Large model application platform oriented to aircraft task intellectualization and construction method thereof

Abstract

The invention discloses a large model application platform construction method oriented to aircraft task intellectualization. The construction method of the platform comprises the following steps: deploying a large language model for local quantification of an aircraft orbital motion calculation vertical field task as a core processing engine; an aircraft orbital motion calculation type task planning workflow constructed based on an open source framework is established, and a complex orbital motion calculation type task is decomposed into fine grit; building an aircraft orbit task real-time correction workflow for monitoring a task execution state and dynamically adjusting a plan; a set of model context protocol tool set required by aircraft orbital motion calculation is assembled, calculation, model, perception and action tools are integrated, and the capability boundary of a special large model is greatly expanded. Besides, an aircraft orbit calculation type task execution specification manual knowledge base and a retrieval auxiliary workflow thereof are also constructed, and an accurate vertical field specification basis is provided for task execution through vectorization storage and semantic retrieval.

Description

Technical Field

[0001] This invention relates to the field of large language model deployment technology, and in particular to a large model application platform for intelligent aircraft missions and its construction method. Background Technology

[0002] Against the backdrop of accelerated global digital transformation, artificial intelligence (AI) technology has become a core engine driving innovation across various industries. With the exponential growth of data volume, significant improvements in computing power, and continuous breakthroughs in algorithm theory, AI technology has gradually evolved from an early rule-driven model to a data-driven intelligent paradigm, with its applications permeating numerous fields such as finance, healthcare, manufacturing, and transportation. Among these, machine learning and deep learning, as the core algorithmic support of AI, provide crucial capabilities for processing complex data and extracting features; large language models, building upon this foundation, have achieved leapfrog development in natural language understanding and generation, propelling human-computer interaction into a new stage; intelligent agent technology, by integrating multi-dimensional capabilities, has achieved autonomous decision-making and collaborative execution; and model context protocols and retrieval enhancement generation technologies further address key issues such as contextual consistency and factual accuracy in model applications. Together, these technologies form the core technical system for current AI applications, providing comprehensive technical support for the implementation of various intelligent applications.

[0003] Machine learning is a core branch of artificial intelligence. Essentially, it uses algorithms to enable computers to automatically learn patterns from data and then use these patterns to predict or make decisions about unknown data. Unlike traditional manual programming, machine learning systems do not rely on explicit manual rule definitions. Instead, they automatically optimize model parameters to fit the data distribution through training on large amounts of labeled or unlabeled data. Based on different learning methods, machine learning can be divided into several main types: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning trains models using labeled sample data and is suitable for prediction tasks such as classification and regression. Unsupervised learning mines potential structural features from unlabeled data and is often used in clustering and dimensionality reduction. Reinforcement learning optimizes decision-making strategies by obtaining reward signals through trial and error between the agent and the environment. Currently, machine learning technology is widely used in image recognition, speech recognition, recommendation systems, risk assessment, and other fields, becoming a fundamental tool for intelligent data processing and providing underlying algorithmic support for more advanced intelligent technologies such as deep learning and large language models.

[0004] Deep learning, a crucial branch of machine learning, utilizes deep neural networks as its core architecture to efficiently process complex, high-dimensional data and extract abstract features. Compared to traditional machine learning algorithms that rely on manually designed features, deep learning, through its multi-layered neural network structure consisting of input, hidden, and output layers, can automatically transform raw data into high-level semantic features. Key technological breakthroughs include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and the Transformer architecture: CNNs, through local receptive fields and weight sharing mechanisms, demonstrate exceptional image feature extraction capabilities in computer vision tasks; RNNs and their variants (such as LSTM and GRU) excel at processing sequential data and are suitable for scenarios like text generation and sentiment analysis in natural language processing; the Transformer architecture, based on self-attention mechanisms, effectively captures long-distance dependencies in data, becoming the core foundation of current large-scale language models. With increased computing power and training data volume, the depth and complexity of deep learning models have continuously improved, achieving breakthroughs in image recognition, natural language processing, autonomous driving, and drug development, propelling artificial intelligence technology into the practical application stage.

[0005] Large Language Models (LLMs) are large-scale pre-trained language models built on deep learning techniques for natural language processing tasks. Their core characteristic lies in mastering the grammatical rules, semantic logic, and world knowledge of human language through pre-training on massive amounts of text data, thereby achieving high-quality natural language understanding and generation. LLMs typically use the Transformer architecture as their basic network structure, pre-training with billions or even trillions of model parameters and massive amounts of general-purpose text data (such as books, web pages, and papers) to form a foundation for generalized language models. During the pre-training phase, the model learns language rules through self-supervised learning tasks (such as masked language modeling and next-sentence prediction). In the application phase, fine-tuning or prompt engineering are used to adapt the model to specific downstream tasks, such as text summarization, machine translation, question answering systems, and code generation. In recent years, large language models, represented by the GPT series, BERT, and LLaMA, have made significant breakthroughs in natural language processing capabilities, driving human-computer interaction towards a more natural and intelligent direction. However, they also face challenges such as limited context windows, factual errors, and insufficient reasoning capabilities.

[0006] An intelligent agent is an intelligent entity with autonomous perception, decision-making, and execution capabilities. It can autonomously achieve specific goals through interaction with its environment and other intelligent agents. Intelligent agent technology integrates technologies from multiple fields, including machine learning, deep learning, natural language processing, and planning algorithms. Its core characteristics include autonomy, responsiveness, sociality, and goal orientation: autonomy means the agent can independently execute tasks without human intervention; responsiveness means it can perceive environmental changes in real time and respond accordingly; sociality means multiple agents can complete complex tasks through collaboration and negotiation; and goal orientation ensures that the agent's behavior always revolves around a preset goal. Depending on the application scenario, intelligent agents can be categorized into intelligent assistants, industrial robots, autonomous driving agents, and virtual digital humans. In practical applications, intelligent agents typically need to integrate large language models to achieve natural language interaction capabilities, ensuring semantic coherence during interaction through model context protocols, thereby enabling the autonomous execution of complex tasks in fields such as customer service, education, and industrial control.

[0007] Model context protocols (MFRs) are a set of standards designed to ensure the effective transmission of contextual information, semantic consistency, and state synchronization of intelligent models during interaction or collaboration. In scenarios such as agent interaction and multi-model collaborative processing of complex tasks, models need to make decisions or generate responses based on contextual data such as historical interaction information and task states. If the transmission of contextual information is not standardized, it can easily lead to semantic ambiguity and decision-making bias. MFRs typically cover core aspects such as the definition standards of contextual information (e.g., format specifications for data such as interaction history, task parameters, and environmental states), transmission mechanisms (e.g., implementation methods for real-time synchronization and asynchronous caching), consistency verification rules (e.g., semantic conflict detection and resolution methods), and lifecycle management (e.g., the timing of context creation, updating, and destruction). Currently, with the rise of applications such as multi-agent collaboration and continuous dialogue systems, the importance of MFRs is increasingly prominent. They not only improve the stability and accuracy of model interaction but also provide a unified interface standard for collaboration between models with different architectures, driving the evolution of intelligent systems from single-model capabilities to multi-model collaborative capabilities.

[0008] Retrieval Augmentation (RAG) is a technical framework that integrates external knowledge base retrieval and generation models, aiming to address the shortcomings of traditional large language models in terms of factual accuracy, timeliness, and professionalism. Traditional large language models derive their knowledge from pre-training data, which suffers from problems such as knowledge lag (inability to cover the latest information), factual bias (learning errors from pre-training data), and insufficient professional knowledge (difficulty in penetrating specific domains). RAG optimizes this through a two-stage "retrieval-generation" process: In the retrieval stage, based on the user's query, relevant factual information is retrieved from external knowledge bases (such as academic papers, industry reports, and real-time news) using tools like search engines and vector databases; in the generation stage, the generation model uses the retrieved information as contextual input, combined with its own language generation capabilities, to generate response content that is both logically sound and factually accurate. Furthermore, RAG can further improve the reliability and relevance of the output content by optimizing retrieval strategies (such as semantic retrieval and multi-turn retrieval) and generation alignment mechanisms (such as information fusion and citation annotation). Currently, this technology has been widely applied in fields such as intelligent question answering, professional consulting, and content creation, becoming one of the key technologies for enhancing the practical value of large language models. Summary of the Invention

[0009] Therefore, embodiments of the present invention provide a large-scale model application platform for intelligent aircraft missions and a method for constructing it.

[0010] According to a first aspect of the embodiments of this application, a method for constructing a large-scale model application platform for intelligent aircraft missions is provided, comprising: S1: Quantize the general domain large model, accelerate its deployment to the local machine through inference, configure the local API call interface, and obtain the local quantized large model; S2: By writing multi-stage prompt word templates, maintaining the model call context, and assembling multiple large model calls in sequence, selection, and loop structure, a dedicated workflow for two sets of large model execution tasks is formed for aircraft path pre-planning and real-time path calibration. S3: Using three steps, namely tool encapsulation, protocol packaging, and framework assembly, the input and output controllable functions are configured into tools usable by large models, and then assembled into a dedicated toolset for spacecraft orbit calculation. S4: Through document extraction, adaptive segmentation, segmented embedding, and semantic retrieval, configure the specification knowledge base that the large model needs to refer to when executing spacecraft tasks, forming a task execution manual knowledge base for storing vertical domain specifications for spacecraft orbital motion calculation and a vertical domain knowledge base retrieval auxiliary workflow for assisting large model decision-making; S5: Based on the local quantization model, deploy the dedicated workflow, aircraft-specific toolset, and standardized knowledge base to form a large model application platform for intelligently executing aircraft missions.

[0011] According to a second aspect of the embodiments of this application, a large model application platform for intelligent aircraft missions is provided, which is constructed by the large model application platform construction method for intelligent aircraft missions described in the first aspect.

[0012] The beneficial effects of this invention are as follows: This application adopts local quantization deployment of large models, which overcomes the problem that large model inference relies on network transmission or large-scale computing power, thereby realizing lightweight operation and offline inference of large models.

[0013] This application adopts a dedicated workflow for large models, which overcomes the problem that large models cannot complete or even experience illusions when a single prompt word template contains too many or too large tasks. This enables large models to execute planning steps in stages, and ultimately complete the complex spacecraft orbit calculation task.

[0014] This application employs a dedicated toolset for aircraft orbit calculation, overcoming the problem that large models cannot directly interact with simulated aircraft mission scenarios, thereby enabling the simulation, visualization, and quantitative evaluation of the entire mission execution process.

[0015] This application employs a knowledge base for aircraft mission execution manuals and a knowledge base for aircraft vertical domains to assist in the retrieval workflow. This overcomes the lack of domain knowledge in large models when performing missions in the aircraft vertical domain, as well as the resulting illusion problem, thereby achieving high interpretability and traceability of the mission execution process.

[0016] This application designs an integrated large model application platform, which overcomes the problem of large models lacking various infrastructure support when performing complex tasks, and thus realizes the mutual cooperation of infrastructure such as workflow, toolset, and knowledge base when large models perform tasks in the vertical domain of aircraft. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for constructing a large-scale model application platform for intelligent aircraft missions, according to an exemplary embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the basic call flow of a large model combined with prompt word project, according to an exemplary embodiment of this application.

[0019] Figure 3 This is a flowchart illustrating a spacecraft path pre-planning workflow according to an exemplary embodiment of this application.

[0020] Figure 4 This is a flowchart illustrating a real-time spacecraft path calibration workflow according to an exemplary embodiment of this application.

[0021] Figure 5 This is a schematic diagram illustrating the composition of a Model Context Protocol (MCP) toolset according to an exemplary embodiment of this application.

[0022] Figure 6 This is a flowchart illustrating the construction and invocation of a space mission specification knowledge base according to an exemplary embodiment of this application.

[0023] Figure 7 This is an overall architecture diagram of a large-scale model application platform for intelligent implementation of complex tasks in a vertical domain, as illustrated in an exemplary embodiment of this application. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the specific content of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention. Contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0025] Figure 1 This is a flowchart illustrating a method for constructing a large-scale model application platform for intelligent aircraft missions, according to an exemplary embodiment of this application. (Reference) Figure 1 The method for constructing a large-scale model application platform for intelligent aircraft missions may include: S1: Quantize the general-domain large model, accelerate its deployment locally via inference, configure the local API call interface, and obtain the locally quantized large model; this step includes the following sub-steps: S11: Select a general-purpose large language model that meets the task requirements of the vertical domain of orbital motion calculation in terms of parameter scale and supported modes, and perform precision quantization on the selected large model; Specifically, considering the task scale, a general-purpose large language model (such as Llama 2-13B) with parameter sizes of 7B-32B and supporting both numerical computation and engineering formula inference modalities is selected, and the INT8 quantization scheme is used for local quantization deployment. This design addresses the specific requirements of orbital motion calculations for model numerical accuracy and modal adaptability. Quantization processing reduces hardware deployment resource consumption, and domain data calibration avoids accuracy loss of general-purpose models in vertical domains, achieving an optimal balance between computational accuracy and deployment efficiency.

[0026] S12: Based on the characteristics of vertical domain data, construct a dataset in a predetermined format to fine-tune the quantized large model; use the Ollam large model local deployment framework to deploy the fine-tuned large model on a local server or terminal device. Specifically, considering the structured characteristics of data in the field of orbital motion computation, the LoRA low-rank adaptation algorithm is used to fine-tune the quantized model, freezing the parameters of the pre-trained layers and training only the domain-specific adaptation layers. The fine-tuned model is then deployed to a local server using the Ollam framework. Adapting the dataset to the characteristics of the vertical domain improves the model's domain adaptability, while LoRA fine-tuning balances training efficiency and model performance. Local deployment of Ollam enables low-latency inference, meeting the real-time requirements of orbital computation, and overall optimizing the model's practicality and deployment flexibility.

[0027] S13: By calling the application API provided by the Ollam framework to call the large model that is deployed locally and is suitable for the vertical field of orbit computing, a large model with local quantization can be obtained. Specifically, a POST request is constructed on the local server terminal using the Ollam framework's built-in RESTful API. After sending the request, a JSON response is received from the local quantization model, completing the model call. The API call simplifies the engineering integration process and adapts to the parameter input requirements of orbit calculation scenarios; the lightweight nature of the quantization model improves response speed, meeting the real-time requirements of orbit calculation and engineering practicality.

[0028] like Figure 2 As shown in the figure, it is a basic call flow diagram of a large model combined with the prompt word project. The large model receives the prompt word input and outputs the result.

[0029] S2: By writing multi-stage prompt word templates and maintaining the model call context, multiple large model calls are assembled in a sequential, selection, and loop structure to form two dedicated workflows for the large model execution tasks: aircraft path pre-planning and real-time path calibration. This step includes the following sub-steps: S21: Write a dedicated prompt word template to receive the user's original input and combine it into prompt words that conform to the input specifications of the large model according to a predetermined format; call the large model through the API to obtain the original results generated by the model, and after parsing the returned results, store them in the model call context for maintenance, forming the basic call workflow of the large model; Specifically, a three-layer prompt word template is designed: the first layer is the parameter input area, the second layer is the task instruction area, and the third layer is the output constraint area. After the user inputs raw data, the system automatically fills in the template to generate standardized prompt words, which are then sent to the local model via Ollam's API. After receiving the returned raw results, the system parses and extracts the values, units, and basis, and stores the input parameters, task type, and output results in the context database, forming a closed-loop workflow of "input-processing-output-storage". This design unifies input standards through templates to avoid model ambiguity; ensures structured results through parsing, facilitating engineering applications; and utilizes historical data for context maintenance to improve computational consistency, reduce repetitive operations, and overall improve the professionalism and efficiency of orbit calculation.

[0030] S22: The first assembly involves multiple large model calls, including: Phase 1: Using prompt template 1, a single complex task is broken down into multiple simple subtasks; Phase 2: Using prompt template 2, steps are planned for each subtask; Phase 3: Using prompt template 3, the planned subtask steps are reviewed to determine if they meet the task objectives and specifications. If not, the process returns to Phase 2 for replanning; Phase 4: Using prompt template 4, the steps of all subtasks are integrated to output a complete, fine-grained task execution planning sequence, forming a dedicated workflow for pre-planning the aircraft path. Specifically, the first stage uses template one (containing "mission objectives + decomposition rules") to decompose the "transfer path planning task from low Earth orbit to Mars parking orbit" into five sub-tasks, including "Hohmann transfer calculation for Earth-Mars transfer orbit" and "design of mid-course correction maneuver parameters." The second stage uses template two (containing "sub-task objectives + step elements") to plan three steps for the "Hohmann transfer calculation": "determining the orbit's aphelion and perihelion—calculating the transfer time window—solving for the velocity increment Δv." The third stage uses template three (containing "review criteria: mission objective matching degree, engineering specification compliance degree") to check whether the steps meet the requirement of propellant reserve ≥ 5%; if not, the plan is returned for replanning. The fourth stage uses template four (containing "timing integration rules + output format") to integrate all steps according to the flight timing, outputting an execution sequence containing the trigger conditions and calculation parameters for each step. This design, by decomposing complex tasks in stages, ensures the standardization of each link through templates, reduces errors through a review mechanism, and integrates to form an executable plan, improves the accuracy and engineering feasibility of the spacecraft's path pre-planning.

[0031] like Figure 3As shown in the figure, this is an example of a pre-planning workflow for spacecraft path planning supported by the platform, taking a spacecraft path planning task as an example. This workflow demonstrates the pre-planning and execution process when this platform is applied to the complex task of spacecraft path planning in the vertical field. The workflow begins by importing the spacecraft's initial state vector and target state vector into the system and converting them into instructions understandable by the Large Language Model (LLM) through prompt word templates. The LLM first receives the prompt words and breaks down the entire task from the starting point to the end point into stages, decomposing it into multiple consecutive sub-segments or "new segment plans". Each new segment plan is submitted to an agent for processing, which is responsible for executing the path planning; during this process, the agent can autonomously call computational tools (such as an orbital dynamics equation calculator) to accurately calculate the feasibility of the path. The agent evaluates each segment plan, searches for paths, and decides whether to accept the segment plan; if accepted, it submits and records the stage result and then proceeds to the next segment plan; if not accepted, it feeds the result back to the LLM, which modifies the plan based on the feedback. Once all sub-plans have been accepted and completed by the agent, LLM performs plan integration, smoothly connecting all stages into a complete and coherent global path plan. Finally, this integrated plan is delivered and the tool is invoked to execute the plan, thereby completing a fully automated and intelligent process from task input to instruction output.

[0032] S23: The second assembly involves multiple large model calls, including: configuring the environment awareness module and periodically collecting task execution environment status information; the first stage analysis uses prompt word template one to analyze the status information and determine whether the task execution is normal; the second stage analysis uses prompt word template two to analyze possible causes of abnormalities, locate the task steps that need to be adjusted, and generate modification suggestions; the review stage uses prompt word template three to review the modification suggestions. If the review is approved, the modifications are executed and the task plan is updated. If the review is not approved, the process returns to the second stage to regenerate suggestions, forming a dedicated workflow for real-time path calibration of the aircraft. Specifically, an environmental perception module consisting of a simulated star sensor and millimeter-wave radar is configured to collect real-time position, velocity, and space environment data of the spacecraft every 10 seconds. In the first phase, using prompt template one (containing "actual state + planned threshold"), the deviation between the actual velocity and the planned value is compared to see if it exceeds 3%, determining whether the mission execution is normal. In the second phase, using template two (containing "abnormal characteristics + possible cause library"), if the velocity deviation exceeds 5%, it is analyzed as an atmospheric drag anomaly, indicating that the "orbital maintenance maneuver steps" need adjustment, generating a modification suggestion of "increasing thrust by 10%". In the review phase, using template three (containing "modification feasibility: propellant reserve ≥8%, attitude stability ≤0.1° / s"), the compliance of the modification suggestion is verified. If approved, the mission plan is updated; otherwise, it is returned for regeneration. This design achieves dynamic response through real-time environmental perception, accurately locates anomalies through phased analysis, and ensures the safety of modifications through a review mechanism, forming a closed-loop calibration flow and improving the reliability and engineering adaptability of real-time spacecraft path adjustments.

[0033] like Figure 4 As shown in the figure, this is a calibration workflow for complex tasks in a vertical domain supported by the platform, taking an aircraft path planning task as an example. This workflow is initiated periodically. The agent first obtains real-time monitoring data by calling tool modules such as sensor interfaces. This data is compared with the expected state of the pre-planned path to form environmental data. Subsequently, through prompt word integration (prompt word 5), the current state, expected state, and deviation information are organized into a structured description and submitted to the Large Language Model (LLM) for evaluation. The core issue is determining whether the current state is normal. If the LLM determines the state is normal and no path deviation or potential risk is detected, the workflow is paused, the system enters a waiting state, and monitors again after one cycle. If the LLM determines the state is abnormal, the calibration process begins. The LLM first accesses the memory module to retrieve relevant coping strategies or constraints. The agent comprehensively analyzes the retrieved information and the current anomaly, and forms a decision basis through prompt word integration (step 5). Subsequently, through prompt word input (prompt word 6), the agent requests the LLM to analyze the root cause of the anomaly and generate correction instructions based on the latest comprehensive information. The correction instructions output by the LLM are received by the agent, which then calls the corresponding tools to execute the instructions. The results of the tool's execution are ultimately used to update the path planning, generating a new path planning scheme, which is then immediately applied to the spacecraft's subsequent navigation and control, thus forming a closed loop of perception-decision-execution-feedback to ensure the mission's high reliability and adaptability in uncertain environments.

[0034] S3: This step involves three steps: tool encapsulation, protocol packaging, and framework assembly. It configures input-output controllable functions into tools usable by large models, and then assembles them into a dedicated toolset for spacecraft orbit calculation. This step includes the following sub-steps: S31: Construct a subset of model tools, wherein the subset of model tools is configured to include small reinforcement learning models for controlling the aircraft's perception and motion; this step includes the following sub-steps: S311: Configuration interface definition unit, used to fix the input and output data format of specialized small models in the field of reinforcement learning aircraft control; Specifically, the interface definition unit first standardizes the data structure. Then, for the reinforcement learning aircraft attitude control mini-model, the input format is fixed as {perceived state: {attitude angle, angular velocity}, control target: {desired attitude angle}}, and the output format is fixed as {control command: {servo angle, execution time}}. Finally, the legality of the input and output is verified through JSON Schema. This design achieves seamless integration between the mini-model and the toolset through standardized formats, reducing integration and adaptation costs; it also ensures the accuracy of reinforcement learning control commands by specifying data precision, enhancing the versatility and reliability of the model toolset subset.

[0035] S312: Configuration tool encapsulation unit, based on the Model Context Protocol (MCP) toolkit framework, encapsulates a small model conforming to the input and output format into a callable tool function; Specifically, the tool encapsulation unit, based on the Model Context Protocol (MCP) toolset framework, first registers the small reinforcement learning model for aircraft attitude control to the framework's model management module and assigns it a unique tool identifier. Then, it defines the tool function structure according to the MCP protocol, including the function name, input parameter list, and output parameter parsing rules. Finally, it encapsulates the model calling logic, automating the input data preprocessing, model inference triggering, and output command formatting processes, forming a tool function that can be directly called by larger models. This design, based on the MCP framework, ensures the standardization of small model encapsulation and compatibility with the overall toolset architecture; encapsulation into tool functions simplifies the calling process, improves the integration efficiency of small models, and the standardized processing logic ensures the accuracy of control command generation, enhancing the collaborative reliability of aircraft perception and motion control.

[0036] S313: Configure prompt word description unit to write corresponding tool function prompt words for each encapsulated model tool, which are used to describe its function when making decisions in large models; Specifically, the prompt description unit is written according to the structure of "tool identifier + core function + input / output description + applicable scenario". For example, the prompt for a tool is "Tool ID: xxx, Function: Based on reinforcement learning algorithm, generate servo angle and thrust control commands according to the real-time attitude perception data and desired attitude target of the aircraft; Input: attitude angle, angular velocity, desired attitude parameters; Output: -30°~30° servo control command, 0-500N thrust command; Applicable scenario: stable control when the aircraft attitude deviation exceeds ±0.5°". Similarly, a corresponding function description is written for the trajectory correction reinforcement learning tool. This design allows large models to quickly and accurately identify the purpose of the tool and the applicable scenario, avoiding misjudgment during calls; the standardized prompt word structure improves decision-making efficiency, ensures that the reinforcement learning tool and task requirements are accurately matched, and enhances the synergy and reliability of aircraft control.

[0037] S314: Configure the execution unit. When the general large model determines that a professional model needs to intervene during the execution of a task, it triggers the small model to perform calculations by calling the model context protocol tool call interface corresponding to the model tool subset, and receives the returned results, thereby realizing the collaborative work between the general large model and the vertical domain small model. Specifically, the execution unit's built-in task triggering logic is invoked. When the general-purpose large model performs the spacecraft orbit maintenance task, it analyzes the sensing data in real time. If the detected orbit radius deviation exceeds ±2km or the attitude angle fluctuation exceeds ±0.5°, it determines that a specialized model needs to intervene. A small orbit correction reinforcement learning model from the model toolkit is called via a RESTful API interface, transmitting a standardized JSON data packet containing real-time orbit parameters and environmental disturbance data. After the small model completes its calculations, it returns results such as thrust adjustment and attitude correction angle. The execution unit verifies the data's validity and then feeds it back to the general-purpose large model to support subsequent task decisions. This design, through intelligent judgment triggering of specialized model intervention, achieves complementarity between the decision-making capabilities of the general-purpose large model and the computational accuracy of the vertical domain small model. This improves the spacecraft's control response speed, ensures the accuracy and collaborative efficiency of orbit task processing, and enhances the system's adaptability to complex scenarios.

[0038] In the construction of the model toolkit, the tool encapsulation stage involves selecting vertical domain models such as orbital anomaly diagnosis and fuel consumption prediction models, encapsulating them into independent calling units via RESTful APIs, and limiting the precision of input parameters and the output format. The protocol packaging stage defines the calling protocol using JSON format, assigns a unique identifier to each small model, and defines a unified request and response body. The framework assembly stage categorizes and integrates the five small models into a model toolkit according to the "diagnosis-prediction-optimization" logic, configuring the call triggering conditions. This design ensures controllable small model calls through encapsulation, unifies the interaction standard of the protocol, and assembles a domain-specific calling logic, adapting to the needs of professional model collaboration in spacecraft orbit calculations, and improving the toolkit's relevance and response efficiency.

[0039] S32: Construct a subset of computational tools, which includes computational tools containing basic computational methods and complex computational formulas, used to encapsulate computational functions to provide mathematical computational capabilities to the intelligent agent; this step includes the following sub-steps: S321: Configures basic calculation tools, providing a calculator tool with basic arithmetic operation functions including addition, subtraction, multiplication, division, exponentiation, logarithms, etc. Specifically, the basic calculation tools use Python to encapsulate independent functions, with dedicated interfaces designed for operations such as addition, subtraction, multiplication, division, exponentiation, and logarithms: the addition tool takes floating-point parameters a and b as input and outputs the sum with three decimal places; the exponentiation tool takes the base x and the exponent y as input and outputs the calculation result of x^y. A unified input / output protocol is used in JSON format, with operation type identifiers clearly marked, and operations are categorized and integrated into "arithmetic operations - exponentiation operations - logarithmic operations." A 50ms timeout threshold is configured to ensure rapid response. This standardized basic calculation interface ensures calculation accuracy and call stability, supports the decomposition and calculation of complex formulas in spacecraft orbit calculations, improves the efficiency of the overall calculation process, and enhances the versatility and extensibility of the toolset.

[0040] S322: Configure complex calculation tools, providing function tools for executing complex calculation formulas in specific domains; Specifically, the complex calculation tools are designed for the needs of spacecraft orbit calculations, encapsulating specialized functions such as Hohmann transfer velocity increment calculation, Kepler's third law orbital period calculation, and centripetal force calculation. For example, the Hohmann transfer Δv tool takes the perigee radius r1 (6500km), apogee radius r2 (42164km), and Earth's gravitational constant μ (3.986×10^5km / s²) as input and outputs Δv1 (2.92km / s) and Δv2 (1.45km / s), with an accuracy controlled within 0.001km / s; the orbital period tool takes the semi-major axis a as input and outputs the period T (in seconds). The input and output formats are defined using the JSON protocol, categorized and integrated according to "orbit transfer - parameter solving - mechanical analysis," and configured with calculation error verification logic. This design adapts to the complex formula solving needs in the orbit calculation field, encapsulates specialized functions to ensure calculation accuracy, categorizes and integrates functions to improve calling efficiency, and performs error verification to avoid data deviation, providing accurate support for core tasks such as orbit planning and enhancing the toolset's domain adaptability and engineering practicality.

[0041] S323: Configuration tool encapsulation unit, which encapsulates both the basic computing tool and the complex computing tool according to the Model Context Protocol (MCP), and configures functional description prompts for them respectively; Specifically, the tool encapsulation unit, based on the Model Context Protocol (MCP), first assigns unique MCP identifiers to basic and complex computational tools, defining a unified function structure. Then, it writes prompts according to "Tool ID + Core Function + Applicable Scenarios," for example: "Tool ID: xxx, Function: Input orbital radius and gravitational constant, solve for Hohmann transfer velocity increment, accuracy 0.001 km / s; Applicable Scenarios: Earth-orbit to geostationary orbit transfer calculation." The basic tool prompts simultaneously clarify the computation type and accuracy. This design achieves standardized tool encapsulation through the MCP protocol, ensuring compatibility with the toolset architecture; the functional prompts facilitate accurate invocation of large models; unified management of basic and complex tools improves orbital computation efficiency and accuracy, and enhances the toolset's versatility and collaborative adaptability.

[0042] S324: Configuration interface calling unit, providing a standardized calling interface for general large models. When the large model encounters computational needs during task planning or execution, it can autonomously decide to call a subset of computational tool interfaces, input parameters, and obtain accurate computational results, thereby ensuring the accuracy and interpretability of the computation process. Specifically, the interface call unit uses a RESTful API to build a standardized interface, defining the request format as JSON, including tool identification, calculation parameters, and result accuracy requirements; the response format includes calculation results, error range, and step descriptions. When the general large model performs an aircraft orbit transfer planning task and requires solving the Hohmann transfer velocity increment, it autonomously calls this interface, inputs preset parameters, and receives the precise Δv value and formula derivation process returned by the tool. Simultaneously, the interface records the call log. This design achieves seamless integration between the large model and the calculation tool through a standardized interface, improves task processing efficiency through autonomous decision-making, and ensures calculation accuracy and interpretability through precise results and process recording, adapting to the stringent requirements of orbit calculation for data accuracy.

[0043] The design comprises the following subsets of computational tools: **Tool Encapsulation Stage:** Core mathematical functions for orbit calculations, such as the Kepler's Third Law calculator and the centripetal force formula function, are selected and encapsulated into controllable input / output interfaces using Python decorators, with parameter precision limited. **Protocol Packaging Stage:** A calling protocol is defined using JSON format, including tool identifiers, input fields, and output fields. **Framework Assembly Stage:** Twelve similar functions are categorized into "Basic Mechanics Calculations - Orbital Parameter Solving - Transfer Orbit Analysis," integrated into the computational tool subset, and call priority rules are configured. This design ensures accurate and controllable computational functions through encapsulation, unifies calling standards through protocols, improves tool calling efficiency through categorized assembly, adapts to the stringent mathematical precision requirements of spacecraft orbit calculations, and enhances the synergy between the toolset and large models.

[0044] S33: Construct a subset of sensing tools, wherein the subset of sensing tools is configured with a method for receiving data from analog sensors to obtain environmental state information; Specifically, the design constructs a subset of sensing tools. The tool encapsulation stage simulates core sensors such as star sensors, millimeter-wave radar, and solar sensors. The star sensor interface outputs attitude angles, and the radar interface outputs atmospheric density, with a unified data sampling frequency of 10Hz. The protocol packaging stage defines the calling protocol using JSON format, specifying message formats including sensor ID, data type, timestamp, and numerical value. The framework assembly stage integrates six types of sensor interfaces categorized as "attitude perception - environmental perception," configuring data anomaly thresholds to form the sensing tool subset. This design, through encapsulating standardized sensor interfaces, ensuring data transmission consistency through protocols, and assembling data for classification and processing, accurately acquires the spacecraft's environmental status, provides real-time and reliable data support for orbit calculations, and improves the stability and adaptability of the system's environmental perception.

[0045] S34: Construct a subset of action tools, wherein the subset of action tools is configured with a method for simulating aircraft coordinate transformation; Specifically, in constructing the action tool subset, the tool encapsulation stage involves: simulating the interfaces of core motion modules such as main thruster thrust control, attitude adjustment servo control, and trajectory correction valve control, and encapsulating control functions. The main thruster thrust input range is limited to 0-500N (accuracy 1N), and the servo angle control range is -30° to 30° (accuracy 0.1°). The protocol packaging stage uses JSON format to define the calling protocol, defining the control command format including device ID, command type, parameter values, and execution duration to ensure stable command transmission. The framework assembly stage integrates four types of control interfaces categorized as "thrust control - attitude adjustment - trajectory correction," configuring trigger logic to form the action tool subset. This design, through encapsulating standardized control interfaces, ensuring accurate command transmission through protocols, and modularizing control logic through assembly, adapts to the motion control requirements of the aircraft, improves the real-time performance and accuracy of trajectory adjustment, and enhances the collaborative adaptability between the toolset and the aircraft control system.

[0046] S35: Standardize and encapsulate each of the above tool subsets through the model context protocol, giving them a unified description, input and output format. Finally, assemble the four tool subsets into a dedicated tool set for spacecraft orbit calculation. Specifically, each of the above tool subsets is standardized and encapsulated through a model context protocol, possessing a unified description, input, and output format. For example... Figure 5As shown in the figure, this is the MCP toolset for complex spacecraft missions supported by the platform, taking spacecraft path planning as an example. This toolset is a core component of the platform of this invention, aiming to encapsulate specialized capabilities into standardized tools for use by Large Language Model (LLM) agents. The toolset is divided into four highly specialized types: First, the computational toolset provides computational capabilities for core algorithms of aerospace dynamics, including a Hohmann transfer calculation tool for optimal orbit transfer, a Lambert orbit change tool for solving two-point boundary value problems, a double-pulse calculation tool for orbit adjustment, and angle measurement approach and angle measurement and ranging approach tools for final rendezvous and docking. These tools ensure the theoretical accuracy and optimality of path planning. Second, the detection toolset simulates the navigation and orbit determination sensors of spacecraft. Among them, the ground orbit determination simulation, GNSS single-machine simulation, and BlackEye single-machine simulation tools are designed to provide agents with real-time or simulated position, velocity, and relative navigation data, enabling them to perceive their own state and surrounding environment, providing information input for decision-making. Then, the motion toolset encapsulates the spacecraft's actuator functions and specifically provides propulsion simulation tools to control the spacecraft's movement along the X, Y, and Z axes. The agent can invoke these tools to simulate or actually execute ignition commands, thereby achieving precise control over its orbit and attitude. Finally, the agent toolset plans to integrate a small reinforcement learning decision-making model for mission-critical tasks. This will be a specialized vertical domain model used to provide reinforcement learning-trained professional decision support to the larger model when handling extremely complex or high-risk real-time decision-making tasks, achieving a deep fusion of general cognition and domain expertise.

[0047] The computational tools subset configures core orbital mechanics calculation functions; the model tools subset integrates multiple scenario-based reinforcement learning model interfaces for orbital anomaly diagnosis and path pre-planning; the perception tools subset includes various tools such as simulated star sensor data parsing and millimeter-wave radar environmental parameter extraction to achieve real-time status data processing; and the action tools subset configures two types of execution tools, such as simulated thruster thrust control and attitude adjustment command generation, which interface with the spacecraft control unit. This design divides the tool subsets according to the entire orbital calculation process, covering the entire chain from perception to computation to analysis to execution, improving the targeting of tool calls; and the specialized configuration adapts to the spacecraft's orbital calculation scenarios, enhancing the practicality and scalability of the toolset and ensuring efficient collaborative operation of the system.

[0048] S4: Through document extraction, adaptive segmentation, segmented embedding, and semantic retrieval, configure the specification knowledge base that the large model needs to refer to when executing spacecraft tasks, forming a task execution manual knowledge base for storing vertical domain specifications for spacecraft orbital motion calculations and a vertical domain knowledge base retrieval auxiliary workflow for assisting large model decision-making; this step includes the following sub-steps: S41: Configure the knowledge base in stages, including: configuring a document import module to support importing and parsing specification documents in various formats and extracting text, tables, and image content; configuring an intelligent segmentation module to segment the imported document content, using regular expression segmentation or natural paragraph segmentation to form multiple knowledge units; configuring a vector storage module to use an embedding model to convert the text of each knowledge unit into a high-dimensional vector and associate the vector with the text content in a vector database; thereby completing the configuration and forming a knowledge base for storing the task execution manual for the vertical domain specification of aircraft orbital motion calculation. Specifically, the document import module supports formats such as PDF, Word, and Excel. After importing documents such as the "Spacecraft Orbit Design Specification," it parses and extracts orbital parameter constraints from the text, thrust performance data from tables, and orbital schematic diagrams from images. The intelligent segmentation module uses regular segmentation (e.g., splitting chapters using "Section X" as a separator) and natural paragraph segmentation (e.g., splitting by technical specification paragraphs) to divide the document into knowledge units such as "Hohmann transfer orbit velocity increment calculation requirements" and "orbit correction maneuver threshold standards." The vector storage module uses the Sentence-BERT embedding model to convert each knowledge unit text into a 768-dimensional high-dimensional vector, which is then stored in the Chroma vector database after being associated with the text content, forming a specification knowledge base for the orbital calculation domain. This design achieves unified integration of multi-format specification documents, improves the refinement of knowledge units through segmentation, and ensures efficient semantic retrieval through vector storage, providing accurate specification support for spacecraft orbit calculation and enhancing the compliance and accuracy of mission execution.

[0049] S42: Establish a retrieval-assisted large-scale model workflow, including: establishing a query triggering workflow, which initiates a query request to the knowledge base when the large-scale model needs standardized knowledge assistance during a task; establishing a question vectorization workflow, which converts natural language questions into query vectors through an embedding model; establishing a similarity calculation workflow, which calculates and sorts the similarity between the query vectors and the vectors of all knowledge units in the knowledge base; establishing a content recall workflow, which selects the original text content corresponding to several knowledge units with the highest similarity; and establishing a content filling workflow, which uses the recalled knowledge text content as context information to fill the prompts in subsequent execution steps, providing the large-scale model with accurate standardized basis; thereby completing the configuration and forming a vertical domain knowledge base retrieval-assisted workflow to assist the large-scale model in decision-making. Specifically, the workflow includes: a query-triggered workflow: when the large model is performing a spacecraft orbit correction task and cannot determine the propellant reserve threshold, it automatically initiates a knowledge base query request for "minimum standard for orbit correction propellant reserve"; a question vectorization workflow: using a Sentence-BERT embedding model, the natural language question is converted into a 768-dimensional query vector; a similarity calculation workflow: using a cosine similarity algorithm, the query vector and the knowledge base vector are calculated and sorted from high to low similarity scores; a content retrieval workflow: selecting the top 5 knowledge units with the highest similarity, and extracting the original text "the single propellant reserve for orbit correction shall not be less than 8%" from the "Spacecraft Propellant Usage Specification"; and a content filling workflow: using this text as context to fill subsequent prompts, providing a basis for the large model's decision-making. This design enables precise on-demand knowledge retrieval, vectorization and similarity calculation ensure result matching, content filling provides standard support for the large model, improves the compliance and accuracy of orbit calculation decisions, and enhances the practicality and adaptability of the workflow.

[0050] like Figure 6 As shown in the diagram, this platform supports a knowledge base of vertical domain specification manuals and documents, adaptable to various complex vertical domains. The vertical domain knowledge base is the core guarantee mechanism for ensuring that large models strictly adhere to domain specifications and avoid "illusions" when performing professional tasks. The entire process can be divided into two main stages: knowledge base construction and large model calling upon the knowledge base.

[0051] Phase 1: Knowledge Base Construction First, specialized knowledge and normative texts from various sources are imported into the vertical domain. These heterogeneous documents from multiple sources are processed by the platform's parsing module and then proceed to the text segmentation stage. The system uses an intelligent segmentation algorithm to break down lengthy documents into multiple semantically complete knowledge blocks. Each knowledge block is then vectorized using an embedding model, transforming it into a high-dimensional numerical representation, i.e., a knowledge vector, which semantically represents the content of the knowledge block. Finally, all these knowledge vectors and their corresponding original text knowledge blocks are stored in pairs in an internal database, thus completing the construction of a semantically meaningful knowledge base that can be efficiently queried.

[0052] Phase Two: Large Model Calls Knowledge Base When the large model requires specialized knowledge to perform specific vertical tasks, it triggers a knowledge retrieval process. The large model converts the current question information into a query request. This question information is also converted into a question vector by the same embedding model. The system then sends this question vector into a vector database, calculates its similarity with all knowledge vectors in the database, and selects the top-k knowledge vectors with the highest similarity. The original knowledge block text corresponding to these top-k vectors is retrieved from the database. Finally, the system combines these retrieved key specification text fragments with the user's original question information to form a context-rich, enhanced prompt, which is then processed by the large model. Based on these precise specification criteria, the large model performs reasoning or operations, outputting a reliable result that conforms to general logic and strictly adheres to domain specifications, completing a closed loop from "knowledge storage" to "knowledge application."

[0053] S5: Based on the aforementioned local quantization model, deploy the dedicated workflow, aircraft-specific toolset, and standardized knowledge base to form a large-scale model application platform for intelligent execution of aircraft missions; this step includes the following sub-steps: S51. Deploy dedicated workflow: Fill in the task conditions and input data into the multi-stage prompts of the platform's dedicated workflow according to the category and stage, and then call the dedicated workflow corresponding to the task in the large model step by step. Specifically, taking the spacecraft's Earth-Mars transfer orbit planning mission as an example, the mission conditions are categorized into "pre-planning" and "real-time calibration" stages, with prompts entered as follows: For the pre-planning stage, "orbit type: Hohmann transfer, accuracy requirement: Δv error ≤ 0.1km / s" is entered, along with input data such as "initial orbit radius 6500km, Mars orbit radius 2.28×10^8km"; for the real-time calibration stage, "anomaly threshold: orbit deviation exceeding 2km triggers calibration" is entered, along with input data such as "real-time velocity 7.8km / s, atmospheric density 1.2×10⁻¹²kg / m³". Subsequently, following the sequence of "pre-planning - sensing data acquisition - anomaly analysis - orbit correction", the large model is invoked step-by-step to execute the corresponding stage's dedicated workflow. This design, by categorizing data into stages, ensures logical coherence, and the step-by-step invocation adapts to the mission execution process, guaranteeing the orderly progress of the workflow, improving the accuracy and efficiency of orbit planning, and enhancing the standardization and controllability of mission processing.

[0054] S52. Deploy aircraft-specific toolset: In the review step of the dedicated workflow, when it is found that the task execution context information of the large model mentions the need for tool support, the platform will perform tool routing and input the tool parameters in the context information to execute the corresponding function, and fill the function output into the model context for the large model to call. Specifically, taking the review process for spacecraft orbit correction tasks as an example: when the review module detects that the task execution context output by the large model includes "the Hohmann transfer velocity increment needs to be calculated to adjust the orbital deviation," the platform automatically triggers the tool routing mechanism to match the "Hohmann transfer Δv calculation function" in the tool subset. Then, it extracts the tool parameters (perigee radius 6500km, apogee radius 42164km, Earth's gravitational constant 3.986×10^5km / s²) from the context information, inputs them into the function to perform the calculation, and obtains results of Δv1=2.92km / s and Δv2=1.45km / s. Finally, the output results are automatically filled into the model context for the large model to call when generating orbit correction instructions later. This design achieves precise matching through tool routing, automatically extracts parameters to reduce manual intervention, and ensures contextual coherence through result backfilling. This ensures that the large model efficiently obtains tool support, improves the accuracy and automation level of spacecraft mission execution, and enhances the collaborative adaptability of the toolset and workflow.

[0055] S53. Deploy the aircraft specification knowledge base: In the review step of the dedicated workflow, when it is found that the task execution context information of the large model mentions the need for additional knowledge support, the platform will call the knowledge base and perform knowledge retrieval. By vectorizing the knowledge requirements and calculating the vector similarity with the data entries in the knowledge base, the platform will obtain the few data with the highest similarity and fill them into the context information for the large model to call. Specifically, taking the review process for spacecraft orbit maintenance missions as an example: When the review module detects a question in the execution context of the large model regarding "whether the propellant reserve for orbit correction meets the standard," the platform automatically calls the spacecraft specification knowledge base. The knowledge requirement "minimum propellant reserve standard for orbit correction" is transformed into a high-dimensional vector using the Sentence-BERT model, and its cosine similarity is calculated with the vector data in the knowledge base. The top three entries with the highest similarity are selected, and the core content of "single propellant reserve for orbit correction shall not be less than 8%" from the "Spacecraft Propellant Usage Specification" is extracted and filled into the model context for the large model to determine whether the propellant meets the mission requirements. This design achieves accurate matching and rapid retrieval of knowledge requirements, providing authoritative specification basis for the large model, avoiding decision-making bias, improving the compliance and accuracy of spacecraft mission execution, and enhancing the collaborative support capabilities of the knowledge base and workflow.

[0056] like Figure 7As shown in the diagram, this platform uses an intelligent agent architecture to call the core large model to execute tasks. The diagram details the intelligent agent framework used by this platform. The intelligent agent of this platform is not a simple interface for calling a large language model, but an autonomous system with comprehensive cognitive capabilities. Its core consists of four modules: the "Tools" module endows the intelligent agent with the ability to interact with the external world and perform operations, such as calling a calculator for precise calculations, using a search engine to obtain real-time information, checking the weather, or parsing user commands; the "Memory" module is divided into short-term memory and long-term memory, enabling the intelligent agent to persistently store historical interaction information and maintain an understanding of the current conversation context, thereby achieving coherent and stateful task processing; the "Reflection" module is the intelligent agent's advanced reasoning engine, using strategies such as thought chain reasoning, task decomposition, and self-questioning and answering to deeply process and analyze information, thereby formulating solutions to complex problems; finally, all these capabilities are integrated under the coordination of the "Action" module, and the intelligent agent executes specific operation commands based on the final decision results. The "Date" function at the top ensures that all operations and decisions are chronological, accurately recorded, and traceable. This architecture enables intelligent agents to possess integrated capabilities such as perception, memory, thinking, and execution, much like human experts, thereby efficiently and reliably completing complex tasks in vertical domains.

[0057] This invention also provides a large-scale model application platform for intelligent aircraft missions, characterized in that the platform is constructed using the aforementioned method for constructing a large-scale model application platform for intelligent aircraft missions.

[0058] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application 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 this application are indicated by the claims.

[0059] It should be understood that this application 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 this application is limited only by the appended claims.

Claims

1. A large model application platform construction method for aircraft task intelligence, characterized in that, Comprise: S1: quantizing general field large models, deploying to local through inference acceleration, configuring local API calling interface, obtaining local quantized large models; S2: writing multi-stage prompt word templates, maintaining model calling context, assembling multiple large model calls in order, selection, and loop structure to form two sets of special work flows for large model execution tasks for aircraft path pre-planning and real-time path calibration; S3: using tool packaging, protocol wrapping, and framework assembly to configure input and output controllable functions as large model usable tools, and then assembling to form aircraft orbit calculation special tool set; S4: through document extraction, adaptive segmentation, segmented embedding, and semantic retrieval, configuring specification knowledge base required for large model execution of aircraft tasks, forming task execution manual knowledge base for storing vertical field specifications for aircraft orbit motion calculation, and vertical field knowledge base retrieval auxiliary work flow for assisting large model decision-making; S5: based on the local quantized large model, deploying the special work flow, aircraft special tool set, and specification knowledge base to form a large model application platform for intelligent execution of aircraft tasks.

2. The large model application platform construction method for aircraft mission intelligence according to claim 1, characterized in that, Quantizing general field large models, deploying to local through inference acceleration, configuring local API calling interface, obtaining local quantized large models, comprising: S11: selecting a general large language model that meets the requirements of orbit motion calculation vertical field tasks in terms of parameter size and supported modalities, and performing precision quantization processing on the selected large model; S12: according to the characteristics of vertical field data, assembling data sets of predetermined format to fine-tune the large model after quantization; using Ollama large model local deployment framework to deploy the fine-tuned large model on local server or terminal device; S13: calling the large model suitable for orbit calculation vertical field deployed on local through application API calling interface provided by Ollama framework, thereby obtaining local quantized large model.

3. The method of claim 1, wherein the method is characterized by: Writing multi-stage prompt word templates, maintaining model calling context, assembling multiple large model calls in order, selection, and loop structure to form two sets of special work flows for large model execution tasks for aircraft path pre-planning and real-time path calibration, comprising: S21: writing special prompt word templates for receiving user original input and combining into prompt words conforming to large model input specifications in predetermined format; through API calling large model, obtaining original results generated by model, and after output analysis of returned results, storing in model calling context maintenance to form large model basic calling work flow; S22: First assembly multiple large model calls, including: the first stage, using prompt template one to decompose a single complex task into multiple simple subtasks; the second stage, using prompt template two to plan the steps for each subtask; the third stage, using prompt template three to review the planned subtask steps, determine whether it meets the task goal and specification, if not, return to the second stage to re-plan; the fourth stage, using prompt template four to integrate all subtask steps, output a complete, fine-grained task execution planning sequence, form a spacecraft path pre-planning dedicated workflow; S23: Second assembly multiple large model calls, including: configure the environment perception module, regularly collect task execution environment state information; the first stage analysis, using prompt template one to analyze the state information, determine whether the task execution is normal; the second stage analysis, using prompt template two to analyze the possible abnormal reasons, locate the task steps that need to be adjusted, generate modification suggestions; the review stage, using prompt template three to review the modification suggestions, if the review is passed, execute the modification and update the task plan, if not, return to the second stage to generate new suggestions, form a spacecraft real-time path calibration dedicated workflow.

4. The method of claim 1, wherein the method is characterized by: Configure input and output controllable functions as tools available to large models, and then assemble them to form a spacecraft orbit calculation dedicated tool set, including: S31: Build a model tool subset, which contains reinforcement learning small models for controlling spacecraft perception and motion; S32: Build a calculation tool subset, which contains calculation tools including basic calculation methods and complex calculation formula tools, used to encapsulate calculation class functions to provide mathematical calculation capabilities for agents; S33: Build a perception tool subset, which contains methods to simulate sensor data reception to obtain environment state information; S34: Build an action tool subset, which contains methods to simulate spacecraft coordinate transformation; S35: Standardize the encapsulation of each tool subset through the model context protocol, with unified description, input and output formats, and finally assemble the four tool subsets to form a spacecraft orbit calculation dedicated tool set.

5. The large model application platform construction method for aircraft mission intelligence according to claim 4, characterized in that, The model tool subset contains reinforcement learning small models for controlling spacecraft perception and motion, including: S311: Configure the interface definition unit to fix the input and output data formats of the reinforcement learning spacecraft control domain professional small model; S312: Configure the tool encapsulation unit based on the model context protocol (MCP) tool set framework to encapsulate small models that meet the input and output formats into a callable tool function; S313: Configure the prompt description unit to write the corresponding tool function prompts for each encapsulated model tool, which are used to describe their functions when the large model makes decisions; S314: Configure the execution unit to call, when the general large model determines that the professional model needs to intervene in the process of executing the task, trigger the small model to perform calculation through the model context protocol to call the model tool subset interface, and receive the returned results, so as to realize the cooperative work of the general large model and the vertical field small model.

6. The large model application platform construction method for aircraft mission intelligence according to claim 4, characterized in that, The calculation tool subset includes a large number of basic calculation methods and complex calculation formula tools, including: S321: Configure the basic calculation tool to provide a calculator tool including basic arithmetic operations such as addition, subtraction, multiplication, division, exponent, logarithm, etc. S322: Configure the complex calculation tool to provide a function tool for executing specific field complex calculation formula; S323: Configure the tool encapsulation unit, encapsulate the basic calculation tool and the complex calculation tool according to the model context protocol (MCP), and configure the function description prompt word for each of them; S324: Configure the interface calling unit to provide a standardized calling interface for the general large model, when the large model encounters calculation demand in the task planning or execution process, it can independently decide to call the calculation tool subset interface, input parameters and obtain accurate calculation results, so as to ensure the accuracy and explainability of the calculation process.

7. The method of claim 1, wherein the method is characterized by: Through document extraction, adaptive segmentation, segmented embedding, semantic retrieval, configure the specification knowledge base required by the large model to execute the aircraft task, form the task execution manual knowledge base for storing the vertical field specification of aircraft orbit motion calculation, and the vertical field knowledge base retrieval auxiliary workflow for assisting the decision of the large model, including: S41: Configure the knowledge base in stages, including: configure the document import module to support importing and parsing specification documents in multiple formats, and extract text, table and picture content therein; configure the intelligent segmentation module to segment the imported document content, which can use regular segmentation or natural paragraph segmentation to form multiple knowledge units; configure the vector storage module to convert the text of each knowledge unit into a high-dimensional vector using an embedding model, and store the vector and the text content in the vector database; thereby completing the configuration and forming the task execution manual knowledge base for storing the vertical field specification of aircraft orbit motion calculation; S42: Build the retrieval auxiliary large model workflow, including: build the query trigger workflow, when the large model needs specification knowledge assistance in the task process, initiate a query request to the knowledge base; build the question vectorization workflow to convert natural language questions into query vectors through an embedding model; build the similarity calculation workflow to calculate and sort the similarity between the query vector and all knowledge unit vectors in the knowledge base; build the content recall workflow to select the original text content corresponding to the knowledge units with the highest similarity; build the content filling workflow to fill the recalled knowledge text content as context information into the prompt word of the subsequent execution step, to provide accurate specification basis for the large model; thereby completing the configuration and forming the vertical field knowledge base retrieval auxiliary workflow for assisting the decision of the large model.

8. The method of claim 1, wherein the method is characterized by: Based on the local quantization large model, the special workflow, the aircraft special tool set and the specification knowledge base are deployed to form a large model application platform for intelligently executing aircraft tasks, comprising: S51, deploying a special workflow: filling the task conditions and input data into the multi-stage prompt words of the platform special workflow according to the category and the stage, and then calling the large model to execute the special workflow corresponding to the task step by step; S52, deploying an aircraft special tool set: in the audit step of the special workflow, when it is found that the task execution context information of the large model mentions the need for tool support, the platform will route the tool and input the tool parameters in the context information to execute the corresponding function, and fill the function output result into the model context for the large model to call; S53, deploying an aircraft specification knowledge base: in the audit step of the special workflow, when it is found that the task execution context information of the large model mentions the need for additional knowledge support, the platform will call the knowledge base and perform knowledge recall, vectorize the knowledge demand and calculate the vector similarity with the data entries in the knowledge base to obtain several data with the highest similarity, fill them into the context information for the large model to call.

9. An application platform for large models oriented to intelligentization of aircraft missions, characterized in that, The platform is constructed by the large model application platform construction method for aircraft task intelligence of any one of claims 1-8.

Images