes: The expert skill semantic space uses a multi-gradient parallel search algorithm to solve for new task types, so that the task semantic vector can simultaneously expand several candidate skill matching paths along three dimensions: task type, fuel constraint, and time constraint. The expansion of the candidate skill matching paths is similar to the directional growth of roots in a multi-branched system towards a high water potential region, that is, each candidate skill matching path extends in parallel along four dimensions in the direction of increasing matching degree. In the multi-gradient parallel search process, computing resources are allocated according to the matching degree between each candidate Skill matching path and the task semantic vector. Candidate Skill matching paths with high matching degree obtain computing resources and are extended first, while candidate Skill matching paths with low matching degree are pruned in advance, so that the search process quickly converges to the optimal Skill combination. The optimal Skill combination is a number of expert Skills and their calling order required to complete the current task. Call each expert skill in the optimal skill combination and pass in parameters such as the orbital six roots, attitude, collision probability, remaining fuel and mission requirements. Each expert skill will work together to solve the problem according to its own standard operating procedure and parameter template to generate a preliminary decision scheme. The preliminary decision-making scheme is compared with the preset expert Skill standard output by a deviation rate. The deviation rate is used to characterize the degree of deviation between the preliminary decision-making scheme and the standard result recognized by expert knowledge. When the deviation rate is less than the deviation threshold, the preliminary decision-making scheme is determined to meet the requirements of expert knowledge and is directly adopted as the satellite-ground collaborative decision. When the deviation rate exceeds the deviation threshold, the preliminary decision scheme is weighted and corrected based on expert knowledge, and the weighted and corrected preliminary decision scheme is sent to a multi-objective optimization program. The multi-objective optimization program takes minimum fuel consumption and shortest response time as optimization objectives. Minimum fuel consumption is used to constrain the remaining fuel consumption of attitude maneuvers or orbital maneuvers, and shortest response time is used to constrain the time cost from decision to on-board execution. Then, Pareto front search is used to find a non-dominated compromise solution among the various optimization objectives, and the compromise solution is used as the satellite-ground collaborative decision. After obtaining the Pareto front solution set using Pareto front search, a compromise solution is selected from the Pareto front solution set as the satellite-ground collaborative decision based on the emphasis of the task semantic vector on the dimensions of fuel constraints and time constraints. The optimal combination of skills called by the satellite-ground collaborative decision, the values ​​of key parameters, and the deviation rate relative to the preset expert skill standard output are given simultaneously. This allows the satellite-ground collaborative decision to be executed on the satellite and its expert knowledge source and multi-objective trade-off basis to be traced.

[0026] Step S203: Generate a decision-making experience model for non-novel task types and utilize historical experience to accelerate decision-making. The specific process includes: When the task semantic vector is determined not to be a new type of task, the expert Skill semantic space generates a decision experience model based on the task semantic vector. The decision experience model is derived from historical tasks of the same task type and their historical best Skill parameters and execution effects, and is used to characterize the mapping relationship between task conditions and optimal decisions under this task type. Based on the decision-making experience model, historical tasks with the same task type, similar orbital positions and similar space weather conditions are retrieved according to the task semantic vector, and the similarity between the current task and each historical task is calculated. The similarity between the current task and each historical task is calculated as follows: Let the current task be T0, and the i-th historical task be Ti, i=1,2,…,N, where N is the number of historical tasks in the decision experience model that belong to the same task type as the current task. First, the features of the current task and each historical task in terms of task type, orbital position and space weather conditions are extracted from the task semantic vector, and the feature components of orbital position and space weather conditions are normalized according to their dimensions. Then, the similarity of task type S1(i), orbital position S2(i) and space weather conditions S3(i) are calculated respectively, and all three are normalized to the interval [0,1]. The similarity of task type is taken as the cosine similarity of the task type feature vectors v0 and vi of the current task and the i-th historical task, S1(i) = (v0·vi) / (‖v0‖·‖vi‖). If a match is successful, that is, if there is a historical task with a similarity greater than the preset similarity threshold, the historical best Skill parameter of the historical task is used as a pre-constraint input into the large language model. The large language model directly generates an accelerated decision-making scheme under the pre-constraint. The accelerated decision-making scheme only needs minor modifications before it can be sent up to the satellite for execution. If the similarity between the current task and the historical task is less than or equal to the preset similarity threshold, then the historical experience is used as the initial solution to guide the optimization direction of the multi-gradient parallel search algorithm, so that the multi-gradient parallel search algorithm can expand the candidate Skill matching path in the region pointed to by the historical experience, thereby reducing the number of iterations and accelerating the convergence speed. The memory weight of historical tasks in the decision-making experience model changes dynamically with time and the effect of calling. The memory weight of the historical best Skill parameter that is repeatedly and successfully called gradually increases, while the memory weight of historical tasks that have not been called for a long time or whose execution effect has deteriorated gradually decreases. This ensures the efficiency of experience reuse while avoiding outdated experience from misleading the current decision. The decision-making experience model is continuously updated with each task execution: when a task is successfully executed, the memory weight of the successful solution with the corresponding expert skill and the historical best skill parameters is strengthened; when a task fails, the memory weight of the corresponding solution is weakened and the semantic space of the expert skill is modified, so that the weakened solution will not be reused in subsequent decisions of the same task type. This forms an experience reuse and self-correction mechanism for non-new task types, and finally generates a space-ground collaborative decision by comparing the preliminary decision solution with the preset expert skill standard output.

[0027] Furthermore, step S3 is implemented through the following process: Step S301: Obtain the execution result data of the satellite-ground collaborative decision-making process, which includes: After the satellite-ground collaborative decision-making system ascends to the satellite, it completes the corresponding attitude maneuver or orbital maneuver. After the maneuver is completed, the actual execution results are packaged and transmitted back to the ground. The actual execution results include attitude deviation residuals, fuel consumption, and maneuver completion time. The ground-based large language model compares the actual execution results with the preliminary decision plan and the preset expert Skill standard output in multiple dimensions, and calculates the decision deviation rate, fuel consumption deviation and response time deviation item by item. The decision deviation rate corresponds to the degree of deviation of the overall decision result, the fuel consumption deviation corresponds to the difference between the fuel consumption and the expected value of the decision, and the response time deviation corresponds to the difference between the maneuver completion time and the expected response time of the decision. Various deviations, along with their corresponding task conditions, are organized into execution deviation memory. This execution deviation memory records the deviation of each space-ground collaborative decision after it is actually executed on the satellite, serving as the basis for parameter modification of the expert Skill semantic space.

[0028] To ensure comparability of deviations of different dimensions, decision deviation rate, fuel consumption deviation and response time deviation are all normalized according to their respective preset benchmarks and organized into an execution deviation memory sequence according to the execution time sequence of satellite-ground collaborative decision-making. The execution deviation memory simultaneously stores the corresponding orbital six-axis number, space weather conditions, and the expert skill combination identifier, so that each execution deviation memory can be traced back to the task conditions and decision-making path that caused the deviation, providing a complete basis for subsequent identification of the deviation source, extraction of new rules, and parameter modification.

[0029] Step S302: Modify the parameters of the expert Skill semantic space based on the execution result data. The specific process includes: The deviation data in the execution deviation memory is used as a coercive signal to identify the source of the deviation and extract new rules; If the deviation originates from a new scenario not covered by the expert skill, that is, the existing expert skill's standard operation process and parameter template do not contain the corresponding handling knowledge for the scenario, then the extracted new rules will be written into the expert skill semantic space incrementally, expanding the scenario coverage of the expert skill semantic space without changing the existing expert skills. If the deviation stems from the parameters of an existing expert skill being unsuitable for the new environment, meaning the handling process is still applicable but the parameter values ​​need to be adjusted, then the parameter template of the corresponding expert skill is updated in the direction of reducing the execution deviation, thus forming parameter modification. While modifying the parameters, the orbital position, season, and space weather conditions at the time of the deviation are recorded, and the above environmental conditions are bound with the corresponding parameter adjustment results and new rules, and stored as epigenetic markers in the expert Skill semantic space. When subsequent missions are in the same or similar orbital positions, seasons and space weather conditions, the expert skills and parameters carrying the corresponding epigenetic markers are called first, so that the semantic space of the expert skills can be adapted to environmental changes in a heritable and reusable manner without changing the core mechanism. There is a clear boundary between incrementally writing new rules and fine-tuning the parameters of existing expert skills: the former is used to expand the task scenarios not yet covered by the semantic space of expert skills, and the latter is used to adjust the parameter template without changing the standard operating procedure. Both need to pass the regression verification of the preset expert skill standard output. That is, the output of the modified expert skill on historical tasks must not be worse than before the modification. If the regression verification fails, the incremental writing or parameter fine-tuning is cancelled, so as to ensure that the parameter modification can continuously adapt to the new environment while maintaining the stability and reliability of the semantic space of expert skills. After parameter modification, the expert Skill semantic space is returned to step S2, which serves as the basis for semantic vector matching and satellite-ground collaborative decision generation in the next cycle. This creates a closed loop consisting of satellite status perception reporting, ground-based large language model and expert knowledge collaborative decision-making, satellite execution result feedback, and expert Skill semantic space parameter modification. The accuracy and environmental adaptability of satellite-ground collaborative decision-making continue to improve with the number of mission executions.

[0030] Please see Figure 2 As shown, the space-ground collaborative decision-making system, which integrates large language models and expert knowledge, includes a space-ground data interaction module, a space-ground collaborative decision-making construction module, and a data self-calibration module. The satellite-to-ground data interaction module is used to deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. The satellite-ground collaborative decision-making construction module is used to construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. The data self-calibration module is used to acquire the execution result data of satellite-ground collaborative decision-making and modify the parameters of the expert Skill semantic space based on the execution result data.

[0031] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A space-ground collaborative decision-making method integrating large language models and expert knowledge, characterized in that: Includes the following steps: Step S1: Deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. Step S2: Construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector in the expert skill semantic space, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. Step S3: Obtain the execution result data of the satellite-ground collaborative decision-making, and modify the parameters of the expert Skill semantic space based on the execution result data.

2. The satellite-ground collaborative decision-making method integrating a large language model and expert knowledge according to claim 1, characterized in that, The process of deploying a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism includes: An attitude reference self-calibration mechanism is formed by deploying a gravity gradient sensor array on the satellite body. The attitude reference self-calibration mechanism continuously measures the attitude deviation angle and angular velocity of the satellite body relative to the local vertical direction, calculates the attitude of the satellite body relative to the local vertical direction by gravity gradient tensor, and performs self-calibration of the attitude with reference to the inertial attitude reference provided by the star sensor, thereby obtaining the real-time attitude and attitude change trend of the satellite body without relying on continuous ground intervention, and then judging the real-time status of the satellite body. The attitude reference self-calibration mechanism is equipped with criteria for triggering the decision reporting process. When the attitude deviation angle exceeds a preset threshold, it is determined that the satellite body has entered a state that requires ground collaborative decision-making, thereby triggering the decision reporting process. The gravity gradient sensor array adopts a redundant configuration. When an individual gravity gradient sensor exhibits a measurement anomaly, the attitude reference self-calibration mechanism identifies the abnormal channel by comparing the consistency of measurement results between different gravity gradient sensors. It then performs a fusion calculation with the inertial attitude reference of the star sensor and the remaining normal channels to ensure continuous output of attitude deviation angle and angular velocity. The fusion calculation uses the inertial attitude reference provided by the star sensor as a long-term reference and the continuous measurement of the gravity gradient sensor array as a high-frequency input. It performs time alignment and error complementation between the two, so that even when the star sensor temporarily fails due to illumination or occlusion, the gravity gradient tensor can still maintain the short-term continuous operation of the attitude reference self-calibration mechanism.

3. The space-ground collaborative decision-making method integrating large language models and expert knowledge according to claim 2, characterized in that, The process of generating uplink and downlink data by determining the real-time status of the satellite through an attitude reference self-calibration mechanism includes: Once the decision reporting process is triggered, the key telemetry data related to this collaborative decision will be packaged. The key telemetry data includes the current orbital six elements, attitude quaternion, gravity gradient tensor, remaining fuel, collision probability, and mission requirement summary. The packaged key telemetry data is marked as satellite uplink and downlink data and transmitted to the ground station via the downlink. The downlink adopts a directional transmission method to reduce satellite-to-ground link overhead and improve the relevance of key telemetry data transmission.

4. The space-ground collaborative decision-making method integrating large language models and expert knowledge according to claim 3, characterized in that, The process of constructing an expert skill semantic space, generating corresponding task semantic vectors based on satellite uplink and downlink data, and determining whether a task semantic vector represents a novel task type includes: After receiving uplink and downlink data from the satellite, the ground station uses a large language model to perform semantic parsing on the key telemetry data and mission requirement summary, and maps the structured telemetry data and the mission requirements in natural language form into a unified mission semantic vector. The mission semantic vector is the unified input for subsequent expert skill matching and decision generation. An expert skill semantic space is pre-constructed. The expert skill semantic space is constructed with a multi-gradient field based on the dimensions of task type, fuel constraint, and time constraint. Each expert skill corresponds to the standard operation process and parameter template for a specific task scenario. The expert Skill semantic space determines whether the task semantic vector is a new type of task. Based on the determination result, a preliminary decision plan is generated, or the expert Skill semantic space generates a decision experience model based on the task semantic vector, and then generates a preliminary decision plan based on the decision experience model.

5. The satellite-ground collaborative decision-making method integrating a large language model and expert knowledge according to claim 4, characterized in that, The process of performing multi-gradient parallel search and generating space-ground collaborative decisions for novel mission types includes: The initial decision scheme is compared with the preset expert Skill standard output by the deviation rate. Based on the comparison results, the initial decision scheme is weighted and corrected according to expert knowledge. The weighted and corrected initial decision scheme is then sent to a multi-objective optimization program. The multi-objective optimization program takes minimum fuel consumption and shortest response time as optimization objectives. Minimum fuel consumption is used to constrain the remaining fuel consumption of attitude maneuvers or orbital maneuvers, and shortest response time is used to constrain the time cost from decision to on-board execution. Then, Pareto front search is used to find a non-dominated compromise solution among the various optimization objectives. This compromise solution is used as the space-ground collaborative decision.

6. The satellite-ground collaborative decision-making method integrating a large language model and expert knowledge according to claim 5, characterized in that, The process of generating decision-making experience models for non-novel task types and leveraging historical experience to accelerate decision-making includes: When the task semantic vector is determined not to be a new type of task, the expert Skill semantic space generates a decision experience model based on the task semantic vector. Based on the decision experience model, historical tasks with the same task type, similar orbital position and similar space weather conditions are retrieved according to the task semantic vector, and the similarity between the current task and each historical task is calculated. If a match is successful, the historical best Skill parameter of the historical task is used as a pre-constraint input into the large language model. The large language model directly generates an accelerated decision-making scheme under the pre-constraint. The accelerated decision-making scheme only needs minor modifications before it can be sent to the satellite for execution. If the similarity between the current task and the historical task is less than or equal to the preset similarity threshold, then the historical experience is used as the initial solution to guide the optimization direction of the multi-gradient parallel search algorithm, so that the multi-gradient parallel search algorithm can expand the candidate skill matching path in the region pointed to by the historical experience, thereby reducing the number of iterations and accelerating the convergence speed.

7. The satellite-ground collaborative decision-making method integrating a large language model and expert knowledge according to claim 6, characterized in that, The process of acquiring the execution result data of satellite-ground collaborative decision-making includes: After the satellite-ground collaborative decision-making system ascends to the satellite, it completes the corresponding attitude maneuver or orbital maneuver. After the maneuver is completed, the actual execution results are packaged and transmitted back to the ground. The actual execution results include attitude deviation residuals, fuel consumption, and maneuver completion time. The large language model compares the actual execution results with the preliminary decision-making scheme and the preset expert Skill standard output in multiple dimensions, and calculates the decision deviation rate, fuel consumption deviation and response time deviation for each item. Various deviations, along with their corresponding task conditions, are organized into execution deviation memory. This execution deviation memory records the deviation of each space-ground collaborative decision after it is actually executed on the satellite, serving as the basis for parameter modification of the expert Skill semantic space. To ensure comparability of deviations across different dimensions, decision deviation rate, fuel consumption deviation, and response time deviation are all normalized according to their respective preset benchmarks and organized into an execution deviation memory sequence based on the execution time sequence of satellite-ground collaborative decision-making.

8. The satellite-ground collaborative decision-making method integrating a large language model and expert knowledge according to claim 7, characterized in that, The process of modifying the parameters of the expert Skill semantic space based on the execution result data includes: The deviation data in the execution deviation memory is used as a stress signal to identify the source of the deviation and extract new rules. While modifying the parameters according to the identification results, the orbital position, season and space weather conditions when the deviation occurs are recorded, and the above environmental conditions are bound with the corresponding parameter adjustment results and new rules.

9. A space-ground collaborative decision-making system integrating a large language model and expert knowledge, used to implement the space-ground collaborative decision-making method integrating a large language model and expert knowledge as described in any one of claims 1-8, characterized in that, It includes a satellite-to-ground data interaction module, a satellite-to-ground collaborative decision-making construction module, and a data self-calibration module; The satellite-to-ground data interaction module is used to deploy a gravity gradient sensor array and a star sensor combination on the satellite to form an attitude reference self-calibration mechanism. The attitude reference self-calibration mechanism is used to determine the real-time status of the satellite and generate uplink and downlink data. The satellite-ground collaborative decision-making construction module is used to construct an expert skill semantic space, generate corresponding task semantic vectors based on satellite uplink and downlink data, determine whether the task semantic vector is a new type of task, generate a decision experience model based on the judgment result and the task semantic vector, generate a preliminary decision scheme based on the decision experience model, and generate a satellite-ground collaborative decision by comparing the preliminary decision scheme with the preset expert skill standard output. The data self-calibration module is used to acquire the execution result data of satellite-ground collaborative decision-making and modify the parameters of the expert Skill semantic space based on the execution result data.

Images