A low-altitude flight intelligent decision method based on atlas retrieval enhancement generation and digital flight rules

Abstract

This invention discloses a method for intelligent decision-making in low-altitude flight based on graph retrieval-enhanced generation and digital flight rules (DFR). First, the DFR text is preprocessed and translated, and entities and relationships are extracted using a large language model to construct a structured knowledge graph, which is then stored in a vector database. Second, the large model is invoked in three cyclical calls: the first call optimizes the prompt word generation template; the second call parses the DFR text based on the template to complete the graph construction; and the third call combines graph retrieval capabilities to perform compliance verification and intelligent route decision-making for low-altitude flight missions. Furthermore, when airspace rules change or temporary notices are issued, knowledge is updated in real time by inserting new nodes into the knowledge graph or modifying edge attributes. This method combines graph retrieval-enhanced generation (GraphRAG) with DFR to improve the interpretability, accuracy, and dynamic adaptability of the large model's decisions, solving complex decision-making problems in low-altitude environments.

Description

Technical Field

[0001] This invention relates to artificial intelligence and intelligent decision-making systems, and in particular to an intelligent decision-making method for low-altitude flight based on Graph Retrieval-Augmented Generation (Graph RAG) and Digital Flight Rules (DFR). Background Technology

[0002] Currently, most UAV flight decisions are based on fixed rule bases, relying on manual path planning or deep reinforcement learning-based path planning techniques to plan and decide on flight missions. Traditional rule-based decision-making methods construct rule tables or logical rule trees based on manually set airspace regulations and flight restrictions, and then perform rule matching and flight plan verification through logical reasoning, decision trees, or expert systems. These systems were widely used in early UTM (Unmanned Traffic Management) and U-Space prototypes, but their scalability and adaptability are limited. Decision systems based on path planning algorithms are the classic path planning methods used by most low-altitude flight scheduling platforms. Examples include 3D path decision-making based on A*, Dijkstra's algorithm, and Rapidly Exploring Random Tree (PRT) algorithms, combined with geographic information data and airspace obstacle information for route generation. These methods have a clear structure and stable execution, but they are slow to respond to dynamic environments and struggle to handle semantic rule changes. Furthermore, path recommendation systems based on reinforcement learning (such as DQN and PPO algorithms) or graph neural networks (GNNs) that have emerged in recent years are beginning to be applied to intelligent flight path generation and obstacle avoidance decision-making. These systems can learn complex flight patterns to a certain extent, but they suffer from high training costs, poor interpretability, insufficient rule constraints, and weak scene transfer capabilities, and have not yet formed stable and controllable industry-level application solutions. Due to these limitations, a unified, efficient, and multi-source information fusion intelligent decision-making framework has not yet been established in the field of UAV flight mission decision-making.

[0003] With the rapid development of urban air traffic (UAM) and large-scale drone flights, low-altitude flight faces increasingly complex airspace structures, spatiotemporal conflicts, and regulatory adaptation issues. Traditional flight decision-making methods relying on manual rule design or local algorithm optimization are no longer sufficient to meet the real-world demands of large-scale collaboration, dynamic adjustment, and intelligent interpretation. Against this backdrop, large-model-driven intelligent decision-making technology has begun to be introduced into low-altitude intelligent management and flight decision-making systems. Currently, the application of large models in intelligent decision-making for low-altitude flight rules is still in its early stages and has not yet formed a complete systematic architecture. Improving the generalization ability of large models, resolving knowledge update mechanisms, and constructing unified semantic representation and rule expression methods are key to meeting the real-time requirements of complex low-altitude airspace.

[0004] Existing technologies are typically based on a fixed rule base, and their working principle is as follows: Figure 1 As shown.

[0005] The decision-making principles for low-altitude flight rules based on traditional methods are as follows: ① Rule-based decision-making methods rely on predefined airspace management rules and logical judgment mechanisms to achieve static verification of flight compliance and identification of path restrictions. ② Path planning algorithm-based decision-making methods generate the shortest or optimal flight path that meets the constraints by constructing an airspace spatial model and calling classic path search algorithms. ③ Reinforcement learning or graph neural network-based decision-making methods autonomously learn flight strategies or route scoring rules by training intelligent agent models or graph structure models to achieve intelligent path selection in dynamic environments.

[0006] The working principle of emerging large-scale model flight decision-making technology is as follows: Figure 2 As shown.

[0007] The decision-making process for low-altitude flight rules based on a large model is as follows: ① Define task requirements and construct prompt words; ② Complete the preprocessing of input information and supplement the context information; ③ Call the large language model to generate responses; ④ Perform structured parsing and output standardization; ⑤ Confirm collaborative output and adopt the decision.

[0008] Existing flight decision-making methods lack adaptability to dynamic environments and complex tasks, struggle to respond promptly to airspace changes, are inefficient, and pose potential flight safety hazards. Meanwhile, with the gradual opening and development of low-altitude airspace, UAVs and other low-altitude aircraft are exhibiting a diversified, multi-type, and high-density development trend, bringing complex issues of airspace coordination, route planning management, and rule adaptation. Flight decision-making based on traditional methods struggles to adjust in real-time according to airspace conditions, has poor adaptability to complex tasks, disconnects path planning from flight rules, and lacks integrated reasoning capabilities.

[0009] First, current low-altitude flight activities are characterized by diversity, a surge in numbers, and complex scenarios, posing unprecedented challenges to the refined management of airspace resources and intelligent decision-making regarding flight rules. In practical applications, UAVs often need to address complex factors such as dynamic airspace demarcation, temporary restricted flight areas, and high-density flight path conflicts, placing higher demands on the understanding of flight rules and compliance judgments. However, existing flight rule decision-making methods still primarily rely on manually set rule bases, classic path planning algorithms, or local intelligent models, which have several limitations.

[0010] Secondly, although large language models (LLMs) have demonstrated powerful capabilities in tasks such as natural language processing and knowledge reasoning, their current application in the field of intelligent decision-making for low-altitude flight rules still has many key shortcomings, making it difficult to meet the high standards required for actual aviation safety management. The "black box" characteristic that large models generally possess makes their decision-making process highly complex and invisible, and the model's reasoning chain difficult to trace. This low interpretability is tolerable in ordinary text generation tasks, but it is extremely fatal in low-altitude flight scenarios involving public safety and personal risks.

[0011] Secondly, the application of large-scale models in the low-altitude domain is mostly still at the stage of general pre-training on open-source corpora and aviation regulation texts, lacking a standardized flight rule system. This results in the generated content of the models often being superficial, lacking accuracy, compliance, and technical depth, making it difficult to handle professional rule judgment tasks. At the same time, the low-altitude flight scenario itself is highly dynamic, with frequent occurrences such as temporary no-fly notices, weather changes, and area adjustments. Existing large-scale models have not yet established an efficient and controllable knowledge update mechanism, making it difficult to integrate into the latest regulations or airspace status in a timely manner, resulting in poor performance in adapting to new scenarios and migrating tasks.

[0012] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0013] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a low-altitude flight intelligent decision-making method based on map retrieval enhancement generation and digital flight rules.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] A method for intelligent decision-making in low-altitude flight based on graph retrieval-enhanced generation and digital flight rules includes the following steps:

[0016] S1. Constructing a Digital Flight Rules Knowledge Graph: Based on the Digital Flight Rules (DFR) text, entities and relationships between entities are extracted using a large language model to construct a structured knowledge graph, which is then stored in a vector database.

[0017] S2. Staged, iterative invocation of the large language model for decision support:

[0018] First call: Execute the prompt word optimization project to generate a prompt word template that includes a description of the task objective and an example of the output format;

[0019] The second call: Based on the optimized prompt word template, the large language model is driven to perform structured parsing of the DFR text, completing entity extraction, relation recognition and knowledge graph construction;

[0020] The third invocation: Combining the enhanced retrieval 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;

[0021] S3. Dynamic Update: In response to changes in airspace rules or temporary announcements, knowledge is updated in real time by inserting new nodes or modifying edge attributes into the knowledge graph.

[0022] Furthermore, the construction of the knowledge graph in step S1 specifically includes:

[0023] The Digital Flight Rules (DFR) text is preprocessed to remove redundant information unrelated to flight rules and generate high-quality DFR corpus.

[0024] The DFR corpus is segmented into structured text units;

[0025] Entity extraction is performed on each text unit using a large language model, and entities with similar semantics are aggregated.

[0026] Relationships between entities are identified based on the syntactic and semantic features of text units, and an initial graph structure is constructed by representing entities with nodes and relationships with edges.

[0027] Clustering nodes based on semantic similarity forms a topic-level structure, and recording original text fragments establishes the traceability of graph elements.

[0028] Furthermore, the construction of knowledge graphs also includes relation fusion processing:

[0029] The identified relationships between entities are fused, including merging duplicate relationship edges pointing to the same entity pair and standardizing the attribute definitions of the edges.

[0030] Establish interaction relationship links between entities in different theme clusters to support complex compliance reasoning and judgment involving multiple theme entities.

[0031] Furthermore, in step S2, the first invocation specifically includes:

[0032] Input the DFR corpus into a large language model to generate a summary description of the corpus topics and subsequent task objectives;

[0033] Text fragments are randomly extracted from the DFR corpus, and the large language model is driven to generate triplet examples containing entities and relations, which serve as a format reference for subsequent entity and relation extraction.

[0034] Furthermore, the prompt word optimization project specifically includes:

[0035] Analyze the topic domain attributes of the DFR corpus;

[0036] Randomly selected text fragments are embedded as example dialogues into the prompt word structure to guide the large language model to understand the output format requirements of entity relation triples.

[0037] Furthermore, in step S2, the second invocation specifically includes:

[0038] The optimized prompt word template generated in the first call is input into the large language model along with the complete preprocessed DFR text.

[0039] The driving large language model segments text into structured units, identifies and extracts entities from each unit;

[0040] The large language model is driven to analyze the relationships between entities, generate associated variables, and semantically bind text units to the identified relationships.

[0041] Based on the extracted entities and their associated relationships, graph nodes and edges are created to form an initial knowledge graph.

[0042] Furthermore, in step S2, the third invocation specifically includes:

[0043] Input the current low-altitude flight mission information into the large language model;

[0044] Enhanced retrieval by combining the knowledge graph: Based on the semantic vectorization of task information, retrieve matching rule fragments, related entities and their compliance constraints from the knowledge graph;

[0045] The large language model is driven to make inference decisions based on retrieval enhancement results. The inference decision-making process includes:

[0046] a) Overall Summary: Make a preliminary assessment of the overall compliance of the task;

[0047] b) Local retrieval: Locate and apply rule fragments in the knowledge graph that are directly related to specific constraints;

[0048] c) Comprehensive Generation: Integrate global summaries with local search results to generate final flight compliance conclusions and route decision recommendations.

[0049] Furthermore, in step S3, the dynamic update specifically includes:

[0050] When flight regulations change or temporary airspace notices are issued, the knowledge graph content can be dynamically adjusted 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 ​​of edges between existing nodes.

[0051] Furthermore, the method also includes:

[0052] Visual interaction: Import the constructed or updated knowledge graph into the graph database system to achieve visual display, and support query and editing operations of nodes and relationships through the interface;

[0053] The visual interaction specifically includes:

[0054] Convert the knowledge graph data into a structured data format file;

[0055] Import the structured data format file into the graph database system;

[0056] The knowledge graph is displayed through a visual interface provided by the graph database system.

[0057] Through the operation interface or database operation toolkit provided by the graph database system, you can perform query, add, delete or modify operations on nodes and relationships.

[0058] A computer program product includes a computer program that, when executed by a processor, implements the aforementioned intelligent decision-making method for low-altitude flight based on graph retrieval enhancement generation and digital flight rules.

[0059] The present invention has the following beneficial effects:

[0060] This invention provides a method for intelligent decision-making in low-altitude flight based on graph retrieval-enhanced generation and digital flight rules (DFR). By deeply integrating the Graph RAG framework with DFR, a structured and traceable knowledge graph of low-altitude flight rules is constructed. A phased, iterative call strategy to a large language model significantly improves the interpretability, compliance, and dynamic adaptability of intelligent decision-making in low-altitude flight. Compared to traditional methods relying on static rule bases, path planning algorithms, or local intelligent models, this invention effectively overcomes the shortcomings of existing technologies, such as slow response to dynamic environments, fragmented rule reasoning, and the "black box" decision-making nature of large models. Specifically: The introduction of the Graph RAG architecture enables large models to retain the advantages of natural language processing while possessing structured retrieval and multi-step reasoning capabilities based on knowledge graphs. The decision-making process can be traced back to graph nodes, greatly enhancing transparency and credibility. The knowledge graph, constructed with 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 schemes comply with core constraints such as airspace speed, altitude, 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 static knowledge bases being unable to adapt to low-altitude, highly time-varying environments. Combined with a visual interaction mechanism, it forms a new paradigm of low-altitude intelligent decision-making that is interpretable, highly generalizable, and dynamically updated.

[0061] This invention utilizes the Graph RAG framework as a retrieval enhancement module for a large-scale model, and uses Knowledge Graph Retrieval Framework (DFR) as the judgment basis and standard for intelligent decision-making in low-altitude flight rules. It achieves decision generation based on knowledge graph retrieval enhancement, thereby improving the generalization ability of the large-scale model, solving the knowledge update mechanism, and constructing a unified semantic representation and rule expression method to meet the real-time requirements of complex low-altitude airspace. By applying a large language model to the field of low-altitude flight rule decision-making, leveraging its powerful general language understanding and generation capabilities, cross-scenario task transfer capabilities, and complex reasoning and logic capabilities, it achieves flight rule decision-making with strong generalization ability and real-time performance. The innovative contributions and key features of this invention are mainly reflected in:

[0062] (1) A method for constructing a knowledge graph of low-altitude digital flight rules based on Graph RAG

[0063] This method uses the original DFR rule text as its corpus and constructs a digital flight rule knowledge graph with a graph structure through large-scale model-driven information extraction and structured organization. This graph is then embedded into a vector database. This allows the large model to perform "traceable" enhanced reasoning by calling upon graph node content when performing low-altitude flight rule judgment and path decision tasks, outputting decision text that better conforms to flight compliance, airspace restrictions, and task adaptability. Simultaneously, both flight suggestions and decisions can be traced back to rule nodes and contextual semantics in the graph, avoiding "black box" output. The graph content can be dynamically updated according to new regulations or temporary notices, exhibiting good adaptability and scalability.

[0064] (2) Large model training method and decision text generation strategy based on loop calling

[0065] This method proposes a large-scale model training approach and decision text generation strategy based on cyclic calls, involving three calls to the large model. The first call optimizes the prompt word, guiding the model to understand the structure of the DFR rule corpus and the task objective, generating high-quality prompt word templates. The second call creates a knowledge graph, driving the model to complete entity extraction, relation construction, and graph initialization based on the optimized prompt words. The third call generates intelligent decisions, performing low-altitude flight compliance analysis and route recommendation with the support of the knowledge graph. This phased, multi-task-linked large-scale model call approach overcomes the technical bottlenecks of traditional large-scale model applications, such as single model functionality and fragmented upstream and downstream information, constructing a new paradigm for intelligent decision-making in flight rules.

[0066] Compared to traditional low-altitude flight decision-making methods that rely on rule bases, path planning algorithms, or local intelligent models, this invention has significant advantages in interpretability, intelligence, and dynamic adaptability:

[0067] (1) By introducing the Graph RAG architecture, large language models can acquire structured knowledge retrieval and reasoning capabilities on the basis of natural language understanding and generation capabilities, effectively making up for the interpretability defects caused by the "black box" problem of large models.

[0068] (2) The introduction of the DFR digital flight rules knowledge graph provides clear legal basis and semantic support for model decision-making, enabling multi-step reasoning and chain judgment at the rule level, thereby improving decision-making accuracy and compliance.

[0069] (3) Knowledge graph databases have high maintainability and support rapid updates to dynamic information such as changes in regulations and temporary restricted flight areas, enabling the model to adapt to dynamic airspace environments and improving its adaptability to the highly time-varying characteristics of low-altitude environments. The visualization and retrieval update mechanism of the graph database (such as Neo4j) not only helps regulators review rules, but also facilitates the transparent presentation of model outputs and the traceability of responsibility.

[0070] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the traditional flight decision-making method based on a fixed rule base.

[0072] Figure 2 This is a schematic diagram of the flight decision-making technology method based on a large model.

[0073] Figure 3 This is a flowchart of an embodiment of the present invention for constructing a DFR (Digital Flight Rules) knowledge graph based on Graph RAG.

[0074] Figure 4 This is a comparative analysis of the Graph DFR embodiment of the present invention with other large model question-answering examples.

[0075] Figure 5 This is a flowchart illustrating the overall process of the intelligent decision-making method for low-altitude flight based on graph retrieval enhancement generation and digital flight rules, as described in this invention. Detailed Implementation

[0076] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0077] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0078] In low-altitude environments, the lack of unified flight rules and effective intelligent decision-making for drones negatively impacts future large-scale, high-density low-altitude drone applications. To address this, this invention proposes a low-altitude flight intelligent decision-making method based on Knowledge Graph Retrieval Enhanced Generation (Graph RAG) and Digital Flight Rules (DFR), aiming to solve complex decision-making problems in dynamic environments within the low-altitude aviation field (especially for drones and urban air traffic). This invention utilizes the Graph RAG framework as a retrieval enhancement module for a large language model, and uses DFR as the judgment basis and standard for intelligent decision-making in low-altitude flight rules within the large language model. This achieves decision generation based on knowledge graph retrieval enhancement, leveraging the powerful general language understanding and generation capabilities, cross-scenario task transfer capabilities, and complex reasoning and logic capabilities of the large language model to achieve flight rule decision-making with strong generalization ability and real-time performance. Through this invention, the behavior of drones in low-altitude airspace can be standardized, including elements such as flight path, speed, altitude, priority, and spacing, providing clear constraints for flight path planning and conflict avoidance. Using knowledge graphs to structurally represent DFR rules and environmental situations enables large models to quickly and accurately retrieve relevant knowledge during decision-making, achieving low-altitude flight compliance verification and efficient intelligent decision-making.

[0079] See Figure 5 This invention provides a method for intelligent decision-making in low-altitude flight based on graph retrieval enhancement generation and digital flight rules, comprising the following steps:

[0080] Step S1: Construct a digital flight rule knowledge graph: Based on the digital flight rule (DFR) text, extract entities and relationships between entities through a large language model, construct a structured knowledge graph, and store the knowledge graph in a vector database.

[0081] In some embodiments, the construction of the knowledge graph in step S1 specifically includes: preprocessing the Digital Flight Rules (DFR) text, cleaning up redundant information unrelated to the flight rules, and generating a high-quality DFR corpus; segmenting the DFR corpus into structured text units; extracting entities from each text unit using a large language model, and aggregating semantically similar entities; identifying relationships between entities based on the syntactic and semantic features of the text units, and constructing an initial graph structure by representing entities with nodes and relationships with edges; clustering nodes according to semantic similarity to form a topic-level structure, and recording original text fragments to establish the traceability of graph elements.

[0082] In some embodiments, the construction of a knowledge graph also includes relation fusion processing: fusing the identified relationships between entities, including merging duplicate relation edges pointing to the same entity pair and standardizing the attribute definitions of the edges; and establishing interaction relationship links between entities in different topic clusters to support complex compliance reasoning and judgment involving multiple topic entities.

[0083] Step S2: Use the large language model in stages for decision support.

[0084] First call: Execute the prompt word optimization project to generate a prompt word template that includes a description of the task objective and an example of the output format;

[0085] The second call: Based on the optimized prompt word template, the large language model is driven to perform structured parsing of the DFR text, completing entity extraction, relation recognition and knowledge graph construction;

[0086] The third invocation: Combining the enhanced retrieval 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.

[0087] In some embodiments, step S2, the first call specifically includes: inputting the DFR corpus into a large language model to generate a summary description of the corpus topic and subsequent task objectives; randomly extracting text fragments from the DFR corpus to drive the large language model to generate triplet examples containing entities and relations, as a format reference for subsequent entity and relation extraction.

[0088] The prompt word optimization project specifically includes: analyzing the topic domain attributes of the DFR corpus; embedding randomly selected text fragments as example dialogues into the prompt word structure to guide the large language model to understand the output format requirements of entity relation triples.

[0089] In some embodiments, step S2 specifically includes: inputting the optimized prompt word template generated in 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, identifying and extracting entities in each unit; driving the large language model to analyze the relationships between entities, generating association variables and semantically binding the text units with the identified relationships; and creating graph nodes and edges based on the extracted entities and bound relationships to form an initial knowledge graph.

[0090] In some embodiments, step S2, the third invocation specifically includes: inputting the current low-altitude flight mission information into the large language model; performing retrieval enhancement by combining the knowledge graph: based on the semantic vectorization of the mission information, retrieving matching rule fragments, related entities, and their compliance constraints from the knowledge graph; driving the large language model to perform reasoning and decision-making based on the retrieval enhancement results, the reasoning and decision-making process including: a) global summary: making a summary judgment on the overall compliance of the mission; b) local retrieval: locating and applying rule fragments in the knowledge graph that are directly related to specific constraints; c) comprehensive generation: integrating the global summary and local retrieval results to generate the final flight compliance conclusion and route decision-making suggestions.

[0091] Step S3, Dynamic Update: In response to changes in airspace rules or temporary announcements, knowledge is updated in real time by inserting new nodes or modifying edge attributes into the knowledge graph.

[0092] In some embodiments, step S3 specifically includes: when changes in flight regulations occur or temporary airspace notices are issued, dynamically adjusting the content of the knowledge graph by inserting nodes representing new rules or states into the knowledge graph, or modifying the attribute values ​​(such as weights and relationship types) of existing edges between nodes, in order to adapt to the latest airspace environment and rule requirements.

[0093] In some embodiments, the low-altitude flight intelligent decision-making method based on graph retrieval enhancement generation and digital flight rules of the present invention further includes: visual interaction: importing the constructed or updated knowledge graph into a graph database system to achieve visual display, and supporting query and editing operations of nodes and relationships through an interface. The visual interaction may specifically 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 a visual interface provided by the graph database system; and performing query, addition, deletion, or modification operations on nodes and relationships through the operation interface or database operation toolkit (such as the py2neo library) provided by the graph database system.

[0094] The following describes specific embodiments of the present invention.

[0095] A method for intelligent decision-making in low-altitude flight based on graph retrieval-enhanced generation and digital flight rules is proposed. A large-scale model training method and decision text generation strategy based on iterative calls are designed, as shown in Table 1. Specifically, the steps and strategies of the large-scale model call are as follows: First, the first call is to perform a prompt word optimization project. The digital flight rules are directly input into the large-scale model as corpus. The large-scale model provides a theme summary by roughly reading the text content, and generates a general descriptive text that summarizes the possible roles and tasks to be achieved in subsequent tasks. It also randomly extracts text examples from the corpus and roughly generates triplet examples of entities and relationships between entities, which serve as the triplet extraction style for subsequent text. Second, the second call realizes knowledge graph creation. The previously constructed prompt word "Prompt" and the complete DFR are used as input. After the large-scale model loads the text, the original text is segmented into structured units, text entities are extracted and associated variables are generated, entities are aggregated to bind text units and relationships, and an initial graph structure is created based on entities and variables. Graph nodes and clustering relationships are generated, completing the creation of the knowledge graph. Finally, the third call is to use the large model as the reasoning engine for intelligent decision-making, and through the retrieval enhancement of the digital flight rule knowledge graph, to achieve the generation of decision text based on three categories: "global summary", "local retrieval" and "comprehensive generation".

[0096] Furthermore, to achieve a visual representation of low-altitude flight rules and improve the dynamic requirements of the atlas for the highly time-varying characteristics of low-altitude scenarios, the atlas is input into Neo4j based on the conversion between Parquet and CSV file formats, enabling a visual representation of low-altitude flight rules. Simultaneously, the py2neo library allows for querying nodes and relationships, as well as adding and deleting nodes as needed. This solution effectively improves the interpretability of intelligent decision-making in large-scale low-altitude models, generating decision judgments based on a standardized flight rule system using DFR (Depth-Flight Representation), and enhancing the time-varying update capability of large-scale models in the low-altitude domain.

[0097] Table 1. Large Model Training Methods and Decision Text Generation Strategies Based on Loop Calling

[0098] Large model call Purpose of call Calling steps First call Prompt word optimization project The digital flight rules are directly input into the large model as corpus. By roughly reading the text content, a theme summary is given, and a general descriptive text is generated, summarizing the possible role and tasks to be achieved in subsequent tasks. By randomly sampling invention examples from the text, examples of triples of entities and relationships between entities are roughly generated as patterns for subsequent text triple extraction. Second call Knowledge Graph Creation Using the previously constructed Prompt and the complete DFR as input, the text is loaded by calling a large model. The original text is then segmented into structured units, text entities are extracted, and associated variables are generated. Aggregation is performed between entities to bind text units to relationships. Based on the entities and variables, an initial graph structure is created, and graph nodes and clustering relationships are generated, thus creating the knowledge graph. Third call Graph-based intelligent decision-making As the reasoning engine for intelligent decision-making, the large model enhances decision text generation based on three categories: "global summarization", "local retrieval" and "comprehensive generation" through retrieval enhancement of the digital flight rule knowledge graph.

[0099] The process of constructing a DFR (Digital Flight Rules) knowledge graph based on Graph RAG is as follows: Figure 3 As shown. The main steps for extracting entities and relationships between entities from the text of the DFR (Digital Flight Rules) and constructing a knowledge graph vector database are as follows:

[0100] 1. DFR Digital Flight Rules Text Processing

[0101] Currently, widely used DFR documents are primarily compiled and published by NASA, written in English, and cover a range of flight rules and operational standards related to low-altitude flight, including route planning, airspace demarcation, flight restrictions, and weather responses. To adapt these DFR documents to intelligent decision-making within a Chinese context, high-quality professional translations were performed on the English documents. However, the original DFR documents contain a large amount of redundant information irrelevant to actual flight rules, such as author information, lists of participating organizations, historical version descriptions, and acknowledgments. If this unfiltered content is directly fed into large-scale model training or inference, it can easily lead to "illusion" problems. Therefore, filtering techniques were used to screen and clean the original text. The final result is a high-quality Chinese DFR flight rule corpus, providing a data foundation for subsequent knowledge graph construction, rule understanding in large-scale models, compliance judgment, and decision-making reasoning.

[0102] 2. Prompt word optimization project

[0103] The prompt words serve as the input controller for the large-scale model's interaction, and their design quality directly determines the accuracy, relevance, and stability of the model's output. In this project, the large-scale model not only extracts entities and relationships from structured and unstructured flight rule corpora to construct a knowledge graph but also serves as the core of intelligent decision-making and reasoning. Therefore, the prompt words must not only possess the accuracy of information extraction but also encompass the contextual constraints of decision-making and reasoning. In the prompt word optimization process, the system first analyzes the subject domain (such as "space science," "microbiology," and "hot news") and language attributes (such as "Chinese" and "English") of the training corpus. Then, it randomly extracts fragments from the corpus and embeds them as "example dialogues" into the prompt word structure to help the large-scale model understand the task objectives and output format. In the "example dialogue," the model roughly generates the entities and relationships between entities in the extracted fragments, forming a triplet output format, which serves as an output reference for subsequent graph creation.

[0104] 3. Construction of DFR Digital Flight Rules Knowledge Graph

[0105] Driven by a large language model, this approach guides the system to automatically extract entities, generate relationships, and organize them into a semantically structured knowledge graph from structured and unstructured rule texts using prompt words. The process begins with pre-processed and translated DFR (Digital Flight Rules) text as input, combined with task templates built during the prompt word optimization project, to guide the large model in semantic reading comprehension. In the text loading and unitization stage, the original rule document is segmented into logically defined structured text units. Subsequently, the model extracts entities from each text unit and aggregates and summarizes similar entities to improve the conciseness and consistency of information expression. Based on syntactic and semantic features, the model automatically identifies relationships between entities, semantically binding text units to form variable association structures. In the graph creation stage, an initial graph structure is built based on entities and their relationships, representing entities as nodes and rule relationships as edges. Simultaneously, nodes are clustered and categorized according to semantic similarity and task domain themes to form a higher-level topic hierarchy. Furthermore, by recording original text fragments, a source mapping is established for each graph element, achieving bidirectional traceability from the graph to the text. Building on this, a relationship fusion processing mechanism is further introduced to merge duplicate edges and standardize edge attributes, and to establish cross-topic entity interaction links to support complex reasoning and multi-step compliance judgments.

[0106] 4. Graph-based intelligent reasoning and decision making

[0107] A newly created DFR (Digital Flight Rules) knowledge graph is embedded as a vector database within a large model, serving as knowledge support for the model's retrieval and enhancement generation. In actual low-altitude flight decision-making tasks, the large model matches semantic vectorization with graph nodes in the knowledge graph to retrieve rule fragments, entity relationships, and compliance restrictions relevant to the current task. After reasoning enhancement, judgments and decisions are made. Compared to traditional decision-making methods based on rule base matching or path algorithms, this method not only retains the advantages of the large model in natural language understanding and generation but also significantly enhances its long-chain rule reasoning capabilities through the external structured knowledge graph. Furthermore, the graph database can dynamically insert new rule nodes or modify edge weights to achieve real-time updates to regulatory changes or airspace notices, overcoming the limitation of traditional models' static knowledge in adapting to changing scenarios.

[0108] Figure 4 This paper presents a comparative analysis of the Graph DFR embodiment of the present invention with other large model question-answering methods. Specific examples are as follows:

[0109] Question 1: What is DFR? (A vague question about the abbreviation)

[0110] 1. The answer from the large language model:

[0111] Design for Recycability

[0112] Digital Frequency Response;

[0113] Demand Forecasting and Replenishment;

[0114] Dynamic Fault Recovery;

[0115] Defensive Football Role.

[0116] 2. Graph RFR's answer:

[0117] DFR can have multiple meanings:

[0118] DFR (Digital Flight Review) is a comprehensive analytical report that integrates concepts and operational models in the aviation field.

[0119] DFR (Digital Flight Operations Regulation) refers to the digital flight operation mode in the aviation industry;

[0120] Digital Flight Rules (DFR) is a framework or system that uses digital rules and automated systems to manage flight intervals and other aspects during flight.

[0121] Question 2: What are the backup plans in case of DFR system failure? (Retrieval and reasoning question)

[0122] 1. The traditional RAG response:

[0123] Existing operating modes (such as VFR (Visual Flight Rules) and IFR (Instrument Flight Rules)) can serve as backup options. Mitigation measures include establishing an appropriate Design Assurance Level (DAL), system redundancy, and collision avoidance system action as a last resort (this system is functionally independent of the DFR and provides separate support).

[0124] 2. Graph RFR's answer:

[0125] In the event of a failure in the DFR (Digital Flight Operations) system, traditional flight separation and air traffic management methods are typically relied upon as backup solutions.

[0126] Manual visual separation control: When the DFR system is unable to provide automated separation recommendations, pilots or controllers will need to rely on personal experience and training to maintain safe flight separations.

[0127] VFR (Visual Flight Rules) and IFR (Instrument Flight Rules) Operation: In certain situations, if the DFR system is unavailable, the aircraft may need to switch to VFR or IFR mode for operation. VFR allows pilots to fly by visually observing their surroundings, while IFR relies on onboard navigation equipment and radio communications to ensure safety.

[0128] Using traditional radar for monitoring: Air traffic control (ATC) may continue to use radar systems to monitor aircraft positions and spacing in order to provide human intervention and guidance.

[0129] Backup communication links: In the event of a DFR system failure, backup voice or data communication links may be required for coordination and information sharing. This includes two-way voice communication, backup versions of automated interval systems, or other auxiliary technologies.

[0130] Figure 4 The graph RFR question answering demonstrated by the graph verifies the knowledge graph's ability to understand professional rules. Its retrieval results are directly transformed into semantic constraints for the decision engine, enabling large models to generate route plans that conform to DFR specifications (such as failover VFR mode) in low-altitude flight missions, thus achieving rule-driven intelligent scheduling.

[0131] In summary, this invention provides a low-altitude flight intelligent decision-making method based on graph retrieval-enhanced generation and digital flight rules. Compared with traditional technologies, the innovative contributions and key features of this invention are mainly reflected in the following aspects:

[0132] (1) A method for constructing a knowledge graph of low-altitude digital flight rules based on Graph RAG

[0133] This method uses the original DFR rule text as its corpus and constructs a digital flight rule knowledge graph with a graph structure through large-scale model-driven information extraction and structured organization. This graph is then embedded into a vector database. This allows the large model to perform "traceable" enhanced reasoning by calling upon graph node content when performing low-altitude flight rule judgment and path decision tasks, outputting decision text that better conforms to flight compliance, airspace restrictions, and task adaptability. Simultaneously, both flight suggestions and decisions can be traced back to rule nodes and contextual semantics in the graph, avoiding "black box" output. The graph content can be dynamically updated according to new regulations or temporary notices, exhibiting good adaptability and scalability.

[0134] (2) Large model training method and decision text generation strategy based on loop calling

[0135] This method proposes a large-scale model training approach and decision text generation strategy based on cyclic calls, involving three calls to the large model. The first call optimizes the prompt word, guiding the model to understand the structure of the DFR rule corpus and the task objective, generating high-quality prompt word templates. The second call creates a knowledge graph, driving the model to complete entity extraction, relation construction, and graph initialization based on the optimized prompt words. The third call generates intelligent decisions, performing low-altitude flight compliance analysis and route recommendation with the support of the knowledge graph. This phased, multi-task-linked large-scale model call approach overcomes the technical bottlenecks of traditional large-scale model applications, such as single model functionality and fragmented upstream and downstream information, constructing a new paradigm for intelligent decision-making in flight rules.

[0136] Compared to traditional low-altitude flight decision-making methods that rely on rule bases, path planning algorithms, or local intelligent models, this invention has significant advantages in terms of interpretability, intelligence, and dynamic adaptability.

[0137] (1) By introducing the Graph RAG architecture, large language models can acquire structured knowledge retrieval and reasoning capabilities on the basis of natural language understanding and generation capabilities, effectively making up for the interpretability defects caused by the "black box" problem of large models.

[0138] (2) The introduction of the DFR digital flight rules knowledge graph provides clear legal basis and semantic support for model decision-making, enabling multi-step reasoning and chain judgment at the rule level, thereby improving decision-making accuracy and compliance.

[0139] (3) Knowledge graph databases have high maintainability and support rapid updates to dynamic information such as changes in regulations and temporary restricted flight areas, enabling the model to adapt to dynamic airspace environments and improving its adaptability to the highly time-varying characteristics of low-altitude environments. The visualization and retrieval update mechanism of the graph database (such as Neo4j) not only helps regulators review rules, but also facilitates the transparent presentation of model outputs and the traceability of responsibility.

[0140] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0141] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0142] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0143] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0144] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0145] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0146] In addition, in the various embodiments of the present invention, each functional unit 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 hardware or in the form of hardware plus software functional units.

[0147] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0149] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0150] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0151] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0152] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent decision-making in low-altitude flight based on graph retrieval-enhanced generation and digital flight rules, characterized in that, Includes the following steps: S1. Constructing a Digital Flight Rules Knowledge Graph: Based on the Digital Flight Rules (DFR) text, entities and relationships between entities are extracted using a large language model to construct a structured knowledge graph, which is then stored in a vector database. S2. Staged, iterative invocation of the large language model for decision support: First call: Execute the prompt word optimization project to generate a prompt word template that includes a description of the task objective and an example of the output format; The second call: Based on the optimized prompt word template, the large language model is driven to perform structured parsing of the DFR text, completing entity extraction, relation recognition and knowledge graph construction; The second call specifically includes: inputting the optimized prompt word template generated in the first call and the complete preprocessed DFR text into the large language model; driving the large language model to identify and extract entities in each unit; driving the large language model to analyze the relationships between entities and semantically binding the text units with the identified relationships; and creating graph nodes and edges based on the extracted entities and the bound relationships to form an initial knowledge graph. The third invocation: Combining the retrieval enhancement capabilities of the knowledge graph, the large language model is driven to perform compliance verification and intelligent route decision-making for low-altitude flight missions; the third invocation specifically includes: inputting the current low-altitude flight mission information into the large language model; performing retrieval enhancement by combining the knowledge graph: retrieving matching rule fragments, related entities, and their compliance constraints from the knowledge graph; and driving the large language model to make inference decisions based on the retrieval enhancement results; S3. Dynamic Update: In response to changes in airspace rules or temporary announcements, knowledge is updated in real time by inserting new nodes or modifying edge attributes into the knowledge graph.

2. The method as described in claim 1, characterized in that, Step S1, which involves constructing the knowledge graph, specifically includes: The Digital Flight Rules (DFR) text is preprocessed to remove redundant information unrelated to flight rules and generate high-quality DFR corpus. The DFR corpus is segmented into structured text units; Entity extraction is performed on each text unit using a large language model, and entities with similar semantics are aggregated. Relationships between entities are identified based on the syntactic and semantic features of text units, and an initial graph structure is constructed by representing entities with nodes and relationships with edges. Clustering nodes based on semantic similarity forms a topic-level structure, and recording original text fragments establishes the traceability of graph elements.

3. The method as described in claim 1, characterized in that, Building a knowledge graph also includes relation fusion processing: The identified relationships between entities are fused, including merging duplicate relationship edges pointing to the same entity pair and standardizing the attribute definitions of the edges. Establish interaction relationship links between entities in different theme clusters to support complex compliance reasoning and judgment involving multiple theme entities.

4. The method according to any one of claims 1 to 3, characterized in that, In step S2, the first call specifically includes: Input the DFR corpus into a large language model to generate a summary description of the corpus topics and subsequent task objectives; Text fragments are randomly extracted from the DFR corpus, and the large language model is driven to generate triplet examples containing entities and relations, which serve as a format reference for subsequent entity and relation 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 topic domain attributes of the DFR corpus; Randomly selected text fragments are embedded as example dialogues into the prompt word structure to guide the large language model to understand the output format requirements of entity relation triples.

6. The method according to any one of claims 1 to 3, characterized in that, In step S2, the second call specifically includes: The large language model is driven to segment text into structured units in order to identify and extract entities in each unit; The large language model is driven to analyze the relationships between entities, generate associated variables, and semantically bind text units to the identified relationships.

7. The method according to any one of claims 1 to 3, characterized in that, In step S2, the third call specifically includes: Based on the semantic vectorization of task information, matching rule fragments, related entities and their compliance constraints are retrieved from the knowledge graph; The reasoning and decision-making process includes: a) Overall Summary: Make a preliminary assessment of 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 summaries with local search results to generate final flight compliance conclusions and route decision recommendations.

8. The method according to any one of claims 1 to 3, characterized in that, In step S3, the dynamic update specifically includes: When flight regulations change or temporary airspace notices are issued, the knowledge graph content can be dynamically adjusted 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 ​​of edges between existing nodes.

9. The method according to any one of claims 1 to 3, characterized in that, Also includes: Visual interaction: Import the constructed or updated knowledge graph into the graph database system to achieve visual display, and support query and editing operations of nodes and relationships through the interface; The visual interaction specifically includes: Convert the knowledge graph data into a structured data format file; Import the structured data format file into the graph database system; The knowledge graph is displayed through a visual interface provided by the graph database system. Through the operation interface or database operation toolkit provided by the graph database system, you can perform query, add, delete or modify operations on nodes and relationships.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements 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.

Images