Low-altitude management task intelligent scheduling method, system and device based on large language model

Abstract

This application discloses an intelligent scheduling method, system, and device for low-altitude airspace management tasks based on a large language model, relating to the field of low-altitude airspace management technology. The method acquires a user's natural language task request, inputs it into a large language model to determine its executability, and if it is not executable, calls a search engine or returns an unprocessable message. If executable, the large language model is scheduled to perform semantic parsing and logical decomposition of the low-altitude airspace management task request, obtaining multiple sub-tasks with execution order logic. Then, the large language model is scheduled to form a mapping relationship between sub-tasks and tools, and the tools are scheduled to execute each sub-task sequentially, obtaining the sub-task execution results. Finally, all sub-task execution results are aggregated into the large language model to generate the final execution result, which is returned to the user. This improves the system's ability to respond to complex scenarios and unexpected situations, reduces development, testing, and maintenance costs, and improves resource utilization efficiency.

Description

Technical Field

[0001] This application relates to the field of low-altitude airspace management technology, specifically to an intelligent scheduling method, system, and device for low-altitude airspace management tasks based on a large language model. Background Technology

[0002] With the rapid rise of the low-altitude economy, low-altitude derivative businesses, such as drone logistics delivery, airbus, low-altitude inspection, and airspace and route management, are becoming increasingly busy. Resource utilization efficiency and task processing capacity in low-altitude airspace have become core bottlenecks for the industry's development. In the early stages of handling low-altitude management tasks, the industry mainly adopted traditional management and scheduling methods, namely, task allocation and resource coordination through manual intervention or simple rule engines. With the improvement of informatization, traditional management and scheduling methods have gradually upgraded to the current system management method, namely, unified scheduling and management of low-altitude aircraft through a centralized information system.

[0003] However, in the context of rapid digitalization and intelligentization, the execution of complex low-altitude missions often requires services spanning multiple fields, scenarios, and environments. For example, a typical low-altitude logistics mission may simultaneously involve multiple stages such as route planning, meteorological data analysis, airspace conflict detection, aircraft status monitoring, and emergency response. While existing system management methods can be applied to generalized, simple task processing, their processing logic is relatively rigid and struggles to effectively address dynamic changes and real-time requirements in complex scenarios.

[0004] To address these challenges, the industry has entered a second phase of development: developing customized solutions based on the specific needs of each application scenario. For example, dedicated route planning modules are developed for drone logistics scenarios, and dedicated safety monitoring modules are developed for airbus scenarios. This customized development model has improved the system's ability to handle complex tasks to some extent. However, customized solutions also suffer from significant rigidity. First, new application scenarios require re-customization, leading to high time and financial costs. Second, replacing or upgrading a functional component in the existing processing logic, such as using a better path planning algorithm, requires redeveloping and testing the entire processing flow, resulting in extremely high maintenance costs. More importantly, when the system encounters unforeseen emergencies, such as sudden severe weather or temporary airspace control, the lack of corresponding response strategies and the inflexible rigid processing logic often paralyzes the system, rendering it unable to respond effectively.

[0005] Therefore, for intelligent scheduling and processing of low-altitude management tasks, there is an urgent need to introduce more advanced intelligent technologies to improve the system's adaptability and generalization capabilities, so as to make flexible and effective responses when facing complex scenarios and emergencies.

[0006] To address the issues of rigid low-altitude airspace management systems and their inability to handle complex scenarios mentioned in the background, the industry has recently begun to explore the introduction of Large Language Models (LLMs) to enhance system intelligence. LLMs demonstrate superior capabilities in natural language understanding, reasoning, and generation. Unlike traditional rule-based systems, LLM-driven task processing exhibits flexibility, adapting to new situations through contextual learning. However, when applying LLMs to the specific technical field of low-altitude airspace management task scheduling, existing solutions still suffer from the following shortcomings: First, when dealing with highly complex, dynamically changing, and multi-constrained low-altitude airspace management tasks, relying solely on a single LLM processing model cannot fully realize its potential, often leading to unreasonable task decomposition, excessively long inference chains, and limited processing efficiency. Second, existing technologies lack a mechanism to effectively combine LLM inference capabilities with external functional APIs, failing to achieve dynamic task decomposition and tool mapping, making it difficult for LLMs to truly implement specific and operable low-altitude airspace management instructions. Finally, existing systems generally employ predefined static workflows, lacking dynamic task orchestration capabilities, and are unable to adaptively adjust to unforeseen emergencies. Summary of the Invention

[0007] The purpose of this application is to provide an intelligent scheduling method, system and device for low-altitude management tasks based on a large language model, which solves the problems of rigidity, poor adaptability and high maintenance cost of existing low-altitude management task processing systems when facing complex scenarios.

[0008] This application is achieved through the following technical solution:

[0009] The first aspect of this application provides an intelligent scheduling method for low-altitude management tasks based on a large language model, including:

[0010] Obtain low-altitude airspace management task requests submitted by users through human-computer interaction; the low-altitude airspace management task requests are natural language task requests describing low-altitude airspace management.

[0011] The low-altitude airspace management task request is input into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task. The executability is either executable or not executable.

[0012] If the executability is deemed non-executable, a preset web search engine will be invoked to retrieve information in an attempt to answer the question, or a response message indicating that the request cannot be processed will be directly returned to the user.

[0013] If the executability is executable, then the large language model is scheduled to perform semantic parsing and logical decomposition on the low-altitude management task request to obtain multiple sub-tasks with execution order logic;

[0014] The large language model is scheduled to match the corresponding tools to be invoked for the multiple subtasks with execution order logic, forming a mapping relationship between subtasks and tools;

[0015] Based on the mapping relationship between tasks and tools and the tool registration list, tools are scheduled to execute each subtask in sequence, and the execution results of the subtasks are obtained.

[0016] The execution results of all subtasks are aggregated into the large language model, which then performs comprehensive analysis, semantic understanding, and logical integration on the execution results of all subtasks to generate the final execution result, which is then returned to the user.

[0017] In one possible implementation, the tool is configured as a six-tuple structure, specifically:

[0018] ;

[0019] in, Representation tools, This represents a unique identifier corresponding to the tool. This represents the natural language function description corresponding to the tool. The type label indicates the tool. This represents the executable entity of the tool. Indicates the input specifications for the tool. This indicates the output specifications of the tool.

[0020] In one possible implementation, it also includes:

[0021] If the output specification of any one of the tools completely matches the input specification of another tool, then the two tools are combined sequentially.

[0022] When multiple tools do not have data dependencies on each other, they are combined into a tool portfolio in parallel.

[0023] During the process of querying and scheduling the large language model, the tool combination can be queried and scheduled.

[0024] In one possible implementation, the low-altitude airspace management task request is input into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task, including:

[0025] The low-altitude airspace management task request is input into a large language model for processing, enabling the large language model to query the attribute information of each tool in the tool registration list; the attribute information includes at least the input requirement set, the output commitment set, and the functional description.

[0026] When the low-altitude airspace management task request matches the input requirement set, output commitment set, and functional description of any tool in the tool registration list, the executableness of the low-altitude airspace management task is determined to be executable; otherwise, it is determined to be unexecutable.

[0027] In one possible implementation, the core service interface of the abstraction layer containing the large language model is set as follows:

[0028] ;

[0029] in, This indicates a service for generating large language models. The prompt information indicates the input to the large language model, and the prompt information includes the semantic description and context data corresponding to the low-altitude management task request; This indicates that the large language model calls the configuration parameter set. This indicates the response returned by the large language model. This indicates that the function maps its input to its output.

[0030] In one possible implementation, the large language model call configuration parameter set is set as follows:

[0031] ;

[0032] in, Indicates temperature parameter; Indicates the maximum number of output tokens; Indicates kernel sampling parameters; Indicates the frequency penalty coefficient; This indicates the presence of a penalty coefficient.

[0033] In one possible implementation, based on the mapping relationship between tasks and tools and the tool registration list, tools are sequentially scheduled to execute each subtask, and the execution results of the subtasks are obtained, including:

[0034] The initialization context information is as follows: ;in, This indicates the context information for initialization. Represents the empty set;

[0035] For each subtask, execution is performed based on the mapping relationship between the tasks and tools, as well as the tool registration list:

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] in, This represents the k-th tool. Let j represent the j-th subtask, j = 1, 2, ..., m; m represent the total number of subtasks. This represents the k-th tool. Let j represent the j-th subtask, where j = 1, 2, ..., m; and m represent the total number of subtasks. This indicates the task and skill mapping operation. This represents the input data for the j-th subtask. This indicates an input extraction operation. This indicates the context after the (j-1)th step is executed. This represents the output of the j-th subtask. Indicates the execution of a skill. This represents the context after step j is executed;

[0041] All subtasks corresponding Together they serve as the result of the subtask execution.

[0042] In one possible implementation, the execution results of all subtasks are aggregated into the large language model, so that the large language model can perform comprehensive analysis, semantic understanding, and logical integration on the execution results of all subtasks to generate the final execution result:

[0043] ;

[0044] in, This indicates the final execution result. Representing a large language model, This indicates a summary / generation operation. This indicates that the last subtask has been completed, and the result includes the output of all subtasks. This indicates the original user task request.

[0045] Based on the same inventive concept, a second aspect of this application discloses an intelligent scheduling system for low-altitude management tasks based on a large language model, comprising:

[0046] The interaction layer is used to obtain low-altitude management task requests submitted by users through human-computer interaction; the low-altitude management task requests are natural language task requests describing low-altitude management.

[0047] The tool layer encapsulates tools for handling discrete tasks in the low-altitude airspace management field and maintains a tool registry list.

[0048] The orchestration layer is used to input the low-altitude airspace management task request into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task; the executability is either executable or not executable. If the executability is not executable, a preset web search engine is invoked to attempt an answer, or a response indicating that the request cannot be processed is directly returned to the user. If the executability is executable, the large language model is scheduled to perform semantic parsing and logical decomposition of the low-altitude airspace management task request, obtaining multiple subtasks with execution order logic. The large language model is then scheduled to match the multiple subtasks with execution order logic with corresponding tools to be invoked, forming a mapping relationship between subtasks and tools. Based on the mapping relationship between tasks and tools and the tool registration list, tools are sequentially scheduled to execute each subtask, obtaining the subtask execution results. All subtask execution results are aggregated into the large language model, allowing the large language model to perform comprehensive analysis, semantic understanding, and logical integration of all subtask execution results, generating a final execution result, which is then returned to the user.

[0049] The infrastructure layer provides computing power support for the interaction layer, tool layer, and orchestration layer, as well as large language model inference capabilities.

[0050] Based on the same inventive concept, a third aspect of this application provides an electronic device, including a processor and a memory;

[0051] The memory stores computer-executed instructions;

[0052] The processor executes computer execution instructions stored in the memory, causing the processor to perform the intelligent scheduling method for low-altitude management tasks based on a large language model as described in any embodiment of the first aspect.

[0053] Compared with the prior art, this application has the following advantages and beneficial effects:

[0054] This application discloses an intelligent scheduling method, system, and device for low-altitude airspace management tasks based on a large language model. The method acquires a user's natural language task request, inputs it into a large language model to determine its executability, and if it is not executable, calls a search engine or returns an unprocessable message. If executable, the large language model is scheduled to perform semantic parsing and logical decomposition of the low-altitude airspace management task request, obtaining multiple sub-tasks with execution order logic. Then, the large language model forms a mapping relationship between sub-tasks and tools, and the tools are scheduled to execute each sub-task sequentially, obtaining the sub-task execution results. Finally, all sub-task execution results are aggregated into the large language model, enabling it to comprehensively analyze, semantically understand, and logically integrate all sub-task execution results, generating the final execution result, which is then returned to the user. This improves the system's ability to respond to complex scenarios and unexpected situations, reduces development, testing, and maintenance costs, and improves resource utilization efficiency. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the exemplary embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0056] Figure 1 A flowchart illustrating an intelligent scheduling method for low-altitude airspace management tasks based on a large language model, provided for embodiments of this application;

[0057] Figure 2 A schematic diagram of an intelligent scheduling system for low-altitude management tasks based on a large language model, provided for an embodiment of this application;

[0058] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application;

[0059] The attached diagram shows the markings and corresponding component names:

[0060] 201-Interaction Layer, 202-Orchestration Layer, 203-Tool Layer, 204-Infrastructure Layer, 301-Memory, 302-Processor, 303-Bus. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0062] like Figure 1As shown, this application provides an intelligent scheduling method for low-altitude management tasks based on a large language model, including:

[0063] S101. Obtain the low-altitude management task request submitted by the user through human-computer interaction; the low-altitude management task request is a natural language task request describing low-altitude management.

[0064] Users can input low-altitude airspace management task requests through a web interface, RESTful API, command-line interface, and mobile application. The low-altitude airspace management task requests are input in natural language, which facilitates subsequent analysis and scheduling using a large language model.

[0065] S102. Input the low-altitude airspace management task request into a large language model for processing, so that the large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task; the executability is either executable or non-executable.

[0066] Optionally, the large language model mentioned in each step of the embodiments of this application can be the same large language model or different large language models. It is preferred to use different large language models so that each large language model can perform a specific function, thereby improving the processing accuracy.

[0067] The tool encapsulates various discrete business capabilities in the field of low-altitude airspace management, implementing them as independently callable functional modules, including but not limited to intelligent route planning tools, real-time meteorological data query tools, airspace conflict detection and early warning tools, aircraft status monitoring tools, and emergency command generation and issuance tools. Each tool registers its functional metadata information in the tool list, allowing the orchestration layer to match and invoke it during task decomposition.

[0068] S103. If the executability is not executable, then a preset network search engine is invoked to retrieve information in an attempt to answer the question, or a response message indicating that the request cannot be processed is directly returned to the user.

[0069] S104. If the executability is executable, then the large language model is scheduled to perform semantic parsing and logical decomposition on the low-altitude management task request to obtain multiple sub-tasks with execution order logic.

[0070] The large language model is scheduled to perform semantic parsing and logical decomposition on the low-altitude management task request, so that each part that can be executed by the tool alone is decomposed into a sub-task. By splitting the low-altitude management task request, multiple sub-tasks with execution order logic can be obtained, thereby improving the adaptability of the low-altitude management system in complex dynamic environments.

[0071] S105. The large language model is scheduled to match the corresponding calling tools for the multiple subtasks with execution order logic, forming a mapping relationship between subtasks and tools;

[0072] S106. Based on the mapping relationship between the tasks and tools and the tool registration list, the tools are scheduled to execute each sub-task in sequence, and the execution results of the sub-tasks are obtained.

[0073] S107. The execution results of all subtasks are aggregated into the large language model, so that the large language model can perform comprehensive analysis, semantic understanding and logical integration of the execution results of all subtasks, generate the final execution result and return it to the user.

[0074] It is worth noting that the large language models involved in each step can be the same model or different models. For example, the large language model in S102 can be a task analysis large language model; the large language model in S104 can be a task decomposition large language model; the large language model in S105 can be a task matching large language model; and the large language model in S107 can be a result aggregation large language model.

[0075] This application provides an intelligent scheduling method for low-altitude management tasks based on a large language model. The intelligent scheduling architecture uses modular functional APIs as pluggable tools, enabling the system to autonomously complete task handling ability judgment, intelligent decomposition, dynamic arrangement of sub-tasks, precise tool mapping and invocation according to the requirements of complex low-altitude tasks, and achieve intelligent aggregation of execution results. This replaces the traditional fixed processing logic and single LLM processing mode, and comprehensively improves the adaptability, processing efficiency and operational stability of the low-altitude management system in complex dynamic environments.

[0076] In one possible implementation, the tool is configured as a six-tuple structure, specifically:

[0077] ;

[0078] in, Representation tools, This represents a unique identifier for a tool, used to index and invoke the tool in the tool list; This represents the natural language functional description of the tool, which helps LLM understand the functional scope of the tool during task decomposition. The tool type label identifies the low-altitude business area to which the skill belongs, such as "flight planning", "meteorological services", "monitoring and alarm" etc. It represents the executable entity of a tool, which encapsulates the specific functional implementation logic. It can be a local function, a remote API call, or a microservice interface. The input specifications for the tool define the set of parameters, parameter types, and format requirements needed to invoke the skill. The output specification of the tool defines the format and content structure of the results returned after the skill is executed.

[0079] In one possible implementation, it also includes:

[0080] If the output specification of any one of the tools completely matches the input specification of another tool, then the two tools are combined sequentially.

[0081] When multiple tools do not have data dependencies on each other, they are combined into a tool portfolio in parallel.

[0082] During the process of querying and scheduling the large language model, the tool combination can be queried and scheduled.

[0083] For example, given two skills and ,like Output specifications and The input specification is a complete match, that is Then the two can form a sequential combination relationship:

[0084] ;

[0085] Its combinatorial semantics is to execute first. , and use its output as The input continues to be executed, and the final output is The execution result; This means that... As Input.

[0086] For multiple skills that do not have data dependencies on each other, a parallel composition relationship can be formed:

[0087] ;

[0088] Parallel composition supports the simultaneous execution of multiple skills, each outputting an independent set of results, which are then merged through a result aggregation mechanism. This modular skill architecture allows the complex operational capabilities of low-altitude management to be broken down into combinations and orchestrations of atomic skills, achieving high reusability of functions and system scalability.

[0089] In one possible implementation, the low-altitude airspace management task request is input into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task, including:

[0090] The low-altitude airspace management task request is input into a large language model for processing, enabling the large language model to query the attribute information of each tool in the tool registration list; the attribute information includes at least the input requirement set, the output commitment set, and the functional description.

[0091] When the low-altitude airspace management task request matches the input requirement set, output commitment set, and functional description of any tool in the tool registration list, the executableness of the low-altitude airspace management task is determined to be executable; otherwise, it is determined to be unexecutable.

[0092] For example, suppose the low-altitude management task request is The tool registration list is as follows any of these tools The attribute information includes at least the input requirement set. Output commitment set .

[0093] The large language model determines the low-altitude management task sequentially using the following quantification criteria. Can the registration list be accessed through the tool? The tools in the toolkit are combined to complete the process:

[0094] (1) Determination of the executability of individual tools;

[0095] For any tool It is capable of handling the task independently. The necessary and sufficient condition is:

[0096] ;

[0097] in, This indicates that the task has been completed. The required set of input information Indicates the task The expected set of output information. Indicates task With tools Exact match, that is: the task All the required input information is in the tool Within the input requirement set, and the tool The output information that can be provided can cover the task. The expected output. When the match rate is 1.0, the tool can be used directly to complete the task; This indicates that both the left and right conditions are true simultaneously.

[0098] (2) Multi-toolchain executability determination;

[0099] If the current task If a task cannot be completed independently by a single tool, it is necessary to determine whether an execution sequence consisting of multiple tools exists. Able to complete the task The tool executes a sequence. The necessary and sufficient conditions for feasibility are:

[0100] The first tool input meets the following conditions:

[0101] ;

[0102] Adjacent tool connection conditions (i.e., the output of the previous tool can be used as the input of the next tool):

[0103] ;

[0104] in Representation tools The set of prerequisite input information necessary for it to perform its function. Representation tools The set of output information that can be provided, and the connection conditions between adjacent tools. The quantification is: the requirement coverage of the output of the preceding tool and the input of the subsequent tool is ≥100%, that is, the output of the preceding tool can fully meet all the input requirements of the subsequent tool.

[0105] The final tool output meets the condition (i.e., the output of the last tool in the sequence can cover the task). (All expected outputs)

[0106] ;

[0107] (3) Comprehensive quantitative score of feasibility;

[0108] The large language model is for the low-altitude management task. With the tool list The feasibility of generating a comprehensive matching score The calculation method is as follows:

[0109] ;

[0110] in, The maximum matching degree of a single tool, i.e. Quantify the degree of matching between the current task and a single tool. For chain-link matching degree, i.e. The smoothness of input-output connections between adjacent tools in a quantitative toolchain is measured, and the minimum value among all connections is taken as the overall chain matching degree. and For the preset weighting coefficients, satisfy This is used to adjust the weight of single-item matching and chain matching in the overall score.

[0111] (4) Feasibility determination threshold criteria;

[0112] like Then determine the low-altitude management task. In the tool registration list Executable in, among which This is a preset executability threshold, with a value range of [0,1].

[0113] like Then determine the low-altitude management task. In the tool registration list This is not executable.

[0114] (5) Optimal tool execution sequence selection;

[0115] When the low-altitude management task is determined When executable, the large language model further selects the tool sequence with the highest comprehensive matching score from all feasible tool execution sequences as the execution plan for the task:

[0116] ;

[0117] in, The set of all tool execution sequences that satisfy the above executability criteria. For the task In the tool execution sequence The matching score is as follows.

[0118] In one possible implementation, the core service interface of the abstraction layer containing the large language model is set as follows:

[0119] ;

[0120] in, This indicates a service for generating large language models. The prompt information indicates the input to the large language model, and the prompt information includes the semantic description and context data corresponding to the low-altitude management task request; This indicates that the large language model calls the configuration parameter set. This indicates the response returned by the large language model. This indicates that the function maps its input to its output.

[0121] In one possible implementation, the large language model call configuration parameter set is set as follows:

[0122] ;

[0123] in, Indicates temperature parameter; Indicates the maximum number of output tokens; Indicates kernel sampling parameters; Indicates the frequency penalty coefficient; This indicates the presence of a penalty coefficient.

[0124] The response returned by the large language model can be:

[0125] ;

[0126] in, This represents the text content generated by LLM; This indicates the resource consumption statistics for this call, including the number of input tokens and the number of output tokens; Indicates the response time of this call.

[0127] By setting the core service interface of the large language model at the abstraction layer, the orchestration layer and the underlying heterogeneous LLM inference engine are decoupled, and a unified LLM abstraction service layer is constructed. This layer defines a standardized set of calling interfaces, shielding the differences between different LLM providers in API style, authentication methods, parameter naming, etc., and providing consistent model service capabilities to the upper layer.

[0128] In one possible implementation, based on the mapping relationship between tasks and tools and the tool registration list, tools are sequentially scheduled to execute each subtask, and the execution results of the subtasks are obtained, including:

[0129] The initialization context information is as follows: ;in, This indicates the context information for initialization. Represents the empty set;

[0130] For each subtask, execution is performed based on the mapping relationship between the tasks and tools, as well as the tool registration list:

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] in, This represents the k-th tool. Let j represent the j-th subtask, j = 1, 2, ..., m; m represent the total number of subtasks. This represents the k-th tool. Let j represent the j-th subtask, where j = 1, 2, ..., m; and m represent the total number of subtasks. This indicates the task and skill mapping operation. This represents the input data for the j-th subtask. This indicates an input extraction operation. This indicates the context after the (j-1)th step is executed. This represents the output of the j-th subtask. Indicates the execution of a skill. This represents the context after step j is executed;

[0136] All subtasks corresponding Together they serve as the result of the subtask execution.

[0137] In one possible implementation, the execution results of all subtasks are aggregated into the large language model, so that the large language model can perform comprehensive analysis, semantic understanding, and logical integration on the execution results of all subtasks to generate the final execution result:

[0138] ;

[0139] in, This indicates the final execution result. Representing a large language model, This indicates a summary / generation operation. This indicates that the last subtask has been completed, and the result includes the output of all subtasks. This indicates the original user task request.

[0140] Compared with the prior art, this application has the following beneficial effects:

[0141] (1) This application takes a large language model as the core of intelligent orchestration. Through task processability judgment, intelligent decomposition and dynamic orchestration mechanism, the system can autonomously understand the intent of complex tasks and flexibly call various functional tools, which significantly improves the ability to respond to complex scenarios and emergencies, and realizes the leap from fixed processing to intelligent adaptation.

[0142] (2) This application adopts a modular skill architecture, which encapsulates various low-altitude management functions into tools that can be independently registered and called, thereby decoupling business functions from scheduling logic. When adding, replacing or upgrading functions, only the tool list needs to be operated, without modifying the core code, which greatly reduces the system development, testing and maintenance costs, while enabling the system to quickly respond to business changes and technological iterations in the low-altitude economy field;

[0143] (3) This application uses LLM to intelligently decompose complex tasks, enabling concurrent scheduling and execution of parallelizable subtasks, which effectively shortens the overall task response time. At the same time, the unified LLM abstraction layer supports flexible switching of heterogeneous inference engines, which can achieve a balance between cost and performance according to business needs, thereby improving the efficiency of system resource utilization.

[0144] like Figure 2 As shown, based on the same inventive concept, this application also provides an intelligent scheduling system for low-altitude management tasks based on a large language model, comprising:

[0145] Interaction layer 201 is used to obtain low-altitude management task requests submitted by users through human-computer interaction; the low-altitude management task requests are natural language task requests describing low-altitude management.

[0146] The orchestration layer 202 is used to input the low-altitude airspace management task request into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task; the executability is either executable or not executable. If the executability is not executable, a preset web search engine is invoked to attempt an answer, or a response indicating that the request cannot be processed is directly returned to the user. If the executability is executable, the large language model is scheduled to perform semantic parsing and logical decomposition of the low-altitude airspace management task request to obtain multiple sub-tasks with execution order logic. The large language model is then scheduled to match the multiple sub-tasks with execution order logic with corresponding tools to be invoked, forming a mapping relationship between sub-tasks and tools. Based on the mapping relationship between tasks and tools and the tool registration list, tools are sequentially scheduled to execute each sub-task, obtaining the sub-task execution results. All sub-task execution results are aggregated into the large language model, allowing the large language model to perform comprehensive analysis, semantic understanding, and logical integration of all sub-task execution results, generating a final execution result, which is then returned to the user.

[0147] Tool layer 203 is used to encapsulate tools for handling discrete tasks in the low-altitude management field and maintain a tool registration list;

[0148] Infrastructure layer 204 is used to provide computing power support for the interaction layer, tool layer and orchestration layer, as well as large language model inference capabilities.

[0149] For example, the interaction layer 201 is located at the top layer of the architecture and provides end users with a variety of task access methods, including but not limited to web interactive interface, RESTful API interface, command line operation interface and mobile application, to receive low-altitude management task requests submitted by users and present the final processing results to users.

[0150] The orchestration layer 202 is the core intelligent hub of this application, consisting of large language model instances. This layer is responsible for semantic understanding of task requests, traversal scanning of the tool list, logical judgment of task processability, intelligent decomposition and execution sequence orchestration of complex tasks, and comprehensive aggregation and generation of multi-source execution results.

[0151] The tool layer 203 encapsulates various discrete business capabilities in the low-altitude airspace management field, implementing them as independently callable functional modules, i.e., single-function tools. These include, but are not limited to, intelligent route planning tools, real-time weather data query tools, airspace conflict detection and early warning tools, aircraft status monitoring tools, and emergency command generation and issuance tools. Each tool registers its functional metadata information in the tool list, allowing the orchestration layer to match and invoke it during task decomposition.

[0152] The infrastructure layer 204 provides underlying computing power support and LLM inference capabilities for the entire system. This layer provides a unified encapsulation of heterogeneous LLM inference engines, supporting both cloud-based large model services and locally deployed open-source large model inference engines. It provides stable and efficient LLM call services for the orchestration layer, achieving decoupling between inference capabilities and business logic.

[0153] This application defines each independently callable functional unit in the tool layer as a skill. A skill is a fundamental atomic module constituting the low-altitude air traffic control task processing capability; each skill encapsulates specific business logic and functional implementation. From a formal perspective, a skill can be represented by the following six-tuple structure:

[0154] ;

[0155] in: A unique identifier for a skill, used to index and invoke that skill in the tool list; It is a natural language functional description of a skill, used by LLM to understand the functional scope of the skill when decomposing tasks; The skill type label is used to identify the low-altitude business area to which the skill belongs, such as "flight planning", "meteorological services", "monitoring and alarm" etc. It is an executable entity for a skill, which encapsulates the specific functional implementation logic. It can be a local function, a remote API call, or a microservice interface. The input specifications for the skill define the set of parameters, parameter types, and format requirements required to invoke the skill. The output specifications for skills define the format and content structure of the results returned after the skill is executed.

[0156] This application supports multiple skill combination modes. Given two skills... and ,like Output specifications and The input specification is a complete match, that is Then the two can form a sequential combination relationship:

[0157] ;

[0158] Its combinatorial semantics is to execute first. , and use its output as The input continues to be executed, and the final output is The execution result, express The input. For multiple skills that do not have data dependencies on each other, a parallel composition relationship can be formed:

[0159] ;

[0160] Parallel composition supports the simultaneous execution of multiple skills, each outputting an independent set of results, which are then merged through a result aggregation mechanism. This modular skill architecture allows the complex operational capabilities of low-altitude management to be broken down into combinations and orchestrations of atomic skills, achieving high reusability of functions and system scalability.

[0161] To decouple the orchestration layer from the underlying heterogeneous LLM inference engines, this application constructs a unified LLM abstraction service layer. This layer defines a standardized set of calling interfaces, shielding the differences between different LLM providers in API style, authentication methods, parameter naming, etc., and providing consistent model service capabilities to the upper layers.

[0162] The core service interface of the LLM abstraction layer is defined as follows:

[0163] ;

[0164] in: The prompts input to the large language model include a semantic description of the user's task and necessary contextual data;

[0165] The formal representation of calling the configuration parameter set for the model is as follows:

[0166] ;

[0167] The meanings of each parameter are as follows: The temperature parameter controls the randomness of the generated results, with a value range of [0,1]. To maximize the number of output tokens, the length of the generated result is limited; The kernel sampling parameter controls the probability accumulation threshold of candidate words during generation; This is a frequency penalty coefficient, which reduces the probability of repeatedly generating the same content; The existence of a penalty coefficient encourages the model to introduce new topics.

[0168] The formal representation of the response returned by LLM is as follows:

[0169] ;

[0170] in: Text content generated for LLM; This provides statistics on resource consumption for this call, including the number of input tokens and the number of output tokens. This is the response time for this call.

[0171] Through this abstract service layer, the orchestration layer can call different LLM backend implementations in a unified way, for example:

[0172] ;

[0173] ;

[0174] When switching the underlying LLM engine, only the configuration items need to be modified, without changing the business logic code of the orchestration layer, thus achieving flexible adaptation of LLM inference capabilities and vendor independence.

[0175] This application uses an intelligent orchestrator as the core execution engine, responsible for driving and controlling the entire task processing flow. The orchestrator receives user task requests, coordinates LLM inference, skill invocation, and result aggregation, forming a complete task processing closed loop. The execution logic of the orchestrator can be formally described as the following mapping relationship:

[0176] ;

[0177] in: Natural language task requests input by the user; A collection of registered tools, containing all available skills. ; Provides a unified inference call capability for the large language model service interface; This is the final output result.

[0178] The complete execution flow of the orchestrator is defined by the following control functions:

[0179] (1) Task Planning. Input the task request Q and the toolset into the LLM, which will perform task analysis and output the task execution plan:

[0180] ;

[0181] in, It is a structured object containing a sequence of subtasks. And each sub-task and its corresponding skill mapping relationship .

[0182] (2) Task execution.

[0183] Traverse the sequence of subtasks and maintain the execution context. Used to pass intermediate results:

[0184]

[0185]

[0186]

[0187]

[0188]

[0189]

[0190]

[0191] (3) Result synthesis.

[0192] Once all subtasks have been executed, the complete context will be... enter Summarize the results and generate the final result. :

[0193] ;

[0194] Through the aforementioned orchestration mechanism, this application achieves intelligent end-to-end driving from task input to result output. The orchestrator, as the core scheduling engine, organically combines the inference capabilities of the LLM with the execution capabilities of the tool layer, enabling complex low-altitude airspace management tasks to be processed efficiently, orderly, and dynamically.

[0195] The beneficial effects of the intelligent scheduling system for low-altitude management tasks based on a large language model provided in this application are similar to those of the methods and technologies described above, and will not be repeated here.

[0196] like Figure 3 As shown, based on the same inventive concept, this application also provides an electronic device, including a processor 302 and a memory 301; the memory 301 and the processor 302 are interconnected via a bus 303.

[0197] The memory 301 stores computer-executed instructions;

[0198] The processor 302 executes the computer execution instructions stored in the memory 301, causing the processor 302 to execute a low-altitude management task intelligent scheduling method based on a large language model as described in any embodiment of this application.

[0199] For specific examples, memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). Furthermore, the processor may include a main processor and coprocessors. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0200] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the intelligent scheduling method for low-altitude management tasks based on a large language model as described in any of the above embodiments.

[0201] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent scheduling method for low-altitude management tasks based on a large language model as described in any of the above embodiments.

[0202] All or part of the steps in the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable memory. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof.

[0203] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processing unit of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0204] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0205] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0206] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed 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 examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0207] 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 method for intelligent scheduling of low-altitude air traffic control tasks based on a large language model, characterized in that, include: Obtain low-altitude management task requests submitted by users through human-computer interaction; The low-altitude airspace management task request is a natural language task request describing low-altitude airspace management. The low-altitude airspace management task request is input into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task. The executability is either executable or not executable. If the executability is deemed non-executable, a preset web search engine will be invoked to retrieve information in an attempt to answer the question, or a response message indicating that the request cannot be processed will be directly returned to the user. If the executability is executable, then the large language model is scheduled to perform semantic parsing and logical decomposition on the low-altitude management task request to obtain multiple sub-tasks with execution order logic; The large language model is scheduled to match the corresponding tools to be invoked for the multiple subtasks with execution order logic, forming a mapping relationship between subtasks and tools; Based on the mapping relationship between tasks and tools and the tool registration list, tools are scheduled to execute each subtask in sequence, and the execution results of the subtasks are obtained. The execution results of all subtasks are aggregated into the large language model, which then performs comprehensive analysis, semantic understanding, and logical integration on the execution results of all subtasks to generate the final execution result, which is then returned to the user.

2. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 1, characterized in that, The tool is configured with a six-tuple structure, specifically: ； in, Representation tools, This represents a unique identifier corresponding to the tool. This represents the natural language function description corresponding to the tool. The type label indicates the tool. This represents the executable entity of the tool. Indicates the input specifications for the tool. This indicates the output specifications of the tool.

3. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 2, characterized in that, Also includes: If the output specification of any one of the tools completely matches the input specification of another tool, then the two tools are combined sequentially. When multiple tools do not have data dependencies on each other, they are combined into a tool portfolio in parallel. During the process of querying and scheduling the large language model, the tool combination can be queried and scheduled.

4. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 1, characterized in that, The low-altitude airspace management task request is input into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task, including: The low-altitude airspace management task request is input into a large language model for processing, enabling the large language model to query the attribute information of each tool in the tool registration list; the attribute information includes at least the input requirement set, the output commitment set, and the functional description. When the low-altitude airspace management task request matches the input requirement set, output commitment set, and functional description of any tool in the tool registration list, the executableness of the low-altitude airspace management task is determined to be executable; otherwise, it is determined to be unexecutable.

5. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 1, characterized in that, The core service interface of the abstraction layer containing the large language model is set as follows: ； in, This indicates a service for generating large language models. The prompt information indicates the input to the large language model, and the prompt information includes the semantic description and context data corresponding to the low-altitude management task request; This indicates that the large language model calls the configuration parameter set. This indicates the response returned by the large language model. This indicates that the function maps its input to its output.

6. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 5, characterized in that, The large language model call configuration parameter set is set as follows: ； in, Indicates temperature parameter; Indicates the maximum number of output tokens; Indicates kernel sampling parameters; Indicates the frequency penalty coefficient; This indicates the presence of a penalty coefficient.

7. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 1, characterized in that, Based on the mapping relationship between tasks and tools and the tool registration list, tools are sequentially scheduled to execute each subtask, and the execution results of the subtasks are obtained, including: The initial context information is as follows: ;in, This indicates the context information for initialization. Represents the empty set; For each subtask, execution is performed based on the mapping relationship between the tasks and tools, as well as the tool registration list: ； ； ； ； in, This represents the k-th tool. Let j represent the j-th subtask, where j = 1, 2, ..., m; and m represent the total number of subtasks. This indicates the task and skill mapping operation. This represents the input data for the j-th subtask. This indicates an input extraction operation. This indicates the context after the (j-1)th step is executed. This represents the output of the j-th subtask. Indicates the execution of a skill. This represents the context after step j is executed; All subtasks corresponding Together they serve as the result of the subtask execution.

8. The intelligent scheduling method for low-altitude management tasks based on a large language model according to claim 1, characterized in that, The execution results of all subtasks are aggregated into the large language model, which then performs comprehensive analysis, semantic understanding, and logical integration on the results to generate the final execution result: ； in, This indicates the final execution result. Representing a large language model, This indicates a summary / generation operation. This indicates that the last subtask has been completed, and the result includes the output of all subtasks. This indicates the original user task request.

9. A low-altitude air traffic control task intelligent scheduling system based on a large language model, characterized in that, include: The interaction layer is used to obtain low-altitude management task requests submitted by users through human-computer interaction; The low-altitude airspace management task request is a natural language task request describing low-altitude airspace management. The tool layer encapsulates tools for handling discrete tasks in the low-altitude airspace management field and maintains a tool registry list. The orchestration layer is used to input the low-altitude airspace management task request into a large language model for processing. The large language model queries the attribute information of each tool in the tool registration list to determine the executability of the low-altitude airspace management task. The executability is either executable or not executable. If the executability is not executable, a preset web search engine is invoked to retrieve information in an attempt to answer the question, or a response message indicating that the request cannot be processed is directly returned to the user. If the executability is executable, the large language model is scheduled to perform semantic parsing and logical decomposition on the low-altitude airspace management task request to obtain multiple subtasks with execution order logic. The large language model is scheduled to match the corresponding tools to be invoked for the multiple subtasks with execution order logic, forming a mapping relationship between subtasks and tools; according to the mapping relationship between tasks and tools and the tool registration list, the tools are scheduled to execute each subtask in sequence to obtain the execution results of the subtasks; The execution results of all subtasks are aggregated into the large language model, so that the large language model can perform comprehensive analysis, semantic understanding and logical integration of the execution results of all subtasks, generate the final execution result and return it to the user; The infrastructure layer provides computing power support for the interaction layer, tool layer, and orchestration layer, as well as large language model inference capabilities.

10. An electronic device, characterized in that, Including processor and memory; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the intelligent scheduling method for low-altitude management tasks based on a large language model as described in any one of claims 1 to 8.

Images