Low-altitude flight intelligent decision-making method based on atlas retrieval enhanced generation and digital flight rules

Abstract

The invention discloses a low-altitude flight intelligent decision-making method based on atlas retrieval enhanced generation and digital flight rules. The method comprises the following steps: firstly, preprocessing and translating a digital flight rule (DFR) text, extracting entities and relationships through a large language model, constructing a structured knowledge graph, and storing the structured knowledge graph in a vector database; and secondly, circularly calling the large model for three times: calling an optimization cue word to generate a template for the first time, calling a template-based analysis DFR text to complete atlas construction for the second time, and calling a combination atlas retrieval capability to execute compliance verification and route intelligent decision-making of a low-altitude flight task for the third time. In addition, when an airspace rule is changed or a temporary announcement exists, real-time updating of knowledge is achieved by inserting a new node into the knowledge graph or modifying edge attributes. According to the method, GraphRAG and DFR combination is generated through map retrieval enhancement, the interpretability, accuracy and dynamic adaptability of large model decision making are improved, and the problem of complex decision making in a low-altitude environment is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence and intelligent decision-making systems, in particular to a low-altitude flight intelligent decision-making method based on graph retrieval-augmented generation (Graph RAG) and digital flight rules (DFR). BACKGROUND

[0002] Current unmanned aerial vehicle flight decisions are mostly based on fixed rule bases, relying on manual path planning or path planning techniques based on deep reinforcement learning to achieve flight task planning and decision-making. Traditional decision-making methods based on rule bases are based on artificially set airspace regulations, flight restrictions, and other conditions, and construct rule tables or logical rule trees, which are matched and verified through logical reasoning, decision trees, or expert systems. This type of system is widely used in early UTM (Unmanned Traffic Management) and U-Space prototypes, but its scalability and adaptability are limited. Decision-making systems based on path planning algorithms are the classic path planning methods used by most current low-altitude flight scheduling platforms, such as three-dimensional path decision-making based on A* algorithm, Dijkstra algorithm, and Rapidly Exploring Random Tree (PRT) algorithm, which generates routes in combination with geographic information data and airspace obstacle information. This type of method has a clear structure and stable execution, but it responds slowly to dynamic environments and is difficult to handle changes in semantic rules. In addition, in recent years, path recommendation systems based on reinforcement learning (such as DQN, PPO algorithm, etc.) or graph neural networks (Graph Neural Network, GNN) have begun to be applied to intelligent route generation and obstacle avoidance decision-making. These systems can learn complex flight patterns to some extent, but they have high training costs, poor interpretability, insufficient rule constraints, weak scene migration ability, and other problems, and have not yet formed a stable, controllable, and industry-level application solution. Due to the above limitations, the current unmanned aerial vehicle flight task decision-making field has not yet formed a unified, efficient, and multi-source information fusion intelligent decision-making framework.

[0003] With the rapid development of urban air mobility (UAM) and large-scale unmanned aerial vehicle (UAV) flights, low-altitude flights face increasingly complex airspace structures, space-time conflicts, and regulatory adaptation problems. Traditional flight decision-making methods relying on manual rule design or local algorithm optimization have been difficult to meet the real-time needs of large-scale coordination, dynamic adjustment, and intelligent interpretation. In this context, large model-driven intelligent decision-making technology has been introduced into low-altitude intelligent management and flight decision-making systems. At present, the application of large models in low-altitude flight rule intelligent decision-making is still in its infancy, and a perfect systematic architecture has not yet been formed. How to improve the generalization ability of large models, solve the knowledge update mechanism, and build a unified semantic representation and rule expression method is the key to meeting the real-time needs of low-altitude complex airspace.

[0004] The existing technology is usually based on a fixed rule base, and its working principle is as shown in Figure 1 .

[0005] The low-altitude flight rule decision principle based on traditional methods: ① The decision method relying on the rule base relies on pre-defined airspace management rules and logical judgment mechanism to realize static verification and path restriction identification of flight compliance. ② The decision method based on path planning algorithm generates the shortest or optimal flight path that meets the constraint conditions by constructing an airspace spatial model and calling a classical path search algorithm. ③ The decision method based on reinforcement learning or graph neural network realizes intelligent path selection in dynamic environment by training intelligent agent models or graph structure models to learn flight strategies or route scoring rules.

[0006] The working principle of the emerging large model flight decision-making technology is as shown in Figure 2 .

[0007] The low-altitude flight rule decision-making process based on large models: ① Clearly define task requirements and build prompt words; ② Complete the preprocessing of input information and supplement context information; ③ Call the large language model to generate responses; ④ Structured structure analysis and output standardization; ⑤ Collaborative output confirmation and decision adoption.

[0008] The existing flight decision-making method lacks the ability to adapt to dynamic environments and complex tasks, making it difficult to respond to airspace changes in a timely manner, resulting in low efficiency and potential flight safety hazards. At the same time, with the gradual opening and development of low-altitude airspace, unmanned aerial vehicles and other low-altitude aircraft are showing a trend of diversification, multi-type, and high-density coexistence, bringing complex airspace coordination, route planning management, and rule adaptation problems. The flight decision-making based on traditional methods is difficult to adjust in real time according to the airspace, has poor adaptability to complex tasks, and lacks the ability to link reasoning.

[0009] Firstly, the current low-altitude flight activities present a situation of diverse types, explosive quantity and complex scenes, and the fine management of airspace resources and the intelligent decision of flight rules are facing unprecedented challenges. In practical applications, unmanned aerial vehicles often need to deal with dynamic airspace setting, temporary flight restricted areas, high-density route conflicts and other complex factors, which puts higher requirements on the understanding and compliance judgment of flight rules. However, the flight rule decision method under the existing technical system still mainly depends on the rule library set by artificial, classical path planning algorithm or local intelligent model, and these methods have many limitations.

[0010] Secondly, although large language models (also known as large models, LLM) have shown great ability in natural language processing and knowledge reasoning tasks, there are still many key deficiencies in their application in the field of intelligent decision of low-altitude flight rules, which cannot meet the high standard requirements of actual aviation safety management. The "black box" characteristics of large models make the decision-making process highly complex and invisible, and the reasoning chain of the model is difficult to trace. Such low interpretability is still tolerable in ordinary text generation tasks, but it is fatal in low-altitude flight scenarios involving public safety and personal risk.

[0011] Thirdly, the application of large models in the low-altitude field mostly still stays in the stage of general pre-training on open source corpus and aviation regulation text, lacking of standardized flight rule system. This leads to the fact that the model-generated content often stays on the surface, lacks of precise compliance and technical depth, and is difficult to perform professional rule judgment tasks. At the same time, the low-altitude flight scene itself is highly dynamic, such as temporary flight restricted announcements, weather changes, regional adjustments, etc. The existing large models have not established an efficient and controllable knowledge updating mechanism, and it is difficult to integrate the latest regulations or airspace status in time, resulting in poor performance in new scene adaptation and task migration.

[0012] It should be noted that the information disclosed in the above background section is only for understanding the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0013] The main purpose of the present application is to overcome the defects in the above background art, and to provide a low-altitude flight intelligent decision method based on graph retrieval enhanced generation and digital flight rules.

[0014] To achieve the above purpose, the following technical solutions are adopted in the present application: A low-altitude flight intelligent decision method based on graph retrieval enhanced generation and digital flight rules, comprising the following steps: S1, constructing a digital flight regulation knowledge graph: based on a digital flight regulation (DFR) text, extracting entities and relationships between entities by a large language model, constructing a structured knowledge graph, and storing the knowledge graph in a vector database; S2, calling a large language model in stages to provide decision support: First call: performing a prompt word optimization project to generate a prompt word template containing a task target description and an output format example; Second call: based on the optimized prompt word template, driving the large language model to perform structured analysis on the DFR text, completing entity extraction, relationship identification, and knowledge graph construction; Third call: combining the knowledge graph's search enhancement capabilities, driving the large language model to perform compliance verification and route intelligent decision-making for low-altitude flight tasks; S3, dynamic updating: responding to airspace rule changes or temporary announcements, updating knowledge in real time by inserting new nodes or modifying edge attributes in the knowledge graph.

[0015] Further, the step S1 of constructing the knowledge graph specifically includes: Pretreating the digital flight regulation (DFR) text, cleaning up redundant information irrelevant to flight regulations, and generating high-quality DFR corpus; Dividing the DFR corpus into structured text units; Extracting entities from each text unit by a large language model and aggregating entities with similar semantics; Identifying relationships between entities based on syntactic and semantic features of text units, and constructing an initial graph structure by representing entities as nodes and relationships as edges; Clustering nodes according to semantic similarity, forming a topic hierarchical structure, and recording original text segments to establish traceability of graph elements.

[0016] Further, constructing the knowledge graph also includes relationship fusion processing: Fusing the identified relationships between entities, including merging duplicate relationship edges pointing to the same entity pair and standardizing attribute definitions of edges; Establishing interactive relationship links between entities in different topic clusters to support complex compliance reasoning involving multiple topic entities.

[0017] Further, in step S2, the first call specifically includes: Inputting the DFR corpus into the large language model to generate a summary description of the corpus topic and a subsequent task target; Randomly extracting text segments from the DFR corpus to drive the large language model to generate triple examples containing entities and relationships as a format reference for subsequent entity relationship extraction.

[0018] Further, the prompt word optimization process specifically includes: Analyzing the subject field attributes of the DFR corpus; Embedding randomly selected text segments as example dialogues into the prompt word structure to guide the large language model to understand the output format requirements of entity relationship triples.

[0019] Further, in step S2, the second call specifically includes: Inputting the optimized prompt word template generated by the first call and the complete preprocessed DFR text into the large language model; Driving the large language model to segment the text into structured units and identify and extract entities in each unit; Driving the large language model to analyze the relationships between entities, generate associated variables, and semantically bind the text units with the identified relationships; Based on the extracted entities and bound relationships, create graph nodes and edges to form an initial knowledge graph.

[0020] Further, in step S2, the third call specifically includes: Inputting the current low-altitude flight task information into the large language model; Retrieval enhancement combined with the knowledge graph: based on the semantic vectorization of task information, retrieve matching rule fragments, related entities, and compliance constraint conditions from the knowledge graph; Driving the large language model to make inference decisions based on the retrieval enhancement results, the inference decision process includes: a) Global summary: making a summary judgment on the overall compliance of the task; b) Local retrieval: locating and applying rule fragments in the knowledge graph directly related to specific constraint conditions; c) Comprehensive generation: integrating global summary and local retrieval results to generate the final flight compliance conclusion and route decision suggestions.

[0021] Further, in step S3, the dynamic update specifically includes: When flight regulations change or temporary airspace notices occur, dynamically adjust the content of the knowledge graph by inserting nodes representing new rules or states into the knowledge graph, or modifying the attribute values of existing node edges, to adapt to the latest airspace environment and rule requirements.

[0022] Further, the method further includes: Visual interaction: importing the constructed or updated knowledge graph into a graph database system to realize visual display, and supporting query and editing operations on nodes and relationships through interfaces; The visual interaction specifically includes: Converting the knowledge graph data into a structured data format file; Importing the structured data format file into a graph database system; Displaying the knowledge graph through a visualization interface provided by the graph database system; Through the operation interface or database operation toolkit provided by the graph database system, the query, addition, deletion or modification operation of the node and the relationship is executed.

[0023] A computer program product comprising a computer program, which, when executed by a processor, implements the low-altitude flight intelligent decision-making method based on graph retrieval enhancement and digital flight rules.

[0024] The present application has the following beneficial effects: The present application provides a low-altitude flight intelligent decision-making method based on graph retrieval enhancement and digital flight rules, which constructs a structured and traceable low-altitude flight rule knowledge graph by deeply integrating the Graph RAG framework and the digital flight rules DFR, and adopts a strategy of calling a large language model in stages, which significantly improves the explainability, compliance and dynamic adaptability of low-altitude flight intelligent decision-making. Compared with the traditional method of relying on static rule library, path planning algorithm or local intelligent model, the present application effectively overcomes the defects of the prior art, such as slow response to dynamic environment, fragmented rule reasoning and "black box" decision-making of large models. Specifically, the introduction of the Graph RAG architecture enables the large model to have structured retrieval and multi-step reasoning capabilities based on the knowledge graph while retaining the advantages of natural language processing, and the decision-making process can be traced back to the graph nodes, greatly enhancing the transparency and credibility; the knowledge graph constructed based on the DFR as a standardized rule system provides clear legal basis and semantic support for model decision-making, ensuring that the generated path planning and conflict avoidance scheme conforms to the core constraints of airspace speed, height, priority and interval; the high maintainability of the knowledge graph supports real-time response to regulatory changes and temporary airspace announcements through node insertion or edge attribute modification, solving the bottleneck of traditional models that static knowledge bases are difficult to adapt to low-altitude high-time-varying environments, and combining with the visual interaction mechanism, forming a new paradigm of low-altitude intelligent decision-making that is explainable, strongly generalizable and dynamically updated.

[0025] The application realizes decision generation based on knowledge graph retrieval enhancement by using the Graph RAG framework as a retrieval enhancement module of a large model, taking DFR as the basis and standard for intelligent decision-making of large model low-altitude flight rules, thereby improving the generalization ability of the large model, solving the knowledge update mechanism, constructing a unified semantic representation and rule expression method, and meeting the real-time needs of low-altitude complex airspace. The application of a large language model in the field of low-altitude flight rule decision-making makes use of its powerful general language understanding and generation ability, cross-scenario task migration ability, and complex reasoning and logic ability to realize flight rule decision-making with strong generalization ability and real-time performance. The innovation contribution and important features of the application mainly include: (1) Graph RAG-based low-altitude digital flight rule knowledge graph construction method The method takes the DFR original rule text as the corpus basis, constructs a digital flight rule knowledge graph with a graph structure through large model-driven information extraction and structured organization, and embeds it in a vector database. When the large model performs low-altitude flight rule judgment and path decision tasks, it can call the graph node content for "traceable" enhanced reasoning, output more compliant flight, airspace restrictions, and task-adaptive decision text. At the same time, flight recommendations and decisions can be traced back to the rule nodes and context semantics in the graph, avoiding "black box" output, and the graph content can be dynamically updated according to new regulations or temporary announcements, with good adaptability and expandability.

[0026] (2) Large model training method and decision text generation strategy based on cyclic calling The method proposes a large model training method and decision text generation strategy based on cyclic calling, which calls the large model three times. The first time is used for prompt optimization engineering to guide the model to understand the DFR rule corpus structure and task goal and generate high-quality prompt word templates. The second time is used for knowledge graph creation to drive the model to complete entity extraction, relationship construction, and graph initialization based on optimized prompts. The third time is used for intelligent decision generation to complete low-altitude flight compliance analysis and route recommendation with the support of the knowledge graph. Through the phased and multi-task linked large model calling method, the technical bottleneck of single model function and fragmented upstream and downstream information in traditional large model application methods is broken through, and a new paradigm for flight rule intelligent decision-making is constructed.

[0027] Compared with traditional low-altitude flight decision-making methods that rely on rule libraries, path planning algorithms, or local intelligent models, the application has significant advantages in explainability, intelligence, and dynamic adaptability: (1) By introducing the Graph RAG architecture, the large language model has structured knowledge retrieval and reasoning capabilities based on natural language understanding and generation capabilities, effectively addressing the explainability defects caused by the "black box" problem of large models.

[0028] (2) The introduction of the DFR digital flight rule knowledge graph enables the model decision to have clear regulatory basis and semantic support, can realize multi-step reasoning and chain judgment at the rule level, and improves the decision accuracy and compliance.

[0029] (3) The knowledge graph database has high maintainability, supports rapid updating of dynamic information such as regulation changes and temporary flight restricted areas, realizes adaptive adjustment of the model to the dynamic airspace environment, and improves the adaptability of the model to the high time-varying characteristics of the low-altitude environment. The visualization expression and retrieval updating mechanism of the graph realized by the graph database (such as Neo4j) not only helps the supervisors to conduct rule auditing, but also facilitates the transparent presentation and responsibility tracing of the model output.

[0030] Other beneficial effects in the embodiments of the present application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a principle diagram of a traditional flight decision method based on a fixed rule base.

[0032] Figure 2 is a principle diagram of a flight decision technology method based on a large model.

[0033] Figure 3 is a flowchart of constructing a DFR digital flight rule knowledge graph based on Graph RAG in the embodiments of the present application.

[0034] Figure 4 is a comparison and analysis of the Graph DFR and the rest of the large model method in the embodiments of the present application.

[0035] Figure 5 is a general flowchart of the low-altitude flight intelligent decision method based on graph retrieval enhancement and generation of digital flight rules. DETAILED DESCRIPTION

[0036] The embodiments of the present application are described in detail below. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope of the present application and its applications.

[0037] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0038] In a low-altitude environment, unmanned aerial vehicles lack unified flight rule control and effective intelligent decision-making, which adversely affects future large-scale high-density low-altitude unmanned aerial vehicle applications. To this end, the present application proposes a low-altitude flight intelligent decision-making method based on knowledge graph retrieval enhancement generation (Graph RAG) and digital flight rules (DFR) to address complex decision-making problems in dynamic environments in the field of low-altitude aviation, especially unmanned aerial vehicles and urban air traffic. The present application uses the Graph RAG framework as a retrieval enhancement module for large language models, uses DFR as the basis and standard for large language model low-altitude flight rule intelligent decision-making, generates decisions based on knowledge graph retrieval enhancement, and uses the powerful general language understanding and generation capabilities, cross-scenario task migration capabilities, and complex reasoning and logic capabilities of large language models to achieve flight rule decisions with strong generalization capabilities and real-time performance. The present application can regulate unmanned aerial vehicle behavior in low-altitude airspace, including route, speed, altitude, priority, and spacing, and provide clear constraints for flight path planning and conflict avoidance. The DFR rules and environmental situation are represented using a knowledge graph structure to support the rapid and accurate retrieval of relevant knowledge by the large model during decision-making, enabling low-altitude flight compliance verification and efficient intelligent decision-making.

[0039] Referring to Figure 5 The present application provides a low-altitude flight intelligent decision-making method based on graph retrieval enhancement generation and digital flight rules, comprising the following steps: Step S1, constructing a digital flight rule knowledge graph: based on digital flight rule (DFR) text, extracting entities and relationships between entities using a large language model, constructing a structured knowledge graph, and storing the knowledge graph in a vector database.

[0040] In some embodiments, the construction of the knowledge graph in step S1 specifically includes: preprocessing the digital flight rule (DFR) text, cleaning up redundant information unrelated to flight rules, and generating high-quality DFR corpus; dividing the DFR corpus into structured text units; extracting entities from each text unit using a large language model and aggregating entities with similar semantics; identifying relationships between entities based on the syntax and semantic features of the text units, and constructing an initial graph structure by representing entities as nodes and relationships as edges; clustering nodes based on semantic similarity to form a topic hierarchy, and recording the original text segments to establish traceability of graph elements.

[0041] In some embodiments, the construction of the knowledge graph also includes relationship fusion processing: the identified relationships between entities are fused, including merging duplicate relationship edges pointing to the same entity pair and standardizing attribute definitions of the edges; establishing interactive relationship links between entities in different topic clusters to support complex compliance reasoning and judgment involving multiple topic entities.

[0042] Step S2, the large language model is called in stages to provide decision support: First call: execute prompt word optimization project to generate a prompt word template containing task target description and output format example; Second call: based on the optimized prompt word template, drive the large language model to perform structured analysis on the DFR text, complete entity extraction, relationship identification and knowledge graph construction; Third call: combined with the retrieval enhancement capability of the knowledge graph, drive the large language model to perform compliance verification and route intelligent decision-making of low-altitude flight tasks.

[0043] In some embodiments, in step S2, the first call specifically includes: inputting the DFR corpus into the large language model to generate a summary description of the corpus topic and a subsequent task target; randomly extracting a text segment from the DFR corpus to drive the large language model to generate a triple example containing entities and relationships as a format reference for subsequent entity relationship extraction.

[0044] The prompt word optimization project specifically includes: analyzing the subject field attributes of the DFR corpus; embedding the randomly extracted text segment as an example dialogue into the prompt word structure to guide the large language model to understand the output format requirements of the entity relationship triple.

[0045] In some embodiments, in step S2, the second call specifically includes: inputting the optimized prompt word template generated by the first call and the complete preprocessed DFR text into the large language model; driving the large language model to segment the text into structured units, identify and extract entities in each unit; driving the large language model to analyze the relationship between entities, generate associated variables and bind the text units with the identified relationships; based on the extracted entities and bound relationships, create graph nodes and edges to form an initial knowledge graph.

[0046] In some embodiments, in step S2, the third call specifically includes: inputting the current low-altitude flight task information into the large language model; combining the knowledge graph for retrieval enhancement: based on the semantic vectorization of the task information, retrieving matching rule fragments, related entities and their compliance constraint conditions from the knowledge graph; driving the large language model to make inference decisions based on the retrieval enhancement results, the inference decision-making process includes: a) Global summary: making a summary judgment on the overall compliance of the task; b) Local retrieval: locating and applying rule fragments in the knowledge graph that are directly related to specific constraint conditions; c) Comprehensive generation: integrating the global summary and local retrieval results to generate the final flight compliance conclusion and route decision-making suggestion.

[0047] Step S3, dynamic updating: in response to changes in airspace rules or temporary announcements, the knowledge is updated in real time by inserting new nodes or modifying edge attributes into the knowledge graph.

[0048] In some embodiments, in step S3, the dynamic updating specifically includes: when changes in flight regulations occur or temporary airspace announcements appear, the dynamic adjustment of the knowledge graph content is realized to adapt to the latest airspace environment and rule requirements by inserting nodes representing new rules or states into the knowledge graph or modifying the attribute values (such as weights, relationship types) of the edges between existing nodes.

[0049] In some embodiments, the low-altitude flight intelligent decision-making method based on graph retrieval enhancement and digital flight rules of the present application further includes: visual interaction: the constructed or updated knowledge graph is imported into a graph database system to realize visual display and support node and relationship query and editing operations through an interface. The visual interaction specifically can include: converting the knowledge graph data into a structured data format file (such as a parquet or csv format file); importing the structured data format file into a graph database system (such as a Neo4j graph database); displaying the knowledge graph through the visual interface provided by the graph database system; and executing the query, addition, deletion or modification operations on the nodes and relationships through the operation interface or database operation toolkit (such as the py2neo library) provided by the graph database system.

[0050] The specific embodiments of the present application are further described below.

[0051] A low-altitude flight intelligent decision-making method based on atlas retrieval enhancement generation and digital flight rules is provided. A large model training method based on cyclic calling and a decision text generation strategy are designed, as shown in Table 1. Specifically, the method steps and strategies of large model calling are as follows: first, the first call is to perform the prompt word optimization project. The digital flight rules are directly input into the large model as the corpus. The large model gives a topic summary by roughly reading the text content, generates a general description text, summarizes the role that may be played in the subsequent task and the task to be achieved, and randomly extracts text invention examples from the corpus. Roughly generate triple examples of entities and relationships between entities as text triple extraction patterns. Second, the second call realizes knowledge graph creation. The prompt word Prompt and the complete DFR are input as inputs. After calling the large model, the original text is segmented into structured units, text entities are extracted and associated variables are generated, entities are aggregated, text units and relationships are bound, initial graph structures are created based on entities and variables, graph nodes and clustering relationships are generated, and knowledge graph creation is completed. Finally, the third call is to let the large model act as an intelligent decision-making reasoning engine. Through digital flight rule knowledge graph retrieval enhancement, three types of decision text generation based on "global summary", "local retrieval" and "comprehensive generation" are realized.

[0052] In addition, in order to realize the visualization of low-altitude flight rules and improve the dynamic requirements of the atlas on the high time-varying characteristics of low-altitude scenes, based on the file conversion of parquet format and csv format, the atlas is input into Neo4j to realize the visualization of low-altitude flight rules. At the same time, through the py2neo library, the query of nodes and relationships can be realized, and the addition and deletion can be realized according to the demand. This scheme can effectively improve the explainability of low-altitude large model intelligent decision-making, generate decision-making based on standardized flight rule system based on DFR, and strengthen the time-varying updating ability of large model in low-altitude field.

[0053] Table 1 Large model training method and decision text generation strategy based on cyclic calling Large model invocation Invocation purpose Invocation steps First invocation Prompt optimization project Input the digital flight rule as a corpus into the large model, give a summary of the theme by roughly reading the text content, and generate a general description text, summarize the possible role and task to be achieved in the subsequent task. By randomly extracting text invention examples, roughly generate entity, entity relationship triple examples as subsequent text triple extraction style. Second invocation Knowledge graph creation Take the previously constructed Prompt and complete DFR as input, after loading the text by invoking the large model, by dividing the original text into structured units, extracting text entities and generating associated variables, and aggregating entities, the binding of text units and relationships is achieved. Based on entities and variables, create an initial graph structure and generate graph nodes and clustering relationships to create a knowledge graph. Third invocation Intelligent decision-making of graph As an inference engine for intelligent decision-making, the large model is enhanced through the retrieval of the digital flight rule knowledge graph, enabling the generation of decision-making text based on three categories: "global summary", "local retrieval", and "comprehensive generation". The process of constructing the DFR digital flight rule knowledge graph based on Graph RAG is shown in Figure 3 The main steps are as follows: 1. DFR digital flight rule text processing The current widely used DFR document is mainly compiled and published by NASA, written in English, covering a series of flight rules and operation standards related to route planning, airspace setting, flight restrictions, weather response, etc. in low-altitude flight. In order to adapt to the intelligent decision-making of low-altitude flight rules in the Chinese context, the English DFR file needs to be translated into high-quality professional translation. At the same time, the original DFR document contains a large amount of redundant information unrelated to actual flight rules, such as author's personal information, participating unit list, historical version description, and thanks, etc. If these contents are not cleaned directly into the large model training or reasoning process, it is easy to cause the "hallucination" problem. Filtering means is used to filter and clean the original text content. Finally, high-quality DFR Chinese flight rule corpus is generated, which provides data basis for subsequent knowledge graph construction, large model rule understanding, compliance judgment and decision-making reasoning.

[0054] 2、Prompt optimization project Prompt is the input controller of large model interaction, and its design quality directly determines the accuracy, relevance and stability of model output. In this project, the large model not only extracts entities and relationships from structured and unstructured flight rule corpus to build knowledge graph, but also serves as the core of intelligent decision-making reasoning. Therefore, the prompt not only needs to have the accuracy of information extraction, but also needs to cover the context constraints of decision-making reasoning. In the prompt optimization project, the system first analyzes the theme and language attributes of the training corpus (such as "space science", "microbiology", "hot news", etc.). Then, randomly extract fragments from the prelude as "example dialogue" to embed into the prompt structure, helping the large model understand the task goal and output format. In the "example dialogue", the model will roughly generate the entities and relationships between entities in the extracted fragments, forming a triple output format, which serves as a reference for the output of subsequent graph creation.

[0055] 3、DFR digital flight rule knowledge graph construction Driven by a large language model, prompts are used to guide it to automatically extract entities, generate relationships, and organize them into a semantically structured knowledge graph from structured and unstructured rule text. This process begins with preprocessed and translated DFR text as input. Combined with the task templates developed during the prompt optimization project, the large model is guided to perform semantic reading comprehension of the text. During the text loading and unitization phase, the original rule document is segmented into logically structured text units. The model then extracts entities from each text unit and aggregates and summarizes similar entities, improving the conciseness and consistency of information representation. Based on syntactic and semantic features, the model automatically identifies relationships between entities and semantically binds text units to form a variable association structure. During the graph creation phase, the initial graph structure is constructed based on the relationships between entities, representing entities as nodes and rule relationships as edges. Nodes are clustered and categorized based on semantic similarity and task domain themes to form a higher-level thematic hierarchy. A provenance mapping is established for each graph element by recording the original text fragments, enabling bidirectional traceability from graph to text. On this basis, a relationship fusion processing mechanism is further introduced to merge duplicate edges, standardize edge attributes, and establish cross-subject entity interaction links to support complex reasoning and multi-step compliance judgment.

[0056] 4. Graph DFR-based graph intelligent reasoning and decision-making The created DFR digital flight rules knowledge graph is embedded in the big model as a vector database, and serves as knowledge support for the big model's retrieval enhancement generation. In actual low-altitude flight decision-making tasks, the big model matches the graph nodes in the knowledge graph through semantic vectorization, retrieves the rule fragments, entity relationships and compliance restrictions related to the current task, and makes judgments and decisions after reasoning enhancement. Compared with traditional decision-making methods based on rule base matching or path algorithms, this method not only retains the advantages of the big model in natural language understanding and generation, but also significantly improves its long-chain rule reasoning capabilities through external structured knowledge graphs. At the same time, the graph database can realize real-time updates of regulatory changes or airspace notices by dynamically inserting new rule nodes or modifying edge weights, solving the limitation that the static knowledge of traditional models cannot adapt to scene changes.

[0057] Figure 4 This paper presents a comparative analysis of question-answering examples using Graph DFR, an embodiment of the present invention, and other large-scale model methods. Specific examples are as follows: Question 1: What is DFR? (A vague question about the acronym) 1. Answers from large language models: Design for Recyclability; Digital Frequency Response Demand Forecasting and Replenishment Dynamic Fault Recovery Defensive Football Role

[0058] 2. Graph RFR's answer: DFR can have multiple meanings: DFR (Digital Flight Review) is a concept and operational mode integration analysis report in the aviation field; DFR (Digital Flight Operations Regulation) refers to the digitalization of flight operations in the aviation field; DFR (Digital Flight Rules) is a framework or system that uses digital rules and automated systems to manage flight intervals during flight.

[0059] Question 2: What is the backup plan when the DFR system fails? (Search Reasoning Question) 1. Traditional RAG's answer: Existing operational modes (such as VFR (Visual Flight Rules) and IFR (Instrument Flight Rules)) can be used as backup plans. Mitigation measures include establishing appropriate design assurance levels (DAL), system redundancy, and as a last resort, collision avoidance system actions (which are functionally separate from the DFR system).

[0060] 2. Graph RFR's answer: In the event of a DFR (Digital Flight Operations) system failure, traditional flight interval and air traffic management methods are typically relied upon as backup plans.

[0061] Manual visual interval control: When the DFR system cannot provide automated interval recommendations, pilots or controllers will need to rely on personal experience and training to maintain safe flight intervals.

[0062] VFR (Visual Flight Rules) and IFR (Instrument Flight Rules) operations: In some cases, if the DFR system is unavailable, aircraft may need to switch to VFR or IFR mode for operation. VFR allows pilots to fly by visually observing the surrounding environment, while IFR relies on onboard navigation equipment and radio communication to ensure safety.

[0063] Legacy radar monitoring: Air traffic control (ATC) can continue to use radar systems to monitor the position and separation of aircraft, providing for human intervention and guidance.

[0064] Backup communication links: In the event of DFR system failure, backup voice or data communication links can be relied upon for coordination and information sharing. This includes two-way voice communication, backup versions of automated separation systems, or other auxiliary technologies.

[0065] Figure 4 The illustrated Graph RFR question and answer verification demonstrates the ability of the knowledge graph to understand professional rules, and the retrieval results are directly converted into semantic constraints for the decision engine, enabling the large model to generate flight path plans that comply with DFR specifications (such as switching to VFR mode in the event of failure) in low-altitude flight tasks, achieving rule-driven intelligent scheduling.

[0066] In summary, the present application provides a method for generating intelligent decisions for low-altitude flight based on graph retrieval and enhanced digital flight rules. Compared to traditional techniques, the innovative contributions and important features of the present application are mainly reflected in the following aspects: (1) Low-altitude digital flight rule knowledge graph construction method based on Graph RAG This method uses DFR original rule text as corpus basis, through large model driven information extraction and structured organization, constructs a digital flight rule knowledge graph with graph structure, and embeds it into a vector database. When the large model executes low-altitude flight rule judgment and path decision tasks, it can call the graph node content for "traceable" enhanced reasoning, outputting decision texts that are more compliant with flight regulations, airspace restrictions, and task adaptability. At the same time, flight recommendations and decisions can be traced back to the rule nodes and context semantics in the graph, avoiding "black box" output, and the graph content can be dynamically updated according to new regulations or temporary announcements, with good adaptability and expandability.

[0067] (2) Large model training method and decision text generation strategy based on cyclic calling This method proposes a large model training method and decision text generation strategy based on cyclic calling, which calls the large model three times. The first time is used for prompt optimization engineering to guide the model to understand the DFR rule corpus structure and task goal, generating high-quality prompt templates. The second time is used for knowledge graph creation, based on optimized prompts to drive the model to complete entity extraction, relationship construction, and graph initialization. The third time is used for intelligent decision generation, under the support of the knowledge graph, to complete low-altitude flight compliance analysis and route recommendation. Through the phased and multi-task linked large model calling method, the technical bottleneck of single-function model and information fragmentation between upstream and downstream in traditional large model application methods is broken through, and a new paradigm for flight rule intelligent decision is constructed.

[0068] Compared with the traditional low-altitude flight decision method relying on rule base, path planning algorithm or local intelligent model, the application has significant advantages in explainability, intelligence and dynamic adaptability.

[0069] (1) By introducing the Graph RAG architecture, the large language model has structured knowledge retrieval and reasoning ability on the basis of natural language understanding and generation ability, effectively making up for the explainability defects caused by the "black box" problem of large models.

[0070] (2) The introduction of DFR digital flight rule knowledge graph makes the model decision have clear legal basis and semantic support, can realize multi-step reasoning and chain judgment at the rule level, and improve the decision accuracy and compliance.

[0071] (3) The knowledge graph database has high maintainability, supports rapid updating of dynamic information such as regulation changes and temporary flight restricted areas, realizes adaptive adjustment of the model to the dynamic airspace environment, and improves the adaptability of the model to the high time-varying characteristics of the low-altitude environment. The visualization expression and retrieval updating mechanism of the graph realized by the graph database (such as Neo4j) not only helps the supervisors to audit the rules, but also facilitates the transparent presentation and responsibility tracing of the model output.

[0072] The embodiment of the application also provides a storage medium for storing a computer program, which is executed to perform at least the method described above.

[0073] The embodiment of the application also provides a control device, which includes a processor and a storage medium for storing a computer program; wherein the processor is used to execute the computer program to perform at least the method described above.

[0074] The embodiment of the application also provides a processor, which executes a computer program to perform at least the method described above.

[0075] The storage medium can be implemented by any type of nonvolatile storage device, or a combination thereof. The nonvolatile memory can be a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Programmable Read-Only Memory (EPROM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a Ferromagnetic Random Access Memory (FRAM), a Flash memory, a magnetic surface storage, an optical disc or a Compact Disc Read-Only Memory (CD-ROM). The magnetic surface storage can be a disk memory or a tape memory. The storage medium described in the embodiments of the present application is intended to include, but is not limited to, these and any other suitable type of memory.

[0076] In several embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented in other manners. The described device embodiments are merely schematic, and the division of the units is merely a logical function division. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0077] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place or distributed on a plurality of network units; some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0078] In addition, each functional unit in the embodiments of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in the form of hardware, or in the form of hardware plus software functional units.

[0079] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program performs the steps of the above-mentioned method embodiments when executed; and the foregoing storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disc or an optical disc and various storage medium capable of storing program codes.

[0080] Alternatively, the integrated unit of the present application can be stored in a computer readable storage medium if it is realized in the form of a software function module and sold or used as an independent product. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, includes several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes a mobile storage device, a ROM, a RAM, a magnetic disc or an optical disc and various storage medium capable of storing program codes.

[0081] The methods disclosed in the several method embodiments of the present application can be combined arbitrarily without conflict to obtain new method embodiments.

[0082] The features disclosed in the several product embodiments of the present application can be combined arbitrarily without conflict to obtain new product embodiments.

[0083] The features disclosed in the several method or device embodiments of the present application can be combined arbitrarily without conflict to obtain new method or device embodiments.

[0084] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For those skilled in the art, without departing from the concept of the present application, a number of equivalent substitutions or obvious modifications can be made, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present application.

Claims

1. A low-altitude flight intelligent decision-making method based on graph retrieval enhancement generation and digital flight rules, characterized by: The following steps are involved: S1. Constructing a Digital Flight Rules Knowledge Graph: Based on the Digital Flight Rules (DFR) text, extract entities and relationships between entities using a large language model, construct a structured knowledge graph, and store the knowledge graph in a vector database. S2: Call the large language model in a phased and cyclic manner for decision support: First call: Execute the prompt word optimization project and generate a prompt word template containing the task target description and output format example; Second call: Based on the optimized prompt word template, the large language model is driven to perform structural analysis on the DFR text to complete entity extraction, relationship recognition, and knowledge graph construction; The third call: Combined with the search enhancement capabilities of the knowledge graph, it drives the large language model to perform compliance verification and intelligent route decision-making for low-altitude flight missions; S3. Dynamic update: In response to airspace rule changes or temporary notices, real-time knowledge updates are achieved by inserting new nodes into the knowledge graph or modifying edge attributes.

2. The method according to claim 1, wherein Constructing the knowledge graph in step S1 specifically includes: Preprocess the digital flight rules (DFR) text to remove redundant information irrelevant to the flight rules and generate high-quality DFR corpus; Segmenting the DFR corpus into structured text units; Entity extraction is performed on each text unit through a large language model, and entities with similar semantics are aggregated; Identify inter-entity relationships based on the syntactic and semantic features of text units, and construct an initial graph structure with nodes representing entities and edges representing relationships. Nodes are clustered according to semantic similarity to form a topic hierarchy, and original text fragments are recorded to establish the traceability of graph elements.

3. The method according to claim 1, wherein Building a knowledge graph also includes relationship fusion processing: Fusing the identified inter-entity relationships, including merging duplicate relationship edges pointing to the same entity pair and standardizing the attribute definitions of edges; Establish interactive relationship links between entities in different topic clusters to support complex compliance reasoning and judgment involving multiple topic entities.

4. The method according to any one of claims 1 to 3, wherein In step S2, the first call specifically includes: Input the DFR corpus into the large language model to generate a summary description of the corpus topic and subsequent task objectives; Text segments are randomly extracted from the DFR corpus to drive the large language model to generate triple examples containing entities and relations, which serve as a format reference for subsequent entity relationship extraction.

5. The method according to any one of claims 1 to 3, characterized in that The prompt word optimization project specifically includes: Analyze the subject domain attributes of DFR corpus; Randomly sampled text snippets are used as example conversations and embedded into the prompt word structure to guide the large language model to understand the output format requirements of entity-relationship triples.

6. The method according to any one of claims 1 to 3, wherein: In step S2, the second call specifically includes: Inputting the optimized prompt word template generated by the first call and the complete pre-processed DFR text into the large language model; Drive the large language model to segment the text into structured units and identify and extract entities in each unit; Drive large language models to analyze relationships between entities, generate associated variables, and semantically bind text units to identified relationships; Based on the extracted entities and bound relationships, graph nodes and edges are created to form the initial knowledge graph.

7. The method according to any one of claims 1 to 3, characterized in that In step S2, the third call specifically includes: Input the current low-altitude flight mission information into the large language model; Combined with the knowledge graph for retrieval enhancement: Based on the semantic vectorization of task information, matching rule fragments, related entities and their compliance constraints are retrieved from the knowledge graph; The large language model is driven to make inference decisions based on the retrieval enhancement results. The inference decision-making process includes: a) Global summary: make a summary judgment on the overall compliance of the task; b) Local retrieval: Locate and apply rule fragments in the knowledge graph that are directly related to specific constraints; c) Comprehensive generation: Integrate global summary and local search results to generate the final flight compliance conclusion and route decision recommendations.

8. The method according to any one of claims 1 to 3, wherein: In step S3, the dynamic update specifically includes: When flight regulations change or temporary airspace notices appear, the knowledge graph content is dynamically adjusted by inserting nodes representing new rules or states into the knowledge graph, or modifying the attribute values ​​of edges between existing nodes to adapt to the latest airspace environment and rule requirements.

9. The method according to any one of claims 1 to 3, wherein: Also includes: Visual interaction: Import the constructed or updated knowledge graph into the graph database system for visual display, and support query and edit operations of nodes and relationships through the interface; The visual interaction specifically includes: Convert the knowledge graph data into a structured data format file; Importing the structured data format file into a graph database system; Display the knowledge graph through the visual interface provided by the graph database system; Query, add, delete, or modify nodes and relationships through the operation interface or database operation toolkit provided by the graph database system.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the low-altitude flight intelligent decision-making method based on graph retrieval enhancement generation and digital flight rules as described in any one of claims 1 to 9 is implemented.

Images