Civil aviation-oriented large model agent interaction system and method

Abstract

The invention belongs to the technical field of civil aviation information, and discloses a civil aviation-oriented large model agent interaction system and method. The method comprises the following steps: acquiring civil aviation data information by using a think tank module, establishing a private professional think tank of an airport, and carrying out query and data acquisition, logic judgment and information extraction in a mode of a built-in intelligent group or a user-defined intelligent group; by setting LLM large language model parameters and roles and utilizing a private professional think tank and an intelligent group, an airport application is constructed, and intelligent question answering and information retrieval are completed. Management and corresponding operation audit log viewing are carried out through system configuration; and capturing and understanding contexts and meanings in the text, recording and associating preorder dialogues, completing question-answer, realizing search of an original text source, and displaying the original text source. According to the civil aviation intelligent agent platform based on the large model, artificial intelligence and civil aviation services are integrated, and an airport is assisted to achieve digital and intelligent transformation.

Description

Technical Field

[0001] This invention belongs to the field of civil aviation information technology, and in particular relates to a large-scale intelligent agent interaction system and method for civil aviation. Background Technology

[0002] With the rapid development of the civil aviation industry and the deepening of digital transformation, airport operations management faces numerous challenges, including massive amounts of information, complex business scenarios, and diversified service demands. Traditional manual information processing methods are insufficient to meet the airport's needs for efficient and accurate information processing. Airports have accumulated a large amount of professional knowledge and business data, but lack effective integration and utilization mechanisms, making it impossible to respond quickly and accurately to various business issues.

[0003] Existing civil aviation systems cannot provide one-click search capabilities and lack sufficient semantic understanding, logical reasoning, and scenario adaptation abilities, making it difficult to meet the personalized needs of airport operations. Therefore, there is an urgent need for a civil aviation intelligent agent platform that can integrate civil aviation expertise, adapt to various business scenarios, and achieve intelligent question answering and efficient management. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this invention discloses a large-scale intelligent agent interaction system and method for civil aviation. The purpose of this invention is to improve the level of intelligence and service quality in civil aviation operations.

[0005] The technical solution is as follows: A large-scale intelligent agent interaction method for civil aviation, comprising the following steps: S1 utilizes the think tank module to collect civil aviation data and information, and establishes a private professional think tank for the airport, containing various types of information and data, which serves as the basic information source for answering user questions in intelligent question-and-answer dialogues; S2 allows airport users to query and retrieve data using the Smart Group module, either through built-in Smart Groups or custom Smart Groups, based on their own business needs. S3 utilizes the intelligent application module to construct actual airport business scenarios based on the LLM large language model. By setting LLM large language model parameters and roles, and using the established private professional think tank and intelligent groups, airport applications are built to complete intelligent question answering and information retrieval. S4, based on the constructed airport application, enables full-process control over the application users, core dependency models, and operational behaviors. It manages users, operation permissions, LLM large model, and views corresponding operation audit logs through the system configuration module. S5, based on management and viewing results, captures and understands the context and meaning in the text through the front-end question and answer module, records and associates previous dialogues, completes the question-and-answer process, realizes the original source search, and displays the results.

[0006] In step S1, within the airport's private professional think tank, information undergoes vectorization, segmentation, organization, and annotation to form a structured knowledge system; specifically including: For multi-source heterogeneous data in the civil aviation field, logical breakpoints in the data are automatically identified through semantic analysis, entity association detection, and data format verification technologies. A domain-enhanced vector generation model is introduced. In the organization and annotation stage, a two-layer structured annotation system is used. The bottom layer uses industry standard ontology to build basic tags. By parsing ontology concepts and matching data, verification tags are generated. The upper layer uses an intelligent tag recommendation engine to automatically generate civil aviation business scenario tags based on civil aviation data matching results, forming a dynamically adapted structured knowledge system, including knowledge structuring, domain adaptability, dynamic updates, semantic relevance, and machine reading.

[0007] In step S1, the airport's private professional think tank can be configured in the intelligent application module. By selecting the associated think tank in the settings, after configuration, the associated think tank and the intelligent application module can deeply collaborate and interact to achieve accurate association and dynamic retrieval of knowledge units. When the intelligent application module initiates a knowledge request, the think tank synchronously returns the matching knowledge unit index. Through semantic similarity algorithm matching, the associated knowledge is accurately located. Irrelevant knowledge is filtered by combining the user's previous dialogue context. And by continuously updating and expanding the think tank content, the platform can ensure that it can respond to the latest business needs and problems in a timely manner. Among them, the think tank content adopts a dynamic iteration mechanism, relies on the semantic understanding capabilities of the LLM large model, automatically identifies newly added civil aviation data, and supports manual uploading of professional materials. Through incremental training modules and integration with existing knowledge graphs, the content is continuously expanded. The dynamic iteration mechanism includes: automatic identification of new data, manual data supplementation, incremental fusion, and continuous content updates; The automatic identification of new data triggers an iterative process by monitoring changes in the data stream in real time, including: (1) Data drift detection; statistical distribution difference: KL divergence or JS divergence is used to measure the difference in distribution between the old and new data; ; when Model updates are triggered when a threshold is reached; for Divergence is used to identify abrupt changes in distribution; For new data in the distribution, For the old data of the distribution, For a new data set with distribution, A set of old, distributed data; (2) Time series outlier detection; The sliding window mean shift is expressed as: ; In the formula, This represents the mean shift value of the sliding window. This is the number of windows. The total sliding time, For a moment, for Data nodes at any given time; if Marked as an abnormal batch The historical standard deviation; The artificial data supplementation includes label confidence assessment and dynamic allocation of expert weights; The confidence assessment of annotations, based on consensus among multiple annotators, is expressed as follows: ; In the formula, This represents the task complexity coefficient. This is the confidence level value. To ensure consistency among multiple annotators, The total number of multiple annotators, For Reynolds operations, For complexity; The expert weights are dynamically allocated based on historical annotation accuracy, adjusting the weights to ensure that experts with high accuracy have a greater impact on the final annotations. The expression is as follows: ; In the formula, The weights adjusted for historical annotation accuracy for The weight of accuracy adjustment is constantly marked. No. Node at Accuracy of time, For the first Node at Accuracy of timing; When incrementally integrating new and old data, it is necessary to balance historical knowledge with new information, including: (1) Incremental update of model parameters, online gradient descent: ; In the formula, In order to be in Online gradient at time intervals, In order to be in Online gradient at time intervals, It is a linear gradient descent function. for Data nodes at any given time In order to be in to The increment of data node parameters at time step. For adaptive learning rate, online gradient Used for models to adapt to new data in real time; (2) Feature importance weighted fusion, dynamic feature weight adjustment: ; In the formula, for The dynamic feature weight adjustment value at time step. for The dynamic feature weight adjustment value at time step. For the first One characteristic, Forgetting factor, For dynamic feature weights; (3) Multimodal data alignment; cross-modal similarity calculation: ; In the formula, For embedding vectors, In order to be in The embedding vector of the modality; The content is continuously updated, which is a key computing strategy to maintain the timeliness of the system; Iteration termination condition determination, information gain saturation detection: ; In the formula, This is the information gain saturation detection value. For the old data One characteristic, For the first time for new data One feature; Real-time expansion of the knowledge graph and propagation of entity relationship confidence. Aggregating multi-source evidence based on path similarity: ; In the formula, Aggregating the confidence propagation path for entity relationships is used to ensure the accuracy of knowledge fusion. The radius of the propagation path, Input data for the first input. Input data for the second input; This represents the confidence propagation path.

[0008] In step S2, the built-in intelligent group includes database query, page search, and professional airport intelligent group, which includes vehicle dispatch, personnel scheduling, and video node acquisition. Customizable intelligent groups are used to independently configure startup parameters, input parameters, and component content according to specific business processes and requirements; Startup parameters are used to configure parameters and component types, input parameters are used to configure parameter names, data types and sources, and component content is written in Python code.

[0009] In step S2, corresponding data queries, logical judgments, and information extractions are performed according to the configuration content, specifically including: The automated processing flow is initiated, first retrieving relevant think tanks according to preset parameters and matching target data; The system uses the logical rules of the intelligent group to perform multi-condition comparison and judgment; LLM large model parsing capabilities are used to extract key information fields from structured data and form a standardized result set that can be directly called.

[0010] In step S3, the intelligent question answering and information retrieval process includes: Step a: The platform supports setting intelligent prompts. By manually customizing and adjusting the prompt content in the intelligent configuration, the chat direction of the LLM large language model can be guided. Variables are supported. The prompts set the core guidance logic, set variable values ​​and descriptions, and support format constraints and value range limits. Step b: The intelligent module realizes intelligent question answering and information retrieval functions. Based on the questions entered by users, combined with the professional knowledge in the think tank and the analysis results of the intelligent group, it generates accurate and detailed answers to meet the needs of various airport operation scenarios. In step c, the system uses an LLM large language model to parse semantics, identify key information related to civil aviation, filter redundant content, accurately extract core elements, and screen knowledge units with matching tags in the think tank. Based on high semantic similarity, it matches related professional knowledge units in the think tank, calls the think tank analysis component to perform logical deduction and data verification, and ensures the consistency between the association conclusion and the verified think tank knowledge units. Then, it integrates the results of both to generate multi-level answers, dynamically fills real-time data with prompt word variables, and ensures that the output is accurate and detailed.

[0011] In step S4, user management is used to manage, enable, and disable platform users; Operation permission management assigns different operation permissions to different users. Based on user identity information, permissions that match the user's responsibilities can be set on the user permission settings page, including management permissions and viewing permissions for think tanks and smart applications. Large model management is used to perform mainstream model management and parameter setting operations on large LLM language models; Mainstream model management is used for the access and version control of mainstream models, and the verification and activation of new models are realized through the model registry. The parameter settings provide options to control the randomness of the output and configure the maximum number of tokens. Operation audit logs are used to view all user operations; during security audits, operation records for specific time periods and user roles are filtered to verify compliance, trace issues, locate abnormal event time points, retrieve related operation sequences, and pinpoint the root cause of problems.

[0012] In step S5, the front-end question-answering module has powerful natural language processing capabilities, which can accurately capture and understand the context and meaning in the text. This includes: using a context-aware model to parse the text context; the context-aware model captures the semantic relationships between the text before and after; identifying user history interactions and question scenario information; parsing contextual logic; identifying referential relationships through inter-sentence association analysis; identifying the specific referents of pronouns; combining the word vector library in the civil aviation field with the knowledge graph; adapting to the scenario; achieving deep semantic understanding; and ensuring cross-turn semantic coherence by using dialogue history caching, context encoding, and intent tracking to achieve multi-turn dialogue context continuity. Record and link preceding dialogues to achieve the goal of answering questions directly and finding the original source, including: Based on the question-and-answer results, the platform enables the original source search function. When generating the question-and-answer results, it simultaneously records the mapping relationship between the answer content and the knowledge unit of the think tank. When the original source search function is triggered, it parses the core semantics and key entities of the current answer, locates the corresponding original knowledge base entry through association mapping, and retrieves the storage path of the entry to display the original source in the form of a reference.

[0013] Furthermore, by recording and linking previous dialogues, the system achieves a direct answer to the question and enables the retrieval of the original source. Further steps include: Custom tests are conducted. Based on the user's input question, matching and related segments are found, and segments that meet the ranking settings and similarity values ​​are displayed according to the settings. After the user inputs the question, the system first performs semantic parsing through the LLM large language model to extract core keywords and intents. Based on the keywords, the system searches the knowledge base, calls the vector similarity algorithm, compares the question vector with the segment vectors in the knowledge base, and calculates the similarity value. Results are filtered according to preset ranking rules, and segments that meet the threshold are retained; When finally displayed, the segmented content, similarity value, and source think tank are shown simultaneously.

[0014] Another objective of this invention is to provide a large-scale intelligent agent interaction system for civil aviation, which implements the aforementioned large-scale intelligent agent interaction method for civil aviation. The system includes: The think tank module is used for structured management of civil aviation data and information, and to establish a private professional think tank for airports, serving as the basic information source for answering user questions in intelligent question-and-answer dialogues; The Smart Group module is used by airport users to query and retrieve data, perform logical judgments, and extract information by using built-in Smart Groups or custom Smart Groups according to their own business needs. The intelligent application module is used to build real-world airport business scenarios based on the LLM large language model. By setting model parameters, roles, associated think tanks, and intelligent groups, it can build airport applications and realize intelligent question answering and information retrieval. The system configuration module is used for user management, operation permission management, large model management, and viewing of corresponding operation audit logs through system configuration; The front-end question-and-answer module is used to accurately capture and understand the context and meaning of text through dialogue, record and link previous dialogues, achieve the goal of answering the question directly, locate the original source, and display it. Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention integrates large-scale model technology with civil aviation business processes to create an AI intelligent agent platform for civil aviation airports. Based on large-scale model capabilities, it incorporates multiple intelligent modules tailored to airport business scenarios, aiming to provide AI out of the box and empower business operations. This further enhances the level of intelligence and service quality in civil aviation operations.

[0015] This invention enables centralized and structured management of various types of civil aviation information, forming a private professional think tank for airports. This provides a rich and accurate knowledge base for intelligent question answering, solving the problems of scattered and difficult-to-use information. Furthermore, users can quickly configure intelligent components according to actual business scenarios, meeting the diverse business needs of airports and improving the flexibility and relevance of business processing. Based on the LLM (Large Language Model), it achieves highly intelligent question answering and information retrieval functions, accurately understanding user questions and generating professional answers, reducing the information analysis difficulty for airport staff and improving management efficiency. This invention's large-model-based civil aviation intelligent agent platform integrates artificial intelligence with civil aviation business, helping airports achieve digital and intelligent transformation.

[0016] Secondly, for airports, intelligent question-and-answer services can replace manual processing of statistics, reducing labor costs; at the same time, the automated data processing of intelligent modules can reduce the time spent on business processes, improve operational efficiency, and indirectly reduce management costs.

[0017] A core challenge that the civil aviation industry has long sought to solve is how to enable intelligent systems to understand civil aviation expertise, adapt to dynamic business scenarios, and meet personalized user needs. Previous technical solutions were either limited to static knowledge base question-and-answer systems or relied on general large-scale models to generate answers, which were prone to generating false responses and unable to connect to airport-specific business data, thus failing to achieve the goals of accurate professional knowledge and flexible scenario adaptation. This invention precisely solves this problem through a collaborative design that combines a knowledge base module to accumulate private professional knowledge with an intelligent group module to adapt to dynamic business logic, achieving deep synergy between professional knowledge and intelligent interaction. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the large-scale intelligent agent interaction method for civil aviation provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the large-scale intelligent agent interaction method for civil aviation provided in the embodiments of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0020] The innovation of this invention lies in its collaborative design: a think tank module to accumulate private professional knowledge, and an intelligent group module to adapt to dynamic business logic, achieving deep synergy between professional knowledge and intelligent interaction. Intelligent question-and-answer services replace manual statistical processing, reducing labor costs and improving operational efficiency.

[0021] Example 1, as Figure 1 As shown, the large-scale intelligent agent interaction method for civil aviation provided in this embodiment of the invention includes: S1 utilizes the think tank module to collect civil aviation data and information, and establishes a private professional think tank for the airport, containing various types of information and data, which serves as the basic information source for answering user questions in intelligent question-and-answer dialogues; S2 allows airport users to query and retrieve data using the Smart Group module, either through built-in Smart Groups or custom Smart Groups, based on their own business needs. S3 utilizes the intelligent application module to construct actual airport business scenarios based on the LLM large language model. By setting LLM large language model parameters and roles, and using the established private professional think tank and intelligent groups, airport applications are built to complete intelligent question answering and information retrieval. S4, based on the constructed airport application, enables full-process control over the application users, core dependency models, and operational behaviors. It manages users, operation permissions, LLM large model, and views corresponding operation audit logs through the system configuration module. S5, based on management and viewing results, captures and understands the context and meaning in the text through the front-end question and answer module, records and associates previous dialogues, completes the question-and-answer process, realizes the original source search, and displays the results.

[0022] For example, in step S1, the think tank module is used to manage civil aviation data information in a structured manner, establishing a private professional think tank for the airport. This think tank contains various types of information data, such as policies, regulations, information, and systems, serving as the basic information source for answering user questions in the intelligent Q&A dialogue. The think tank module collects civil aviation data information through drag-and-drop or user uploads, covering multiple areas such as airport operations, flight management, and passenger services, thus constructing the airport's private professional think tank.

[0023] Information from the airport's private professional think tank, after being vectorized, segmented, organized, and labeled, forms a structured knowledge system, specifically including: For multi-source heterogeneous data in the civil aviation field, logical breakpoints in the data are automatically identified through technologies such as semantic analysis, entity association detection, and data format verification. A domain-enhanced vector generation model is introduced. During the data processing and annotation phase, a two-layer structured annotation system is used. The bottom layer uses industry-standard ontology to construct basic tags, generating validation tags by parsing ontology concepts and matching data. The upper layer uses an intelligent tag recommendation engine to automatically generate civil aviation business scenario tags based on civil aviation data matching results, forming a dynamically adaptable structured knowledge system. This includes knowledge structuring, domain adaptability, dynamic updability, semantic relevance, and machine readability. The domain-enhanced vector generation model is an artificial intelligence model based on a neural network structure. It integrates domain knowledge and vector generation technology, aiming to generate high-quality vector representations for specific domains to construct a dynamically adaptable structured knowledge system.

[0024] The airport's private professional think tank supports configuration within the intelligent application module. By selecting and associating think tanks in the settings, the associated think tanks and the intelligent application module can deeply collaborate and interact, enabling precise association and dynamic retrieval of knowledge units. When the intelligent application module initiates a knowledge request, the think tank synchronously returns a matching knowledge unit index. Through semantic similarity algorithms, it accurately locates related knowledge and filters out irrelevant knowledge by considering the user's prior dialogue context. Furthermore, by continuously updating and expanding the think tank content, the platform ensures timely responses to the latest business needs and issues.

[0025] The think tank content adopts a dynamic iteration mechanism (automatic identification of new data, manual data supplementation, incremental fusion, and continuous content updates). Relying on the semantic understanding capabilities of the LLM large model, it automatically identifies new data in the civil aviation field, while also supporting manual uploading of professional materials. Through incremental training modules, it integrates with existing knowledge graphs to achieve continuous content expansion.

[0026] The dynamic iteration mechanism includes: automatic identification of new data, manual data supplementation, incremental fusion, and continuous content updates; The automatic identification of new data triggers an iterative process by monitoring changes in the data stream in real time, including: (1) Data drift detection; statistical distribution difference: KL divergence or JS divergence is used to measure the difference in distribution between the old and new data; ; when Model updates are triggered when a threshold is reached; for Divergence is used to identify abrupt changes in distribution; For new data in the distribution, For the old data of the distribution, For a new data set with distribution, A set of old, distributed data; (2) Time series outlier detection; The sliding window mean shift is expressed as: ; In the formula, This represents the mean shift value of the sliding window. This is the number of windows. The total sliding time, For a moment, for Data nodes at any given time; if Marked as an abnormal batch The historical standard deviation; The artificial data supplementation includes label confidence assessment and dynamic allocation of expert weights; The confidence assessment of annotations, based on consensus among multiple annotators, is expressed as follows: ; In the formula, This represents the task complexity coefficient. This is the confidence level value. To ensure consistency among multiple annotators, The total number of multiple annotators, For Reynolds operations, For complexity; The expert weights are dynamically allocated based on historical annotation accuracy, adjusting the weights to ensure that experts with high accuracy have a greater impact on the final annotations. The expression is as follows: ; In the formula, The weights adjusted for historical annotation accuracy for The weight of accuracy adjustment is constantly marked. No. Node at Accuracy of time, For the first Node at Accuracy of timing; When incrementally integrating new and old data, it is necessary to balance historical knowledge with new information, including: (1) Incremental update of model parameters, online gradient descent: ; In the formula, In order to be in Online gradient at time intervals, In order to be in Online gradient at time intervals, It is a linear gradient descent function. for Data nodes at any given time In order to be in to The increment of data node parameters at time step. For adaptive learning rate, online gradient Used for models to adapt to new data in real time; (2) Feature importance weighted fusion, dynamic feature weight adjustment: ; In the formula, for The dynamic feature weight adjustment value at time step. for The dynamic feature weight adjustment value at time step. For the first One characteristic, Forgetting factor, For dynamic feature weights; (3) Multimodal data alignment; cross-modal similarity calculation: ; In the formula, For embedding vectors, In order to be in The embedding vector of the modality; The content is continuously updated, which is a key computing strategy to maintain the timeliness of the system; Iteration termination condition determination, information gain saturation detection: ; In the formula, This is the information gain saturation detection value. For the old data One characteristic, For the first time for new data One feature; Real-time expansion of the knowledge graph and propagation of entity relationship confidence. Aggregating multi-source evidence based on path similarity: ; In the formula, Aggregating the confidence propagation path for entity relationships is used to ensure the accuracy of knowledge fusion. The radius of the propagation path, Input data for the first input. Input data for the second input; This represents the confidence propagation path.

[0027] For example, in step S2, airport users can flexibly choose built-in smart groups or custom smart groups according to their own business needs.

[0028] The built-in intelligent components include three types: database query, page search, and professional airport intelligent components. Professional airport intelligent components are intelligent components pre-designed for common civil aviation business scenarios, such as vehicle dispatching, personnel scheduling, and video node acquisition. Customizable intelligent groups allow users to configure startup parameters, input parameters, and component content independently according to specific business processes and needs; The configuration uses a visual interface, and the startup parameters can be configured with parameters and component types. The input parameters can be configured with parameter names, data types, and sources. The component content can be written in Python code.

[0029] The platform performs corresponding data queries, logical judgments, and information extraction operations according to the configuration, specifically including: The automated processing flow is initiated by first searching related think tanks according to preset parameters (startup parameters and input parameters) and matching target data. Then, multi-condition comparison and judgment are performed through the logical rules of the intelligent group (built-in intelligent group or custom intelligent group); Finally, based on the LLM large model parsing capability, key information fields are extracted from the structured data (by parsing the structured data through the LLM large model, key civil aviation information fields such as flight number, take-off and landing time, boarding gate, and baggage data are identified and extracted), forming a standardized result set that can be directly called.

[0030] It is evident that the intelligent module can quickly and accurately acquire the required data and analyze and process it to support airport business decisions.

[0031] For example, in step S3, a real airport business scenario is constructed based on the LLM large language model in the intelligent application module. Users can build personalized airport applications by setting LLM large language model parameters (by selecting AI model settings for each intelligent group in the intelligent application module, including various mainstream large models in the industry), defining roles, and associating think tanks and intelligent groups. Multiple applications can be set up according to current needs, and different parameters, roles, and private professional think tanks and intelligent groups can be assigned to each application.

[0032] For example, building a personalized airport application includes: Step 1, set the parameters of the LLM large language model; Step 2, Define Roles: By assigning a specific role or identity to the LLM (Large Language Model), the output of the LLM is guided to better meet the needs of specific tasks or scenarios. For example, the role can be set as "Database Expert," requiring only that it generate SQL statements according to user requirements, without providing any other textual explanations.

[0033] Step 3: Set up a private professional think tank and a think tank as described in step S1.

[0034] For example, completing intelligent question answering and information retrieval includes: Step a: The platform supports setting intelligent prompts. Users can manually customize and adjust the prompt content in the intelligent configuration to guide the LLM large language model's chat direction. Variables are supported. Prompts can set core guidance logic, variable values ​​and descriptions (variable values ​​are dynamically replaceable key information placeholders), and format constraints and value range limitations are supported (specifying the output format and optional variable values).

[0035] Step b: The intelligent application module implements intelligent question answering and information retrieval functions. Based on the user's input question, combined with the professional knowledge in the think tank and the analysis results of the think tank, it generates accurate and detailed answers (matching knowledge unit indexes according to the question and using semantic similarity algorithms to accurately locate related knowledge), meeting the needs of various scenarios such as airport operations.

[0036] In step c, the system uses an LLM (Large Language Model) to parse semantics, identify key civil aviation-related information, filter redundant content, accurately extract core elements, and screen knowledge units with matching tags in the think tank. Based on high semantic similarity, it matches related professional knowledge units in the think tank and calls the think tank analysis component to perform logical deduction and data verification to ensure the consistency between the related conclusions and the verified think tank knowledge units. Then, it integrates the results of both to generate a multi-level answer, namely "core conclusion + supporting explanation + supplementary details". Real-time data is dynamically filled in through prompt word variables (after calling the data interface to obtain real-time information, the data is automatically matched and filled into the corresponding variables) to ensure that the output is accurate and complete in detail.

[0037] For example, in step S4, the system configuration module is mainly responsible for platform management, and realizes user management, operation permission management, LLM large language model management and viewing of corresponding operation audit logs through system configuration.

[0038] User management allows for the management of platform users, including enabling and disabling them; Operation permission management can assign different operation permissions to different users. Based on user identity information, permissions that match the user's responsibilities can be set on the user permission settings page, including management permissions and viewing permissions for think tanks and smart applications. The granularity of permissions can be refined to a single think tank or a single smart application, ensuring the security and privacy of platform data. Large model management allows for mainstream model management and parameter settings for large LLM language models. Among them, it supports the access and version control of mainstream models, and realizes the verification and activation of new models through the model registry; The parameter settings provide configuration options for core parameters such as output randomness and maximum token count; the operation audit log viewing function records all user operations, facilitating security auditing and problem tracing.

[0039] During security audits, users can customize and filter operation records for specific time periods and user roles to verify the compliance of operations. In the problem tracing process, the time nodes of abnormal events can be located based on the problem log records, and related operation sequences can be retrieved to accurately locate the root cause of the problem.

[0040] For example, in step S5, the front-end question-answering module (front-end display module) has powerful natural language processing capabilities, which can accurately capture and understand the context and meaning in the text, including: using a context-aware model to parse the text context, the context-aware model captures the semantic relationships between the text before and after, identifies information such as user history interactions and question scenarios, parses the context logic, identifies referential relationships through inter-sentence association analysis, identifies the specific referents of pronouns, combines the word vector library in the civil aviation field with the knowledge graph, adapts to the scenario, achieves deep semantic understanding, and achieves the continuation of multi-turn dialogue context through dialogue history caching, context encoding, and intent tracking to ensure cross-turn semantic coherence. In step S5, the preceding dialogue is recorded and linked to achieve the goal of matching the question with the answer and locating the original source, including: For example, the front-end question-and-answer module (front-end display module) records and associates previous dialogues to achieve context-aware intelligent question-and-answer. When a user raises a new question, the LLM large language model matches the context and combines the knowledge of the think tank to generate an associated answer, ensuring that the question is answered and is finally displayed. The platform can perform source retrieval based on question-and-answer results. When generating question-and-answer results, the platform simultaneously records the mapping relationship between the answer content and the knowledge unit of the think tank. When the source retrieval function is triggered, the core semantics and key entities of the current answer are parsed using LLM. After parsing, the vector library is associated, and the corresponding original knowledge base entry is located through vector matching and tag mapping. At the same time, the storage path of the entry is retrieved, and the source of the original text is displayed in the form of a reference. Users can view the knowledge source of the answer content and support one-click jump to view the full content online. This helps airport staff quickly understand the source of information, lowers the analysis threshold, and improves management efficiency.

[0041] For example, the platform can also perform custom tests, extracting core elements of the question based on user input, matching segment semantics and tag positioning, finding matching related segments, and displaying segments that meet ranking settings and similarity values ​​according to the settings, including: After the user inputs a question, the system first performs semantic parsing using the LLM large language model to identify key information, classify and determine the user's core needs, extract core keywords and intent, and search the knowledge base based on keywords. The question and knowledge base are segmented and converted into vectors. The vector similarity algorithm is called to compare the question vector with the segment vectors in the knowledge base, calculate the similarity value (similarity value less than 1), and rank them.

[0042] Results are filtered according to preset ranking rules (such as descending similarity), and segments that meet the threshold (configurable, such as ≥0.6) are retained; The final display will simultaneously show segmented content, similarity scores, and the source think tank. The platform supports third-party system embedding, allowing the code to be embedded into third-party systems for use.

[0043] For example, Figure 2 This is the principle of the large-scale intelligent agent interaction method for civil aviation provided in the embodiments of the present invention.

[0044] Example 2: The large-scale intelligent agent interaction system for civil aviation provided in this embodiment of the invention includes: The think tank module manages civil aviation data and information in a structured manner, establishing a private professional think tank for airports. It contains various types of information and data, such as policies, regulations, information, and systems, serving as the basic information source for answering user questions in intelligent question-and-answer dialogues.

[0045] The Smart Group module allows airport users to query and retrieve data, perform logical judgments, extract information, or perform other operations based on their own business needs through built-in Smart Groups (database query, page search, professional airport Smart Groups) or custom Smart Groups.

[0046] The intelligent application module is based on the actual airport business scenario built on the LLM large language model. By setting models, roles, associated think tanks, intelligent groups, etc., it builds airport applications to realize intelligent question answering and information retrieval, meeting the needs of various scenarios.

[0047] The system configuration module provides overall platform management, including user management, operation permission management, large model management, and viewing corresponding operation audit logs. Large model management supports various mainstream large models in the industry.

[0048] The front-end question-and-answer module accurately captures and understands the context and meaning in the text through dialogue, records and links previous dialogues, realizes the answer to the question, realizes the search for the original source, and displays it, helping airports lower the analysis threshold and improve management efficiency.

[0049] To further illustrate the effects of the embodiments of the present invention, the following experiments were conducted.

[0050] Create a think tank named "CHATBI Airport Edition" in the think tank module, and upload 20 documents from the Civil Aviation Administration and existing airport regulations and policies, and view the vectorized segmentation results.

[0051] In the Smart Application module, create a Smart Application named Airport BI Management, configure the AI ​​model (DeepSeek), system role (Civil Aviation Industry Data Analysis Expert), Smart Group (basic information, AI dialogue, SQL judgment, generate database SQL, judge, specify reply, SQL query, parse chart), airport database, generation rules (table description, field description, reference sample SQL, etc.), and associate it with the CHATBI Airport Edition think tank. Initiate a dialogue in the front-end display module, such as daily flight statistics, the top 5 airlines by number of flights, generate a pie chart, on-time rate calculation rules, view the original text, etc. After testing the knowledge base processing stage, the accuracy of vectorized segmentation was ≥95%, the response time for dynamic question and answer updates was ≤3 seconds, the accuracy of question and answer was ≥98%, and the accuracy of viewing the original text was also good.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A large-scale intelligent agent interaction method for civil aviation, characterized in that, The method includes the following steps: S1 utilizes the think tank module to collect civil aviation data and information, and establishes a private professional think tank for the airport, containing various types of information and data, which serves as the basic information source for answering user questions in intelligent question-and-answer dialogues; S2 allows airport users to query and retrieve data using the Smart Group module, either through built-in Smart Groups or custom Smart Groups, based on their own business needs. S3 utilizes the intelligent application module to build actual airport business scenarios based on the LLM large language model. By setting LLM large language model parameters and roles, and using the established private professional think tank and intelligent groups, airport applications are built to complete intelligent question answering and information retrieval. S4, based on the constructed airport application, enables full-process control over the application users, core dependency models, and operational behaviors. It manages users, operation permissions, LLM large model management, and views corresponding operation audit logs through the system configuration module. S5, based on management and viewing results, captures and understands the context and meaning in the text through the front-end question and answer module, records and associates previous dialogues, completes the question-and-answer process to find the original source of the text, and displays it.

2. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S1, within the airport's private professional think tank, information undergoes vectorization, segmentation, organization, and annotation to form a structured knowledge system; specifically including: For multi-source heterogeneous data in the civil aviation field, logical breakpoints in the data are automatically identified through semantic analysis, entity association detection, and data format verification technologies. A domain-enhanced vector generation model is introduced. In the organization and annotation stage, a two-layer structured annotation system is used. The bottom layer uses industry standard ontology to build basic tags. By parsing ontology concepts and matching data, verification tags are generated. The upper layer uses an intelligent tag recommendation engine to automatically generate civil aviation business scenario tags based on civil aviation data matching results, forming a dynamically adapted structured knowledge system, including knowledge structuring, domain adaptability, dynamic updates, semantic relevance, and machine reading.

3. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S1, the airport's private professional think tank can be configured in the intelligent application module. By selecting the associated think tank in the settings, after the configuration is completed, the associated think tank and the intelligent application module can conduct deep collaborative interaction to achieve accurate association and dynamic retrieval of knowledge units. When the intelligent application module initiates a knowledge request, the think tank synchronously returns the matching knowledge unit index. Through semantic similarity algorithm matching, the associated knowledge is accurately located. Combined with the user's previous dialogue context, irrelevant knowledge is filtered out. And by continuously updating and expanding the think tank content, the platform ensures that it can respond promptly to the latest business needs and issues; Among them, the think tank content adopts a dynamic iteration mechanism, relies on the semantic understanding capabilities of the LLM large model, automatically identifies newly added civil aviation data, and supports manual uploading of professional materials. Through incremental training modules and integration with existing knowledge graphs, the content is continuously expanded. The dynamic iteration mechanism includes: automatic identification of new data, manual data supplementation, incremental fusion, and continuous content updates; The automatic identification of new data triggers an iterative process by monitoring changes in the data stream in real time, including: (1) Data drift detection; statistical distribution difference: KL divergence or JS divergence is used to measure the difference in distribution between the old and new data; ； when Model updates are triggered when a threshold is reached; for Divergence is used to identify abrupt changes in distribution; For new data in the distribution, For the old data of the distribution, For a new set of distributed data, A set of old, distributed data; (2) Time series outlier detection; The sliding window mean shift is expressed as: ； In the formula, This represents the mean shift value of the sliding window. This is the number of windows. The total sliding time, For a moment, for Data nodes at any given time; if Marked as an abnormal batch The historical standard deviation; The artificial data supplementation includes label confidence assessment and dynamic allocation of expert weights; The confidence assessment of annotations, based on consensus among multiple annotators, is expressed as follows: ； In the formula, This represents the task complexity coefficient. This is the confidence level value. To ensure consistency among multiple annotators, The total number of multiple annotators. For Reynolds operations, For complexity; The expert weights are dynamically allocated based on historical annotation accuracy, adjusting the weights to ensure that experts with high accuracy have a greater impact on the final annotations. The expression is as follows: ； In the formula, The weights adjusted for historical annotation accuracy for The weight of accuracy adjustment is constantly marked. No. Node at Accuracy of time, For the first Node at Accuracy of timing; When incrementally integrating new and old data, it is necessary to balance historical knowledge with new information, including: (1) Incremental update of model parameters, online gradient descent: ； In the formula, In order to be in Online gradient at time intervals, In order to be in Online gradient at time intervals, It is a linear gradient descent function. for Data nodes at any given time In order to be in to The increment of data node parameters at time step. For adaptive learning rate, online gradient Used for models to adapt to new data in real time; (2) Feature importance weighted fusion, dynamic feature weight adjustment: ； In the formula, for The dynamic feature weight adjustment value at time step, for The dynamic feature weight adjustment value at time step, For the first One characteristic, Forgetting factor, For dynamic feature weights; (3) Multimodal data alignment; cross-modal similarity calculation: ； In the formula, For embedding vectors, In order to be in The embedding vector of the modality; The content is continuously updated, which is a key computing strategy to maintain the timeliness of the system; Iteration termination condition determination, information gain saturation detection: ； In the formula, This is the information gain saturation detection value. For the first time of the old data One characteristic, For the first time for new data One feature; Real-time expansion of the knowledge graph and propagation of entity relationship confidence. Aggregating multi-source evidence based on path similarity: ； In the formula, Aggregating the confidence propagation path for entity relationships is used to ensure the accuracy of knowledge fusion. The radius of the propagation path, Input data for the first input. Input data for the second input; This represents the confidence propagation path.

4. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S2, the built-in intelligent group includes database query, page search, and professional airport intelligent group, which includes vehicle dispatch, personnel scheduling, and video node acquisition. Customizable intelligent groups are used to independently configure startup parameters, input parameters, and component content according to specific business processes and requirements; Startup parameters are used to configure parameters and component types, input parameters are used to configure parameter names, data types and sources, and component content is written in Python code.

5. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S2, corresponding data queries, logical judgments, and information extractions are performed according to the configuration content, specifically including: The automated processing flow is initiated, first retrieving relevant think tanks according to preset parameters and matching target data; The system uses the logical rules of the intelligent group to perform multi-condition comparison and judgment; LLM large model parsing capabilities are used to extract key information fields from structured data and form a standardized result set that can be directly called.

6. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S3, the intelligent question answering and information retrieval process includes: Step a: The platform supports setting intelligent prompts. By manually customizing and adjusting the prompt content in the intelligent configuration, the chat direction of the LLM large language model can be guided. Variables are supported. The prompts set the core guidance logic, set variable values ​​and descriptions, and support format constraints and value range limits. Step b: The intelligent module realizes intelligent question answering and information retrieval functions. Based on the questions entered by users, combined with the professional knowledge in the think tank and the analysis results of the intelligent group, it generates accurate and detailed answers to meet the needs of various airport operation scenarios. In step c, the system uses an LLM large language model to parse semantics, identify key information related to civil aviation, filter redundant content, accurately extract core elements, and screen knowledge units with matching tags in the think tank. Based on high semantic similarity, it matches related professional knowledge units in the think tank, calls the think tank analysis component to perform logical deduction and data verification, and ensures the consistency between the association conclusion and the verified think tank knowledge units. Then, it integrates the results of both to generate multi-level answers, dynamically fills real-time data with prompt word variables, and ensures that the output is accurate and detailed.

7. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S4, user management is used to manage, enable, and disable platform users; Operation permission management assigns different operation permissions to different users. Based on user identity information, permissions that match the user's responsibilities can be set on the user permission settings page, including management permissions and viewing permissions for think tanks and smart applications. Large model management is used to perform mainstream model management and parameter setting operations on large LLM language models; Mainstream model management is used for the access and version control of mainstream models, and the verification and activation of new models are realized through the model registry. The parameter settings provide options to control the randomness of the output and configure the maximum number of tokens. The operation audit log is used to view all user operations recorded. During security audits, operation records for specific time periods and user roles are screened to verify compliance. In the problem tracing process, the time nodes of abnormal events are located, related operation sequences are retrieved, and the root cause of the problem is identified.

8. The large-scale intelligent agent interaction method for civil aviation according to claim 1, characterized in that, In step S5, the front-end question-answering module has powerful natural language processing capabilities, which can accurately capture and understand the context and meaning in the text. This includes: using a context-aware model to parse the text context; the context-aware model captures the semantic relationships between the text before and after; identifying user history interactions and question scenario information; parsing contextual logic; identifying referential relationships through inter-sentence association analysis; identifying the specific referents of pronouns; combining the word vector library in the civil aviation field with the knowledge graph; adapting to the scenario; achieving deep semantic understanding; and ensuring cross-turn semantic coherence by using dialogue history caching, context encoding, and intent tracking to achieve multi-turn dialogue context continuity. Record and link preceding dialogues to achieve the goal of answering questions directly and finding the original source, including: Based on the question-and-answer results, the platform enables the original source search function. When generating the question-and-answer results, it simultaneously records the mapping relationship between the answer content and the knowledge unit of the think tank. When the original source search function is triggered, it parses the core semantics and key entities of the current answer, locates the corresponding original knowledge base entry through association mapping, and retrieves the storage path of the entry to display the original source in the form of a reference.

9. The large-scale intelligent agent interaction method for civil aviation according to claim 8, characterized in that, Record and link previous dialogues to achieve the question-and-answer correspondence and enable the search for the original source. Further features include: Custom tests are conducted. Based on the user's input question, matching and related segments are found, and segments that meet the ranking settings and similarity values ​​are displayed according to the settings. After the user inputs the question, the system first performs semantic parsing through the LLM large language model to extract core keywords and intents. Based on the keywords, the system searches the knowledge base, calls the vector similarity algorithm, compares the question vector with the segment vectors in the knowledge base, and calculates the similarity value. Results are filtered according to preset ranking rules, and segments that meet the threshold are retained; When finally displayed, the segmented content, similarity value, and source think tank are shown simultaneously.

10. A large-scale intelligent agent interaction system for civil aviation, characterized in that, The system implements the large-scale intelligent agent interaction method for civil aviation as described in any one of claims 1-9, and the system includes: The think tank module is used for structured management of civil aviation data and information, and to establish a private professional think tank for airports, serving as the basic information source for answering user questions in intelligent question-and-answer dialogues; The Smart Group module is used by airport users to query and retrieve data, perform logical judgments, and extract information by using built-in or custom Smart Groups according to their own business needs. The intelligent application module is used to build real-world airport business scenarios based on the LLM large language model. By setting model parameters, roles, associated think tanks, and intelligent groups, it can build airport applications and realize intelligent question answering and information retrieval. The system configuration module is used for user management, operation permission management, large model management, and viewing of corresponding operation audit logs through system configuration; The front-end question-and-answer module is used to accurately capture and understand the context and meaning in the text through dialogue and question-and-answer, record and associate previous dialogues, realize the answer to the question, realize the source of the original text, and display it.

Images