Spacecraft formation flying mission planning method based on large language model and tool set cooperation

Abstract

The application particularly relates to a spacecraft formation flight task planning method based on cooperation of a large language model and a professional tool set, and comprises the following steps: constructing a three-layer task planning architecture of the professional tool set, the LLM and a high-level dynamic execution engine; constructing the professional tool set for the spacecraft formation flight; decomposing a user instruction into a meta-task sequence through a complex task decomposition model; and calling the tool set to execute the meta-task in sequence through a task sequence execution model. The application provides a more efficient, flexible and intelligent solution with scene generalization capability for the spacecraft formation flight task planning problem.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a spacecraft formation flight mission planning method that integrates a large language model and toolset. Background Technology

[0002] In related technologies, spacecraft formation flying has important applications in multiple fields such as deep space exploration, scientific observation, and communication enhancement, enabling high-precision measurements and flexible deployments that cannot be achieved by a single spacecraft. Mission planning for formation flying is crucial in these missions, as it directly determines how spacecraft cooperate to achieve predetermined objectives. However, mission planning for spacecraft formation flying faces multiple challenges, including the complexity of orbital dynamics, coordination of multi-stage missions, and optimal control under thrust constraints. While traditional spacecraft formation flying mission planning methods can effectively solve mission planning problems in specific scenarios, their inherent limitations cannot be ignored: First, current methods rely on manually designed rules or cost functions, making it difficult to handle undefined complex situations and resulting in insufficient adaptability; second, mission adjustments require repeated modeling, making it difficult to quickly complete planning based on actual mission requirements and resulting in insufficient flexibility; third, they lack natural language processing capabilities, hindering convenient language interaction with users. Therefore, the problem of spacecraft formation flying mission planning urgently requires more efficient and intelligent solutions.

[0003] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a spacecraft formation flight mission planning method that combines a large language model with a toolset, which can effectively overcome the shortcomings of existing technologies.

[0005] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0006] According to a first aspect of the present invention, a method for spacecraft formation flight mission planning based on the collaboration of a large language model and toolset is provided, the method comprising:

[0007] The dynamic execution engine business layer obtains the current pending tasks, generates task decomposition prompts based on the pending tasks, tool information, task decomposition prompt templates, and preset manual task decomposition examples, and calls the task decomposition model to decompose and process the pending tasks; wherein, the tool information is obtained by the dynamic execution engine business layer from the toolset business layer.

[0008] The large language model business layer uses a task decomposition model, combined with the task decomposition prompts, to decompose the task to be processed and obtain the corresponding meta-task sequence; and returns the meta-task sequence to the dynamic execution engine business layer; wherein, the meta-task sequence includes several sequentially executed meta-tasks; the meta-tasks are tasks that can be completed without calling tools or with a single call to tools;

[0009] The dynamic execution engine business layer constructs task execution prompts based on the meta-task sequence, tool information, and tool call prompt template; and calls the task sequence execution model to process the task prompts.

[0010] The large language model business layer utilizes the task sequence execution model, combines task execution prompts to call the tools corresponding to the meta-tasks to the toolset business layer to complete the execution of the meta-tasks, integrates the task data corresponding to each meta-task to generate task planning data, and returns the task planning data to the dynamic execution engine business layer.

[0011] In some exemplary embodiments, the method further includes: when the dynamic execution engine business layer obtains a task to be processed, obtaining the currently available tools in the toolset from the toolset business layer to determine the tool information.

[0012] In some exemplary embodiments, the large language model business layer utilizes a task sequence execution model, combined with task execution prompts, to call the tool corresponding to the meta-task from the toolset business layer to complete the execution of the meta-task, including:

[0013] The task sequence execution model executes each meta-task in sequence and identifies whether the current meta-task requires a tool call.

[0014] When it is determined that the current meta-task requires a tool call, the corresponding tool name and tool input parameters are determined; and the corresponding tool is called from the toolset business layer according to the tool name, and the tool input parameters are processed to obtain the tool's feedback data.

[0015] In some exemplary embodiments,

[0016] Based on the tool name, the corresponding tool is invoked at the toolset's business layer, and the tool's input parameters are processed to obtain the tool's feedback data, including:

[0017] The task sequence execution model feeds back the tool name and tool input parameters to the dynamic execution engine business layer;

[0018] The dynamic execution engine business layer calls the corresponding tool from the toolset business layer according to the tool name, processes the tool input parameters, and obtains the feedback data from the tool to the dynamic execution engine business layer;

[0019] The dynamic execution engine business layer sends the tool's feedback data to the large language model business layer, so that the task sequence execution model can obtain the tool's feedback data.

[0020] In some exemplary embodiments, when the dynamic execution engine business layer invokes the corresponding tool from the toolset business layer based on the tool name and processes the tool input parameters, the method further includes:

[0021] The tools in the toolset business layer process the tool input parameters and send the feedback data to the task sequence execution model.

[0022] In some exemplary embodiments, when the task sequence execution model executes each meta-task sequentially and identifies whether the current meta-task requires a tool call, the method further includes:

[0023] The task sequence execution model returns the tool and tool input parameters corresponding to the current meta-task based on the actual execution results of the meta-task sequence and the preceding meta-tasks.

[0024] In some exemplary embodiments, the method further includes:

[0025] A pre-built task planning architecture is provided, comprising a toolset business layer, a large language model business layer, and a dynamic execution engine business layer configured sequentially. The toolset business layer integrates multiple tools for processing meta-tasks and feeds task data back to the large language model business layer and / or the dynamic execution engine business layer. The large language model business layer is used to parse natural language instructions and generate meta-task sequences. The dynamic execution engine business layer is used to acquire tasks to be processed and coordinate the real-time interaction between the toolset business layer and the large language model business layer.

[0026] Configure standardized tool interfaces for the toolset business layer, including: tool name, tool function description, and parameter format; wherein, the parameter format includes the parameter name, parameter description, and parameter data format followed by each parameter node;

[0027] Configure a multi-dimensional parameter selection mechanism, including: a multi-layered nested array dimension constraint mechanism for constraining high-dimensional state vectors, a finite range constraint mechanism for explicitly defining the legal input value range, a default parameter completion mechanism based on domain knowledge, and a parameter polymorphism adaptation mechanism for achieving parameter polymorphism adaptation.

[0028] In some exemplary embodiments, the toolset business layer includes:

[0029] Web search tools are used to access the internet and search for data.

[0030] Knowledge graph tools are used to query local knowledge bases;

[0031] A formation-scale design tool for recommending formation arm lengths based on the required gravitational wave frequency bands to be detected;

[0032] Formation configuration generation tool, used to generate formation configurations based on formation arm length and spacecraft phase angle;

[0033] A pulse sequence calculation tool used to solve for the pulse sequence required for orbit transfer;

[0034] The trajectory drawing tool is used to draw the transfer trajectory of a spacecraft under pulse action;

[0035] Stability analysis tools are used to assess the long-term stability of formations under perturbation environments.

[0036] According to a second aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the above-described spacecraft formation flight mission planning method that combines a large language model and a toolset is implemented.

[0037] According to a third aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described spacecraft formation flight mission planning method that combines a large language model with a toolset.

[0038] According to a fourth aspect of the present invention, an electronic device is provided, comprising:

[0039] Processor; and

[0040] Memory for storing the executable instructions of the processor;

[0041] The processor is configured to implement the aforementioned spacecraft formation flight mission planning method that combines a large language model and toolset when executing the executable instructions.

[0042] The spacecraft formation flight mission planning method provided by embodiments of this invention, which integrates a large language model and toolset, constructs a mission planning architecture based on a toolset business layer, a large language model business layer, and a dynamic execution engine business layer. The large language model business layer and the dynamic execution engine business layer lower the professional threshold for users through natural language interaction, allowing non-expert users to dynamically define and adjust mission requirements through multi-turn dialogues. By designing a large language model business layer capable of parsing natural language commands and generating meta-task sequences, the reasoning process can be explicitly presented, significantly improving the execution transparency of complex tasks. By integrating multiple tools for mission planning into the toolset business layer, which can be freely invoked by the large language model business layer and the dynamic execution engine business layer, efficient and flexible handling of spacecraft formation flight mission planning problems is achieved. This provides an intelligent solution with scenario generalization capabilities.

[0043] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0044] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0045] Figure 1 This illustration schematically shows an exemplary embodiment of the present invention: a spacecraft formation flight mission planning method that integrates a large language model and toolset.

[0046] Figure 2 The diagram illustrates an overall task planning framework according to an exemplary embodiment of the present invention.

[0047] Figure 3 This diagram illustrates a specific process of end-to-end task planning according to an exemplary embodiment of the present invention.

[0048] Figure 4 The diagram illustrates the composition of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0050] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0051] Among related technologies, LLM has shown great potential in complex mission planning, with its core advantages reflected in the following three aspects: First, it lowers the professional threshold for users through natural language interaction, allowing non-expert users to dynamically define and adjust mission requirements through multi-turn dialogues; second, with the help of prompting strategies such as thought chains, LLM can explicitly present the reasoning process, significantly improving the interpretability of planning results; and third, its internalized "world knowledge" and cross-domain generalization ability support autonomous adaptation to unexpected scenarios. Nevertheless, in the field of spacecraft formation flight mission planning, the application of LLM still faces the following challenges: First, the capabilities of LLM itself cannot meet the high-precision modeling requirements of aerospace dynamics; second, for specific mission scenarios, there is a lack of high-quality training data, resulting in insufficient domain knowledge for LLM.

[0052] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a spacecraft formation flight mission planning method that integrates a large language model and toolset. (Reference) Figure 1 As shown, it can specifically include:

[0053] Step S11: The dynamic execution engine business layer obtains the current pending tasks, generates task decomposition prompts based on the pending tasks, tool information, task decomposition prompt template, and preset manual task decomposition examples; and calls the task decomposition model to decompose and process the pending tasks; wherein, the tool information is obtained by the dynamic execution engine business layer from the toolset business layer.

[0054] Step S12: The large language model business layer uses the task decomposition model and the task decomposition prompt words to decompose the task to be processed and obtain the corresponding meta-task sequence; and returns the meta-task sequence to the dynamic execution engine business layer; wherein, the meta-task sequence includes several sequentially executed meta-tasks; the meta-tasks are tasks that can be completed without calling tools or with a single call to tools.

[0055] Step S13: The dynamic execution engine business layer constructs task execution prompts based on the meta-task sequence, tool information, and tool call prompt template; and calls the task sequence execution model to process the task prompts.

[0056] Step S14: The large language model business layer uses the task sequence execution model, combined with the task execution prompt words, to call the tool corresponding to the meta-task to the toolset business layer to complete the execution of the meta-task, and integrates the task data corresponding to each meta-task to generate task planning data; and returns the task planning data to the dynamic execution engine business layer.

[0057] The following will describe in more detail each step of the spacecraft formation flight mission planning method that combines a large language model and toolset in this exemplary embodiment, with reference to the accompanying drawings and embodiments.

[0058] In this example implementation, the method may include:

[0059] Step S21: Pre-construct a task planning architecture; wherein, the task planning architecture includes a toolset business layer, a large language model business layer, and a dynamic execution engine business layer configured sequentially; the toolset business layer integrates multiple tools for processing meta-tasks and feeds task data back to the large language model business layer and / or the dynamic execution engine business layer; the large language model business layer is used to parse natural language instructions and generate meta-task sequences; the dynamic execution engine business layer is used to acquire tasks to be processed and coordinate the real-time interaction between the toolset business layer and the large language model business layer;

[0060] Step S22: Configure standardized tool interfaces for the toolset business layer, including: tool name, tool function description, and parameter format; wherein, the parameter format includes the parameter name, parameter description, and parameter data format followed by each parameter node;

[0061] Step S23: Configure a multi-dimensional parameter selection mechanism, including: a multi-layer nested array dimension constraint mechanism for constraining high-dimensional state vectors, a finite range constraint mechanism for specifying the legal input value range, a default parameter completion mechanism based on domain knowledge, and a parameter polymorphism adaptation mechanism for achieving parameter polymorphism adaptation.

[0062] Specifically, a task planning architecture can be built on the server side. The user inputs the current task data on their terminal device and sends it to the server. The server then executes the task planning architecture, analyzes the task data, generates corresponding task planning data, and sends this data back to the user's terminal device. Alternatively, the task planning architecture can be built on a terminal device with a certain computing power, and the user's input task data can be processed locally.

[0063] The toolset business layer, located at the bottom of the architecture, is equipped with a professional toolset that integrates multiple tools for task planning and can be called by the large language model business layer and the dynamic execution engine business layer.

[0064] The large language model business layer, located in the middle layer of the architecture, is configured with a large language model (LLM) capable of parsing natural language instructions and generating meta-task sequences. Furthermore, it can utilize specialized toolsets to perform orbital dynamics calculations, optimization solutions, constraint verification, and more.

[0065] The high-level dynamic execution engine business layer can interact with users, receive user input, and provide feedback on task planning results; it also coordinates the real-time interaction between the LLM and the toolset. Based on this business architecture, the calling rules between various business layers can be configured.

[0066] For example, when building a task planning architecture for the toolset business layer, a standardized definition system for tool interfaces can be configured, including tool names. Tool Function Description Parameter format For the parameter format, each parameter node follows the parameter name. Parameter Description Parameter data format The standard format.

[0067] For example, the standard format of a tool interface may include:

[0068] {

[0069] "name for human": A1,

[0070] "name for model": A1,

[0071] "description": A2,

[0072] "ags": A3

[0073] }

[0074] The standard format of parameters may include:

[0075] ags":[

[0076] {

[0077] "name": B1,

[0078] "description": B2,

[0079] "'schema": B3("min / maxItems"+"enum"+ "default"+"oneOf")

[0080] } ]

[0082] In addition, a multi-dimensional parameter selection mechanism can be built for the toolset business layer, including:

[0083] 1) Multi-level nested array dimension constraint mechanism. High-dimensional state vectors are handled by nested array dimension constraints. Each level constrains the array dimension through the "minItems" and "maxItems" fields, which solves the problem of mismatch between high-dimensional input parameters and the format required by the tool.

[0084] 2) Finite value range constraint mechanism. For input parameters in a finite space, the legal input value range is clearly defined through the "enum" field, ensuring the controllability of the input space.

[0085] 3) Domain-Knowledge-Based Default Parameter Completion Mechanism. Users typically issue high-level, coarse-grained commands, making it unreasonable and unrealistic to expect the user's semantics to include all the information of the tool parameters. Therefore, the "default" field specifies default parameters to resolve the contradiction between the ambiguity of user semantics and the completeness of the required tool parameters.

[0086] 4) Task-oriented parameter polymorphism adaptation mechanism. Depending on the actual task, the parameter mode of the tool sometimes needs to be flexibly selected. The parameter mode can be switched through the "oneOf" field, realizing parameter polymorphism adaptation.

[0087] For example, the toolset business layer includes:

[0088] Web search tools are used to access the internet and search for data.

[0089] Knowledge graph tools are used to query local knowledge bases;

[0090] A formation-scale design tool for recommending formation arm lengths based on the required gravitational wave frequency bands to be detected;

[0091] Formation configuration generation tool, used to generate formation configurations based on formation arm length and spacecraft phase angle;

[0092] A pulse sequence calculation tool used to solve for the pulse sequence required for orbit transfer;

[0093] The trajectory drawing tool is used to draw the transfer trajectory of a spacecraft under pulse action;

[0094] Stability analysis tools are used to assess the long-term stability of formations under perturbation environments.

[0095] Specifically, the web search tool is a general-purpose search engine interface based on the SerpAPI (Google Search Engine Results API). Its main function is to achieve real-time retrieval of internet information through API calls. This tool can cover diverse search needs, including encyclopedic knowledge queries, current affairs news tracking, and academic resource acquisition. Users only need to enter search keywords or phrases, and the tool will return relevant search results. The input parameters for the web search tool are shown in Table 1.

[0096] Table 1

[0097]

[0098] When a task relies on dynamic information that is highly real-time or beyond the scope of LLM's inherent knowledge, web search tools should be used. For example, when faced with a question such as "Please introduce the latest progress and expected timeline of China's gravitational wave detection 'Taiji Program'", LLM's own knowledge may not be able to provide a sufficiently novel or accurate answer. In this case, this tool should be used, and the user's question should be entered into "search_query" to obtain the latest information published by authoritative sources to ensure the timeliness and accuracy of the answer.

[0099] Specifically, for the knowledge graph tool, the formation knowledge graph is a local expert knowledge base built on structured data, supporting professional queries on mission objectives, constraints, and the physical meaning and reference thresholds of formation stability indices in typical space-based gravitational wave mission scenarios. The input parameters for the formation knowledge graph tool are shown in Table 2.

[0100] Table 2

[0101]

[0102] This tool specifies the valid value range of input parameters through JSON Schema fragments, effectively ensuring that the input parameters provided by LLM do not exceed the limits. For example, the data format of the "metric_name" parameter adopts a limited value range constraint mechanism. Through the field "enum": ["bight length change", "bight length change rate", "breathing angle change", "formation phase angle change", "formation tilt angle change", "all metrics"], it is indicated that LLM can only query the above five types of metrics.

[0103] When user tasks involve sophisticated professional parameters, metric definitions, or domain-specific knowledge, relying solely on LLM generation or general web searches cannot guarantee accuracy and professionalism. In such cases, a formation knowledge graph tool should be invoked. For example, when handling tasks such as "Please provide recommended stability metrics for gravitational wave detection formations," this tool should be used, with "Stability Metrics" entered in "query_type," "All Metrics" in "metric_name," and "Reference Threshold" in "attribute_type" to obtain a reliable and professional answer.

[0104] Specifically, for the formation-scale design tool, in space-based gravitational wave detection missions, formation scale is one of the key factors determining the detection center frequency band. Based on the physical mapping relationship between gravitational wave detection frequency bands and formation scale (or arm length), the formation-scale design tool provides nominal arm length recommendations for different scientific objectives. This tool draws on the engineering design parameters of major international space-based gravitational wave detection programs (such as LISA, Taiji, and DECIGO) to construct a mapping model from detection frequency bands to nominal formation arm lengths. The input parameters of the formation-scale design tool are shown in Table 3.

[0105] Table 3

[0106]

[0107] When a user task involves designing a formation configuration for a specific detection frequency band, the scale design tool needs to be invoked. For example, when dealing with a task such as "The center frequency band to be detected is 0.5Hz, please design the formation arm length", this tool needs to be invoked, and "0.5Hz" needs to be entered in "frequency_type" to obtain the recommended nominal formation arm length.

[0108] Specifically, the formation configuration generation tool calculates the state vectors of the three spacecraft in the target formation, i.e., the formation configuration, based on the nominal arm length of the formation and the phase angles of the three spacecraft in the formation plane using the classical CW equations. The input parameters of the formation configuration generation tool are shown in Table 4.

[0109] Table 4

[0110]

[0111] This tool uses a domain-knowledge-based default parameter completion mechanism. For the phase angle parameter "phase_angle", if the user does not specify it, LLM will automatically set the phase angle to 0°, 120° and 240° according to the preset JSON Schema fragment ""default": [0, 120, 240]" in the tool parameter format.

[0112] This tool should be invoked when the user task involves generating a specific formation configuration. For example, when dealing with a task such as "The nominal arm length of the target formation is 3 million km, please design the formation configuration", this tool needs to be invoked, and "3000000" should be entered in "arm_length" and the default value "[0, 120, 240]" should be entered in "phase_angle" to obtain the required formation configuration, i.e., the state vector of each spacecraft.

[0113] Specifically, the pulse sequence calculation tool calculates the required pulse sequence based on the initial and final times of the orbital transfer and the initial and final states of the three spacecraft in the formation, using an optimized algorithm. The input parameters for the pulse sequence calculation tool are shown in Table 5.

[0114] Table 5

[0115]

[0116] It is worth noting that if the user does not specify the end time tf of the orbital transfer, LLM will automatically set the transfer time to 90 days based on the tool's preset fields, i.e., tf"=t0+90". Furthermore, to ensure that the "initial_state_vector" and "target_state_vector" provided by LLM conform to the tool's required format, this tool employs a multi-level nested array dimension constraint mechanism. Through the JSON Schema fragment "{"minItems": 3,"maxItems": 3,"Items":{"minItems": 6,"maxItems": 6}}", the dimensions of the input parameters are explicitly constrained, indicating that LLM should provide a nested array structure containing 3 spacecraft, each with a 6-dimensional state vector.

[0117] This tool is needed when the user task involves calculating the pulses required for orbit transfer. For example, when handling tasks such as "At the initial moment, three spacecraft are at the origin, and the target formation configuration is XX. Please calculate the pulse sequence required for orbit transfer," this tool needs to be called. Enter "0" in "t0", enter the default value "90" in "tf", and enter the state data that meets the above dimension and format requirements in "initial_state_vector" and "target_state_vector" to obtain the corresponding pulse sequence calculation results.

[0118] Specifically, for the trajectory plotting tool, addressing the visualization and analysis needs of spacecraft during orbital maneuvers, this tool provides users with numerical simulations and multiple visualizations of the transfer trajectories of three spacecraft based on the initial and final moments and initial state of the orbital transfer, as well as the pulse sequence applied during the transfer process. The input parameters for the trajectory plotting tool are shown in Table 6. This tool also uses the "minItems" and "maxItems" fields to constrain the dimensions of the input parameters.

[0119] Table 6

[0120]

[0121] This tool should be used when user tasks involve visual analysis and result verification of the orbit transfer process. For example, when dealing with tasks such as "The initial time of the orbit transfer is 0, the duration is 60 days, the applied pulse sequence is pulse sequence XX1, and the initial and final states are XX2 and XX3 respectively. Please draw the transfer trajectories of the three spacecraft," the trajectory drawing tool needs to be enabled. Enter "0" in "t0," "60" in "tf," and enter data that meets the dimension and format requirements in "initial_state_vector," "pulse_time," and "delta_V" to generate the corresponding trajectory simulation results and graphical output.

[0122] Specifically, the formation stability analysis tool assesses the long-term stability of a formation under perturbation environments by calculating the maximum divergence of its geometric parameters during its on-orbit mission. The input parameters for the formation stability analysis tool are shown in Table 7.

[0123] Table 7

[0124]

[0125] This tool employs a task-oriented parameter polymorphism adaptation mechanism, supporting both divergence monitoring and threshold verification modes. Specifically, the tool's parameter format uses the JSON Schema fragment ""oneOf": [{"type": "number},{"const": "-"}]" to prompt the LLM to flexibly switch modes according to requirements. When the LLM enters "-" in the parameter representing the corresponding threshold indicator, the divergence monitoring mode is enabled, and the tool returns the maximum value of the divergence of each geometric parameter. When a specific indicator threshold is entered, the threshold verification mode is enabled. If the divergence of the formation's geometric parameters exceeds the stability indicator threshold, the tool returns the fault time and the corresponding abnormal configuration.

[0126] This tool is required when user tasks involve stability analysis and feasibility verification of formations. For example, when dealing with tasks such as "Can a formation with a nominal length of 3 million km meet the stability requirement of less than 1% change in arm length within two years after its construction?", this tool should be enabled. Enter "0" in "t0", "730" in "tf", "3000000" in "arm_length", the default value "[0, 120, 240]" in "phase_angle", the default value "-20" in "alpha", and "0.01" in "threshold_delta_L". Enter "-" in other parameters representing threshold indicators to generate the corresponding stability analysis report.

[0127] In step S11, the dynamic execution engine business layer obtains the current pending tasks, generates task decomposition prompts based on the pending tasks, tool information, task decomposition prompt templates, and preset manual task decomposition examples, and calls the task decomposition model to decompose and process the pending tasks; wherein, the tool information is obtained by the dynamic execution engine business layer from the toolset business layer.

[0128] Specifically, users can input the pending task requiring task planning through the terminal device's interface. The top-level dynamic execution engine business layer of the task planning framework receives this pending task and creates a corresponding task decomposition prompt generation task. When processing this task decomposition prompt generation task, task decomposition prompts can be generated based on the pending task, tool information, task decomposition prompt template, and preset manual task decomposition examples. Then, a call to the task decomposition model can be initiated to the large language model business layer, sending the pending task and task decomposition prompts to the dynamic execution engine business layer so that the task decomposition model can process the current decomposition task.

[0129] For example, the method further includes: when the dynamic execution engine business layer obtains a task to be processed, it obtains the currently available tools in the toolset from the toolset business layer to determine the tool information.

[0130] Specifically, before generating task decomposition prompts, the dynamic execution engine business layer can first read the availability status of each tool from the toolset business layer, and generate tool information for each tool based on the availability status.

[0131] For example, a complex task decomposition model breaks down the user-inputted original problem or task into a sequence of meta-tasks. Each meta-task represents a simple task that can be completed without tool calls or with only a single tool call. To achieve this task decomposition process, the prompts provided to the complex task decomposition model include the following information:

[0132] 1) Names and functions of each tool;

[0133] 2) Example of manual task breakdown;

[0134] 3) Standard prompt template to enable LLM (Large Language Model) to decompose tasks: Please decompose the user input question into a sequence of meta-tasks based on {tool information}, where each meta-task is a simple task that cannot be further decomposed. To help you with task decomposition, I have provided an example: {Manual Task Decomposition Example}.

[0135] In step S12, the large language model business layer uses the task decomposition model and the task decomposition prompt words to decompose the task to be processed and obtain the corresponding meta-task sequence; and returns the meta-task sequence to the dynamic execution engine business layer; wherein, the meta-task sequence includes several sequentially executed meta-tasks; the meta-tasks are tasks that can be completed without calling tools or with a single call to tools.

[0136] For example, after receiving the pending task and task decomposition prompts from the dynamic execution engine business layer, the large language model business layer can create the current decomposition task and call the task decomposition model to perform task decomposition processing based on the task decomposition prompts, obtaining the corresponding meta-task sequence. The meta-task sequence includes several sequentially executed meta-tasks; and each meta-task is a task that can be completed without calling tools or with only a single tool call. After generating the meta-task sequence, the large language model business layer can feed the meta-task sequence back to the dynamic execution engine business layer.

[0137] For example, the complex task decomposition model breaks down the user-inputted question or task into a sequence of meta-tasks Q. Each meta-task represents a simple task that can be completed without or with only a single tool call. For the space-based gravitational wave detection mission scenario, to implement this task decomposition process, the prompts in the complex task decomposition model need to be adaptively designed, as shown in Table 8.

[0138] Table 8

[0139]

[0140] In step S13, the dynamic execution engine business layer constructs task execution prompts based on the meta-task sequence, tool information, and tool call prompt template; and calls the task sequence execution model to process the task prompts.

[0141] For example, after obtaining the meta-task sequence, the dynamic execution engine business layer can construct task execution prompts based on the current meta-task sequence tool information and tool call prompt templates, and call the task sequence execution model to the large language model business layer so that the task sequence execution model can execute the meta-task sequence and obtain the task data of each meta-task.

[0142] For example, the prompts for the task sequence execution model include the following information:

[0143] 1) Detailed information about the tool, including the tool name, tool function, and the format of the input parameters required by the tool;

[0144] 2) The meta-task sequence returned by the complex task decomposition model;

[0145] 3) Enable LLM to have a standard prompt template for calling external tools to complete tasks: Please answer the user's input question {original question + meta-task sequence} based on {tool information}.

[0146] In addition, the format of the returned data can be pre-configured, including:

[0147] Thought: (You need to think about what you need to do now)

[0148] Action: (The action to be taken; one of the tools)

[0149] Action Input: (Input parameters required by the tool)

[0150] Observation: (The result after calling the tool)

[0151] ...

[0152] Thought: (You now know the final answer)

[0153] Final Answer: (Provide the final answer)

[0154] Configuration rules include: Thought / Action / Action Input / Observation can be repeated 0 or more times; when no suitable tool is available or the current task does not require calling a tool, the output should not include Action / ActionInput.

[0155] In step S14, the large language model business layer uses the task sequence execution model, combined with task execution prompts, to call the tools corresponding to the meta-tasks to the toolset business layer to complete the execution of the meta-tasks, and integrates the task data corresponding to each meta-task to generate task planning data; and returns the task planning data to the dynamic execution engine business layer.

[0156] For example, after receiving the task execution prompt from the dynamic execution engine business layer, the large language model business layer can create a sequence execution task, enable the task sequence execution model, and input the task execution prompt into the task sequence execution model.

[0157] The task sequence execution model, based on the decomposed meta-task sequence, provides the tools and input parameters required to complete the current meta-task. The program automatically parses the LLM's response, thereby invoking the actual tools. Once all meta-tasks are completed, the task sequence execution model integrates the tool outputs and generates the final task planning scheme.

[0158] For example, step S14 described above may include:

[0159] Step S31: The task sequence execution model executes each meta-task in sequence and identifies whether the current meta-task requires tool invocation.

[0160] Step S32: When it is determined that the current meta-task requires tool invocation, determine the corresponding tool name and tool input parameters; and invoke the corresponding tool to the tool set business layer according to the tool name, and process the tool input parameters to obtain the tool's feedback data.

[0161] For example, when the dynamic execution engine business layer calls the corresponding tool to the toolset business layer according to the tool name and processes the tool input parameters, the method further includes: the tool in the toolset business layer processes the tool input parameters and sends the feedback data to the task sequence execution model.

[0162] Specifically, for the task sequence execution model, when executing the meta-task sequence in order according to the task execution prompts, the currently executing meta-task can be judged in real time to determine whether a tool call is needed. If a tool call is required, the corresponding tool name and parameters are determined based on the meta-task information. The corresponding tool can then be directly called from the toolset business layer based on the tool name, and the tool input parameters are input into the corresponding tool to obtain its output data. (Reference) Figure 2 As shown, the tool's output data can be directly fed back to the task sequence execution model.

[0163] Alternatively, in some exemplary embodiments, the corresponding tool is invoked from the toolset business layer based on the tool name, and the tool input parameters are processed to obtain the tool's feedback data, including:

[0164] Step S41: Feed back the tool name and tool input parameters to the dynamic execution engine business layer;

[0165] Step S42: The dynamic execution engine business layer calls the corresponding tool to the toolset business layer according to the tool name, processes the tool input parameters, and obtains the feedback data from the tool to the dynamic execution engine business layer.

[0166] In step S4, the dynamic execution engine business layer sends the tool's feedback data to the large language model business layer so that the task sequence execution model can obtain the tool's feedback data.

[0167] Specifically, refer to Figure 3 As shown, for the task sequence execution model, when the execution of a meta-task requires calling a tool, the tool can be called from the dynamic execution engine business layer to the toolset business layer, and the feedback data generated by the tool can be fed back to the task sequence execution model through the dynamic execution engine business layer.

[0168] For example, when the task sequence execution model executes each meta-task in sequence and identifies whether the current meta-task requires tool invocation, the method further includes: the task sequence execution model returns the tool and tool input parameters corresponding to the completion of the current meta-task based on the meta-task sequence and the actual execution results of the preceding meta-tasks.

[0169] For example, the task sequence execution model returns the tools and tool input parameters required to complete the current meta-task in JSON format, based on the decomposed meta-task sequence Q and the actual execution results R of the preceding meta-tasks. (See reference) Figure 3 As shown, the dynamic execution engine, i.e., the main program, automatically parses the LLM's response to complete the actual tool call. Once all meta-tasks are completed, the task sequence execution model integration tool outputs the results and generates the final task planning scheme. For the space-based gravitational wave detection mission scenario, to achieve this task execution process, the prompts in the task sequence execution model need to be adaptively designed, as shown in Table 9.

[0170] Table 9

[0171]

[0172] For example, QWEN3-4b is used as a complex task decomposition model and QWEN3-32b is used as a task sequence execution model. The prompt words for task decomposition and execution are constructed according to Table 8 and Table 9, respectively.

[0173] For example, the original task to be processed by the user input may include: At the initial moment, three spacecraft are located in the center of the formation. Please design a 6-year gravitational wave detection mission plan with a target frequency of approximately 0.5 Hz.

[0174] By calling the task decomposition model of the large language model's business layer and inputting the task decomposition prompts corresponding to the task to be processed into the task decomposition model, the corresponding task decomposition results can be obtained, including:

[0175] Complex issue: Yes;

[0176] Subtask: Obtain formation scale using satellite formation scale design tools;

[0177] Subtask: Use satellite formation configuration design tools to obtain the nominal formation status;

[0178] Subtask: Uses a pulse sequence calculation tool to obtain the pulse sequence required to reach the nominal state from the initial state;

[0179] Subtask: Use knowledge graph tools to query reference values ​​for formation stability metrics;

[0180] Subtask: Use formation stability analysis tools to determine whether the formation meets the reference indicators within 6 years after its construction is completed;

[0181] Subtask: (If the criteria are not met) Use the pulse sequence calculation tool to find the pulse sequence required for the current abnormal state to reach the nominal state;

[0182] Subtask: Use the formation stability analysis tool to determine whether the correction meets the reference indicators;

[0183] Based on the meta-task sequence, tool information, and tool call prompt template, construct task execution prompts, which may include:

[0184] The final mission involves three spacecraft initially positioned at the center of the formation. Please design a 6-year gravitational wave detection mission, targeting a frequency of approximately 0.5 Hz. The execution sequence is:

[0185] 1. Obtain the formation scale using a satellite formation scale design tool;

[0186] 2. Use satellite formation configuration design tools to obtain the nominal formation status;

[0187] 3. Use a pulse sequence calculation tool to obtain the pulse sequence required to reach the nominal state from the initial state;

[0188] 4. Use knowledge graph tools to query reference thresholds for formation stability indicators;

[0189] 5. Use formation stability analysis tools to determine whether the formation meets the reference indicators within 6 years after its construction is completed;

[0190] 6. (If the criteria are not met) Use a pulse sequence calculation tool to obtain the pulse sequence required for the current abnormal state to reach the nominal state;

[0191] 7. Use formation stability analysis tools to determine whether the corrections meet the reference indicators. The current task is to use satellite formation scale design tools to obtain the formation scale.

[0192] Action: scale designer

[0193] Action Input: {"frequency band": "0.1-1Hz"}

[0194] Observation:XX

[0195] Based on the results of previous missions, the current task is to use satellite formation configuration design tools to determine the nominal formation status.

[0196] Thought: Based on the scale design results, the configuration design tool needs to be called to generate the nominal state vector of the equilateral triangle formation. Input arm length 3,000,000km and default phase angle [0°, 120°, 240°].

[0197] Action: Configuration Designer

[0198] Action Input:{"arm length":3000000,"phase angle":[0.120.240]}

[0199] Observation: XX

[0200] ...

[0201] Based on the results of the previous task, the current task is to output the final result.

[0202] Final Answer: XX

[0203] The method provided in this invention constructs a three-layer collaborative architecture. The bottom layer, the toolset business layer, defines the functions and specific implementation methods of each tool. The middle layer, the large language model business layer, uses a complex task decomposition model to break down user instructions into meta-task sequences. Through a task sequence execution model, the toolset is called sequentially to execute the meta-tasks. This method provides a novel technical path and solution for spacecraft formation flight mission planning. Compared to traditional mission planning methods, which suffer from insufficient adaptability to new scenarios, excessively high requirements for the structure of input data, difficulty in real-time user interaction, and poor interpretability of results, the core advantages of this invention are as follows: First, it lowers the professional threshold for users through natural language interaction, allowing non-expert users to dynamically define and adjust task requirements through multi-turn dialogues; second, with the help of prompting strategies such as Chain-of-Thought (CoT), LLM can explicitly present the reasoning process, significantly improving the execution transparency of complex tasks; third, the "world knowledge" internalized by LLM and its cross-domain generalization capabilities support autonomous adaptation to unexpected scenarios. Therefore, this invention provides a more efficient, flexible, and scenario-generalizable intelligent solution for spacecraft formation flight mission planning.

[0204] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may, for example, be executed synchronously or asynchronously in multiple modules.

[0205] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0206] Figure 4 A schematic diagram of an electronic device suitable for implementing embodiments of the present invention is shown.

[0207] It should be noted that, Figure 4 The electronic device 1000 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0208] like Figure 4 As shown, the electronic device 1000 includes a Central Processing Unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 1002 or programs loaded from storage section 1008 into Random Access Memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An Input / Output (I / O) interface 1005 is also connected to the bus 1004. Furthermore, the electronic device 1000 also includes an FPGA device and a System-on-a-Chip (SoC) device.

[0209] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.

[0210] For example, the aforementioned electronic device could be a host computer.

[0211] In particular, according to embodiments of the present invention, the processes described below with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit (CPU) 1001, it performs various functions defined in the system of this application.

[0212] Specifically, the aforementioned electronic devices can be airborne intelligent electronic devices, such as airborne video processing equipment.

[0213] It should be noted that the storage medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein computer-readable program code is carried. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0214] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0215] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0216] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The aforementioned storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments. For example, the electronic device may perform... Figure 1 The steps of the method shown.

[0217] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0218] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0219] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0220] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A spacecraft formation flight mission planning method based on large language model and tool set cooperation, characterized in that, The method comprises: pre-constructing a task planning architecture; wherein the task planning architecture comprises a tool set service layer, a large language model service layer and a dynamic execution engine service layer configured in sequence; the tool set service layer integrates a plurality of tools for processing meta-tasks and feeds back task data to the large language model service layer and / or the dynamic execution engine service layer; the large language model service layer is used for analyzing natural language instructions and generating a meta-task sequence; and the dynamic execution engine service layer is used for obtaining a task to be processed and coordinating real-time interaction between the tool set service layer and the large language model service layer; configuring a standardized tool interface for the tool set service layer, including a tool name, a tool function description and a parameter format; wherein the parameter format comprises a parameter name, a parameter description and a parameter data format followed by each parameter node; configuring a multi-dimensional parameter selection mechanism for the tool set service layer, including a multi-layer nested array dimension constraint mechanism for constraining a high-dimensional state vector, a finite value domain constraint mechanism for explicitly defining a legal input value domain, a default parameter completion mechanism based on domain knowledge and a parameter polymorphic adaptation mechanism for realizing parameter polymorphic adaptation; the dynamic execution engine service layer obtains a current task to be processed, generates a task decomposition prompt word according to the task to be processed, tool information, a task decomposition prompt template and a preset artificial task decomposition example, and calls a task decomposition model to decompose and process the task to be processed; wherein the tool information is obtained by the dynamic execution engine service layer from the tool set service layer; the large language model service layer uses the task decomposition model to decompose the task to be processed in combination with the task decomposition prompt word, obtains a corresponding meta-task sequence, and returns the meta-task sequence to the dynamic execution engine service layer; wherein the meta-task sequence comprises a plurality of sequentially executed meta-tasks; and the meta-task is a task that can be completed without calling a tool or by calling a tool once; the dynamic execution engine service layer constructs a task execution prompt word according to the meta-task sequence, tool information and a tool calling prompt template, and calls a task sequence execution model to process the task execution prompt word; the large language model service layer uses the task sequence execution model to call a tool corresponding to a meta-task from the tool set service layer in combination with the task execution prompt word to complete the execution of the meta-task, integrates task data corresponding to each meta-task to generate task planning data, and returns the task planning data to the dynamic execution engine service layer; wherein when the execution of the meta-task requires calling a tool, the tool is called from the tool set service layer by the dynamic execution engine service layer, and feedback data generated by the tool is fed back to the task sequence execution model by the dynamic execution engine service layer.

2. The method of claim 1, wherein, The method further comprises: when the dynamic execution engine service layer obtains a task to be processed, obtaining available tools in the tool set from the tool set service layer to determine the tool information.

3. The method of claim 1, wherein, the large language model service layer uses the task sequence execution model to call a tool corresponding to a meta-task from the tool set service layer in combination with the task execution prompt word to complete the execution of the meta-task, comprising: The task sequence execution model executes each subtask in sequence and identifies whether the current subtask requires a tool call; When it is determined that the current subtask requires a tool call, the corresponding tool name and tool input parameter are determined, and the corresponding tool is called from the tool set business layer according to the tool name, and the tool input parameter is processed to obtain feedback data of the tool.

4. The method of claim 3, wherein, According to the tool name, the tool set business layer is called to obtain the feedback data of the tool, which includes: The tool name and tool input parameter are fed back to the dynamic execution engine business layer by the task sequence execution model; The dynamic execution engine business layer calls the corresponding tool from the tool set business layer according to the tool name, and processes the tool input parameter to obtain the feedback data of the tool to the dynamic execution engine business layer; The dynamic execution engine business layer sends the feedback data of the tool to the large language model business layer for the task sequence execution model to obtain the feedback data of the tool.

5. The method of claim 4, wherein, When the dynamic execution engine business layer calls the corresponding tool from the tool set business layer according to the tool name and processes the tool input parameter, the method further includes: The tool set business layer processes the tool input parameter and sends the feedback data to the task sequence execution model.

6. The method of claim 3, wherein, When the task sequence execution model executes each subtask in sequence and identifies whether the current subtask requires a tool call, the method further includes: The task sequence execution model returns the tool corresponding to the completion of the current subtask and the tool input parameter according to the subtask sequence and the actual execution result of the previous subtask.

7. The method of claim 1, wherein, The tool set business layer includes: A network search tool for accessing the Internet and searching for data; A knowledge graph tool for querying a local knowledge base; A formation scale design tool for recommending a formation arm length according to a gravitational wave frequency band to be detected; A formation configuration generation tool for generating a formation configuration according to the formation arm length and a spacecraft phase angle; A pulse sequence calculation tool for solving a pulse sequence required for orbit transfer; A trajectory drawing tool for drawing a transfer trajectory of a spacecraft under the action of a pulse; A stability analysis tool for evaluating the long-term stability of the formation under perturbation environment.

8. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the spacecraft formation flight task planning method of the large language model and the tool set cooperation according to any one of claims 1 to 7.

9. An electronic device, comprising: It includes: A processor; And A memory for storing executable instructions of the processor; Wherein, the processor is configured to implement the spacecraft formation flight task planning method of the large language model and the tool set cooperation according to any one of claims 1 to 7 by executing the executable instructions.

Images