An airport data intelligent processing method, device and electronic equipment

Abstract

The application provides an airport data intelligent processing method and device and electronic equipment. The method comprises the following steps: obtaining a question text for airport data; performing intention recognition on the question text by an intention recognition agent in a multi-agent collaborative processing system to obtain a target intention category; arranging a plurality of agents in the multi-agent collaborative processing system according to the target intention category by a general control planning agent in the multi-agent collaborative processing system to obtain the plurality of agents after arrangement; and processing the airport data according to the question text by the plurality of agents after arrangement to obtain a processing result of the airport data. The multi-agent collaborative processing system can dynamically assign the intention category of a complex question text to a plurality of agents capable of processing structured subtasks. The plurality of agents can realize accurate analysis of cross-domain data association through task cascading and distributed cooperation, thereby significantly improving the processing capability of complex composite intentions.

Description

Technical Field

[0001] This application relates to the technical fields of artificial intelligence, multi-agent collaboration, and airport data processing. Specifically, it relates to an intelligent airport data processing method, apparatus, and electronic device. Background Technology

[0002] In current airport operations, Large Language Models (LLMs) are widely used for customer service interaction, flight information queries, and emergency response due to their powerful natural language understanding and generation capabilities. However, the limitations of LLM training data make it difficult for the model to cover the diverse and complex scenarios in airport operations, resulting in insufficient generalization ability for issues such as fuzzy temporal representations and cross-domain correlations outside the training data. Therefore, solutions relying solely on LLMs are insufficient for handling complex and compounded intent text. Summary of the Invention

[0003] The purpose of this application is to provide an airport data intelligent processing method, apparatus, and electronic device to improve the problem of handling complex and compound intentions in problem text.

[0004] This application provides an intelligent processing method for airport data, comprising: acquiring question text for airport data; identifying the intent of the question text using an intent recognition agent in a multi-agent collaborative processing system to obtain a target intent category; orchestrating multiple agents in the multi-agent collaborative processing system according to the target intent category using a master control planning agent in the multi-agent collaborative processing system to obtain orchestrated agents; and processing the airport data using the orchestrated agents based on the question text to obtain the processing result of the airport data. In the implementation of the above scheme, the intent recognition agent in the multi-agent collaborative processing system identifies the intent category of the question text, and the master control planning agent orchestrates multiple agents in the multi-agent collaborative processing system using the target intent category. This allows the multi-agent collaborative processing system to dynamically assign the intent category of complex question text to multiple agents capable of processing structured subtasks. These orchestrated agents can achieve accurate parsing of cross-domain data associations through task cascading and distributed collaboration, thereby overcoming the shortcomings of a single model in simultaneously addressing intent recognition and dynamic prediction, ultimately significantly improving the processing capability of complex composite intents.

[0005] Optionally, in this embodiment of the application, after the intent recognition agent in the multi-agent collaborative processing system performs intent recognition on the question text, the method further includes: obtaining the confidence value of the target intent category; if the confidence value of the target intent category is less than a first confidence threshold, then calling a preset language model to perform intent recognition on the question text, obtaining a candidate intent category and the confidence value of the candidate intent category, wherein the model parameters of the preset language model are less than the model parameters of the post-trained language model used in the intent recognition agent; if the candidate intent category is different from the target intent category, and the confidence value of the candidate intent category and the confidence value of the target intent category meet preset error correction conditions, then the target intent category is corrected through the self-reflection mechanism of the intent recognition agent, wherein the preset error correction conditions include: both the confidence value of the candidate intent category and the confidence value of the target intent category are less than a second confidence threshold, or the difference between the confidence value of the candidate intent category and the confidence value of the target intent category is less than a third confidence threshold. In implementing the above scheme, the accuracy of intent recognition is significantly improved by introducing a dual-model collaboration mechanism of post-trained language model and pre-trained language model, along with a dynamic error correction strategy. When the confidence level of the initial intent recognition agent falls below a first threshold, the system calls the pre-trained language model with smaller parameters to generate candidate intent categories and triggers the error correction mechanism by comparing the confidence level difference between the candidate category and the target category. This design avoids misjudgments caused by fluctuations in the confidence level of a single model. Especially in complex or ambiguous problem scenarios (such as users asking "comparison of delay rates during the Spring Festival travel rush"), it can effectively correct misclassifications with low confidence levels, thereby improving the final accuracy of intent recognition.

[0006] Optionally, in this embodiment, the target intent category includes open-ended questions, and the multiple agents include open-ended question-answering agents. Before the orchestrated agents process the airport data based on the question text, the process further includes: calling the Large Language Model (LLM) in the open-ended question-answering agents through the overall control planning agent, so that the LLM infers the airport data based on the question text to obtain the answer text; and summarizing the answer text through the overall control planning agent to obtain the processing result of the airport data. In the implementation of the above scheme, the overall control planning agent coordinates the Large Language Model (LLM) in the open-ended question-answering agents to infer and generate answers for open-ended questions on the airport data, and summarizes the answer text to form the processing result, thereby significantly improving the processing efficiency and accuracy of open-ended questions in complex scenarios. Furthermore, the centralized scheduling mechanism of the overall control planning agent ensures the rationality of task allocation, the deep reasoning capability of the LLM can accurately parse the question intent and generate structured answers, and the answer summarization process further integrates multi-dimensional information, avoiding the limitations of a single agent, and ultimately achieving efficient and comprehensive analysis of airport data.

[0007] Optionally, in this embodiment, the multiple agents include: a spatiotemporal alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, and a trusted computing agent. The orchestrated agents process airport data based on the question text, including: mapping the relative time in the question text to standard date and time using the spatiotemporal alignment agent; mapping the relative spatial alignment in the question text to standard regional locations using the spatiotemporal alignment agent; retrieving domain knowledge from the vertical domain knowledge base corresponding to the airport data using the domain knowledge agent; determining the table nouns and field list in the target data warehouse using the data selection agent based on the question text; generating an SQL query statement using the code generation agent based on the standard date and time, standard regional locations, domain knowledge, and / or the table nouns and field list in the target data warehouse; and executing the SQL query statement in the target data warehouse using the trusted computing agent to obtain the query results corresponding to the question text. In the implementation of the above scheme, the intelligent processing of airport data is achieved through a multi-agent collaborative mechanism, significantly improving processing efficiency and result reliability in complex query scenarios. Among them, the spatiotemporal alignment agent accurately parses and standardizes the spatiotemporal information in the problem, solving the challenge of aligning multi-source data; the domain knowledge agent enhances semantic understanding capabilities by combining vertical domain expertise; the data selection agent and code generation agent reduce the risk of human intervention through automated table structure parsing and SQL generation; and the trusted computing agent ensures the verifiability and security of the query process. This end-to-end process, through modular division of labor and intelligent collaboration, not only reduces the complexity of cross-modal data processing but also achieves high generalization through standardized interfaces and domain adaptation, ultimately significantly shortening the response cycle of complex queries while ensuring data accuracy.

[0008] Optionally, in this embodiment, the target intent category includes: structured data query, and the multiple agents also include: visualization agent; after obtaining the query results corresponding to the question text, the method further includes: generating a visualization script by the visualization agent based on the visualization description in the question text and the query results, the visualization script being used to visualize the query results during execution; performing static security review, feasibility assessment review, and / or logical verification review on the visualization script by the trusted computing agent to obtain the visualization review result; if the visualization review result passes, the visualization script is sent or executed. In the implementation of the above scheme, by introducing a collaborative mechanism between the visualization agent and the trusted computing agent, the efficiency and security of visualization processing of structured data query results are significantly improved. The visualization agent automatically generates a visualization script based on the description in the question text and the query results, realizing automated mapping from data to graphical display and reducing the complexity of manually writing scripts; while the trusted computing agent ensures the reliability of the generated script at the syntax, permission, and logic levels through static security review, feasibility assessment, and logical verification, effectively avoiding potential security risks and execution errors. This layered collaboration model not only ensures the accuracy and compliance of data visualization, but also shortens the time required for traditional manual verification through automated review. Ultimately, while ensuring system security, it achieves fast and intelligent visualization of query results.

[0009] Optionally, in this embodiment, the target intent category includes: data prediction; the multiple agents further include: a data cleaning agent and a feature engineering agent; the process of processing airport data based on the question text by the orchestrated multiple agents further includes: cleaning the airport data using the data cleaning agent to obtain cleaned airport data; generating a feature processing script using the feature engineering agent based on the feature descriptions in the question text and the cleaned airport data, the feature processing script being used to perform feature derivation, feature filtering, and / or feature encoding operations on the airport data; performing static security review, feasibility assessment review, and / or logical verification review on the feature processing script using a trusted computing agent to obtain the feature review result; if the feature review result passes, executing the feature processing script to obtain the data features of the airport data, and predicting the data features of the airport data using a machine learning model determined based on the question text to obtain the processing result of the airport data, the machine learning model being trained using historical airport data. In the implementation of the above scheme, the intelligent prediction processing of airport data is achieved through a multi-agent collaborative mechanism, significantly improving the data quality and model reliability of the prediction task. The data cleaning agent automatically identifies and corrects missing values, outliers, and duplicates in the original data, effectively eliminating noise interference. The feature engineering agent dynamically generates feature processing scripts based on the problem text, accurately extracting high-value information through feature derivation, filtering, and encoding, enhancing the expressive power of the model input. The trusted computing agent ensures the script's compliance in terms of syntax, permissions, and business logic through static security reviews, feasibility assessments, and logical verification, mitigating potential execution risks. Finally, the high-quality data, after cleaning and feature optimization, is input into a machine learning model trained on historical data, achieving accurate predictions of airport data. This process, through modular division of labor and intelligent collaboration, not only ensures the standardization and security of data processing but also significantly improves the accuracy and generalization ability of the prediction model through deep optimization of feature engineering.

[0010] Optionally, in this embodiment, the target intent category includes: domain knowledge question answering; the multiple agents include: a spatiotemporal alignment agent and a domain knowledge agent; the multiple agents, after being orchestrated, process the airport data according to the question text, including: mapping the relative time in the question text to standard date and time using the spatiotemporal alignment agent; mapping the relative spatial alignment in the question text to standard regional locations using the spatiotemporal alignment agent; retrieving multiple knowledge fragments of the airport data from a preset domain knowledge base using the domain knowledge agent, the preset domain knowledge base being constructed by converting preset unstructured data into content vectors; and aggregating the multiple knowledge fragments of the airport data using the overall control planning agent to obtain the processing result of the airport data.

[0011] In implementing the above solution, a multi-agent collaborative mechanism was used to achieve intelligent processing of airport domain knowledge-based question answering, significantly improving the accuracy and response efficiency of the question answering system in complex scenarios. The spatiotemporal alignment agent precisely maps the ambiguous time and relative spatial descriptions in the question text into standardized time and geographic coordinates, resolving the ambiguity of spatiotemporal information in natural language. The domain knowledge agent, based on a vectorized pre-defined knowledge base, can quickly retrieve structured knowledge fragments related to the question, achieving efficient conversion of unstructured text into professional information. The overall control and planning agent aggregates multi-source knowledge fragments and combines them with contextual logic to generate a final answer that conforms to business rules. The collaborative work of these three agents not only supports accurate positioning and correlation analysis of dynamic airport information (such as real-time flight status) and static information (such as facility locations), but also reduces the need for manual intervention through automated processing, ultimately achieving efficient intelligent responses under complex spatiotemporal semantics while ensuring the authority of the answer.

[0012] This application also provides an intelligent airport data processing device, comprising: a question text acquisition module for acquiring question text related to airport data; an intent category recognition module for performing intent recognition on the question text using an intent recognition agent in a multi-agent collaborative processing system to obtain a target intent category; an intelligent collaborative orchestration module for orchestrating multiple agents in the multi-agent collaborative processing system according to the target intent category using a master control planning agent in the multi-agent collaborative processing system to obtain orchestrated multiple agents; and an airport data processing module for processing airport data based on the question text using the orchestrated multiple agents to obtain the processing result of the airport data.

[0013] Optionally, in this embodiment of the application, the airport data intelligent processing device further includes: a target confidence value acquisition module, used to acquire the confidence value of the target intent category; a candidate confidence value acquisition module, used to, if the confidence value of the target intent category is less than a first confidence threshold, call a preset language model to perform intent recognition on the question text, obtain a candidate intent category and the confidence value of the candidate intent category, wherein the model parameters of the preset language model are less than the model parameters of the post-trained language model used in the intent recognition agent; and a target intent category error correction module, used to, if the candidate intent category is different from the target intent category, and the confidence value of the candidate intent category and the confidence value of the target intent category meet preset error correction conditions, then correct the target intent category through the self-reflection mechanism of the intent recognition agent, wherein the preset error correction conditions include: both the confidence value of the candidate intent category and the confidence value of the target intent category are less than a second confidence threshold, or the difference between the confidence value of the candidate intent category and the confidence value of the target intent category is less than a third confidence threshold.

[0014] Optionally, in this embodiment, the target intent category includes: open-ended questions; multiple agents include: open-ended question-answering agents; the airport data intelligent processing device further includes: an LLM model invocation module, used to invoke the large language model LLM in the open-ended question-answering agent through the overall control planning agent, so that the large language model LLM can infer the airport data based on the question text to obtain the answer text; and an answer text aggregation module, used to aggregate the answer text through the overall control planning agent to obtain the processing result of the airport data.

[0015] Optionally, in this embodiment, the multiple agents include: a spatiotemporal alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, and a trusted computing agent; the airport data processing module includes: a time alignment mapping submodule, used to map the relative time alignment in the question text to standard date and time using the spatiotemporal alignment agent; a spatial alignment mapping submodule, used to map the relative spatial alignment in the question text to standard regional locations using the spatiotemporal alignment agent; a domain knowledge retrieval submodule, used to retrieve domain knowledge from the vertical domain knowledge base corresponding to the airport data using the domain knowledge agent based on the question text; a database table field determination submodule, used to determine the table nouns and field list in the target data warehouse using the data selection agent based on the question text; a query statement generation submodule, used to generate SQL query statements using the code generation agent based on standard date and time, standard regional locations, domain knowledge, and / or the table nouns and field list in the target data warehouse; and a query statement execution submodule, used to execute the SQL query statements in the target data warehouse using the trusted computing agent to obtain the query results corresponding to the question text.

[0016] Optionally, in this embodiment, the target intent category includes: structured data query; the multiple agents also include: visualization agent; the airport data processing module further includes: a visual script generation submodule, used to generate a visual script based on the visual description in the question text and the query results through the visualization agent, the visual script being used to visualize the query results during execution; a visual script review submodule, used to perform static security review, feasibility assessment review and / or logical verification review on the visual script through a trusted computing agent to obtain a visual review result; and a script sending and execution submodule, used to send or execute the visual script if the visual review result passes.

[0017] Optionally, in this embodiment, the target intent category includes: data prediction; the multiple agents also include: a data cleaning agent and a feature engineering agent; the airport data processing module further includes: an airport data cleaning submodule, used to clean airport data using the data cleaning agent to obtain cleaned airport data; a feature script generation submodule, used to generate a feature processing script using the feature engineering agent based on the feature description in the question text and the cleaned airport data, the feature processing script being used to perform feature derivation, feature filtering, and / or feature encoding processing operations on the airport data; a feature script review submodule, used to perform static security review, feasibility assessment review, and / or logical verification review on the feature processing script using a trusted computing agent to obtain feature review results; and a feature script execution submodule, used to execute the feature processing script if the feature review results pass, to obtain the data features of the airport data, and to predict the data features of the airport data using a machine learning model determined based on the question text to obtain the processing results of the airport data, the machine learning model being trained using historical airport data.

[0018] Optionally, in this embodiment, the target intent category includes: domain knowledge question answering; multiple agents include: spatiotemporal alignment agent and domain knowledge agent; the airport data processing module includes: a time alignment mapping submodule, used to map the relative time alignment in the question text to standard date and time through the spatiotemporal alignment agent; a spatial alignment mapping submodule, used to map the relative spatial alignment in the question text to standard area location through the spatiotemporal alignment agent; a knowledge fragment retrieval submodule, used to retrieve multiple knowledge fragments of airport data from a preset domain knowledge base through the domain knowledge agent, the preset domain knowledge base being constructed by converting preset unstructured data into content vectors; and a knowledge fragment aggregation submodule, used to aggregate multiple knowledge fragments of airport data through the overall control planning agent to obtain the processing result of airport data.

[0019] This application also provides an electronic device, including a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform the methods described above.

[0020] This application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to perform the methods described above.

[0021] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The illustration shows a flowchart of the intelligent airport data processing method provided in an embodiment of this application;

[0024] Figure 2 The diagram shown is a schematic representation of the core functions of the multi-agent cooperative processing system provided in the embodiments of this application.

[0025] Figure 3 The diagram shown is a flowchart illustrating the overall control planning agent orchestration and invocation of multiple agents provided in an embodiment of this application.

[0026] Figure 4 The diagram shown is a structural schematic of the airport data intelligent processing device provided in an embodiment of this application;

[0027] Figure 5 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the embodiments of this application are for illustrative and descriptive purposes only and are not intended to limit the protection scope of the embodiments of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the embodiments of this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of the embodiments of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0029] Furthermore, the described embodiments are merely a part of the embodiments of this application, and not all of them. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of this application, but merely to illustrate selected embodiments of this application.

[0030] It is understood that the terms "first" and "second" in the embodiments of this application are used to distinguish similar objects. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different. In the description of the embodiments of this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship. The term "multiple" refers to two or more (including two), and similarly, "multiple groups" refers to two or more groups (including two groups).

[0031] It should be noted that the airport data intelligent processing method provided in this application embodiment can be executed by an electronic device. Here, electronic device refers to a device terminal or server with the function of executing computer programs. Device terminals include, for example, smartphones, personal computers, tablets, personal digital assistants, or mobile internet devices. Servers refer to devices that provide computing services through a network. Servers include, for example, x86 servers and non-x86 servers. Non-x86 servers include, for example, mainframes, minicomputers, and UNIX servers.

[0032] In current airport operations, data processing largely relies on Large Language Models (LLMs). While LLMs possess powerful natural language understanding and generation capabilities and are widely used in customer service interactions, flight information queries, and emergency response, practical applications have revealed that purely LLM-based solutions are prone to overfitting or undergeneralization when handling complex, multi-layered intents. Long-term research has shown that the limitations of LLM training data make it difficult for the model to cover the diverse and complex scenarios in airport operations (such as comparing flight delay rates and resource scheduling recommendations during the Spring Festival travel rush with 2019). This results in insufficient generalization ability for fuzzy time representations and cross-domain correlations outside the training data. Therefore, solutions solely relying on LLMs are ill-suited for handling complex, multi-layered intent text.

[0033] For the above issues, please refer to Figure 1The illustrated flowchart illustrates the intelligent airport data processing method provided in this application embodiment. The main idea of ​​this method is to identify the intent category of the question text using an intent recognition agent in a multi-agent collaborative processing system, and then orchestrate multiple agents in the system using a central control planning agent targeting the intent category. This allows the multi-agent collaborative processing system to dynamically assign the intent category of complex question texts to multiple agents capable of handling structured subtasks, overcoming the limitations of a single model in simultaneously addressing intent recognition and dynamic prediction, ultimately significantly improving the processing capability for complex composite intents. Implementation methods of the aforementioned intelligent airport data processing method may include:

[0034] Step S110: Obtain the question text for airport data.

[0035] Query text refers to natural language requests or queries made by users regarding airport data. It typically contains vague expressions, complex intents, or specific business requirements, and needs to be parsed by the system and transformed into an executable task. For example, the query text mentioned above could be a structured data query, such as "What flights are there from Chengdu to Guangzhou today?", or a complex intent like "Compare flight delay rates and resource scheduling suggestions during the Spring Festival travel rush compared to 2019."

[0036] Step S120: The intent recognition agent in the multi-agent collaborative processing system is used to perform intent recognition on the question text to obtain the target intent category.

[0037] Please see Figure 2 The diagram illustrates the core functions of the multi-agent collaborative processing system provided in this application embodiment. The multi-agent collaborative processing system refers to a software system coordinated by multiple agent intelligent agents. This system may include: a general control planning agent, an intent recognition agent, a spatiotemporal alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, a trusted computing agent, a visualization agent, a data cleaning agent, a feature engineering agent, and a modeling agent, etc. These multiple agents can call each other through predefined prompts and interfaces. All requests and responses between these agents follow the JSON pattern, specifically using a Model Context Protocol (MCP), etc.

[0038] The intent recognition agent is a core component of a multi-agent system. It is responsible for extracting user intent from the question text using natural language processing techniques (such as rule matching, machine learning, or deep learning) and classifying it into predefined intent categories. Essentially, the intent recognition agent transforms the user's ambiguous natural language requests into structured intents that the system can understand, thereby driving subsequent task orchestration. This intent recognition agent can perform the recognition when invoked by the overall control and planning agent; for example, the overall control and planning agent might call upon the intent recognition agent to perform intent recognition on the question text and obtain the target intent category.

[0039] The target intent category is the intent category obtained by the intent recognition agent through intent recognition of the question text. Intent categories can include: open-ended questions, structured data queries, data prediction, and domain knowledge question answering, etc.

[0040] Step S130: The overall control planning agent in the multi-agent collaborative processing system orchestrates multiple agents in the multi-agent collaborative processing system according to the target intent category, thereby obtaining multiple orchestrated agents.

[0041] The central control and planning agent acts as the coordination hub of a multi-agent system. It is responsible for dynamically orchestrating the collaborative processes of multiple sub-agents based on the output of the intent recognition agent, ensuring efficient task execution. The central control and planning agent transforms abstract intents into concrete sub-tasks and assigns them to appropriate agents (such as trusted computing agents, data cleaning agents, and feature engineering agents), while also managing task priorities and resource allocation. Furthermore, the central control and planning agent is responsible for invoking and orchestrating various agents. For example, the intent recognition agent can be invoked by the central control and planning agent, allowing it to identify the target intent category from the question text.

[0042] The aforementioned orchestrated Agents are a collection of Agents that collaborate to complete specific tasks according to the instructions of the central planning Agent. Each Agent focuses on an independent function (such as data retrieval, computation, and visualization). Each Agent can be independently upgraded and extended through this modular design, and communicates through standardized interfaces to achieve the decomposition and collaboration of complex tasks. The aforementioned orchestrated Agents can be represented in JSON format as follows: {"stage":"central planning","agent_chain":["spatiotemporal alignment","domain knowledge","data selection","code generation","trusted computing","data execution","visualization"],"decision_log":{"fallback_plan":""}}.

[0043] Step S140: The airport data is processed by multiple agents according to the question text to obtain the processing results of the airport data.

[0044] In the implementation of the above scheme, the intent category of the question text is identified by the intent recognition agent in the multi-agent collaborative processing system. The target intent category of the overall control planning agent is used to orchestrate multiple agents in the multi-agent collaborative processing system. This enables the multi-agent collaborative processing system to dynamically assign the intent category of complex question text to multiple agents that can process structured sub-tasks. The orchestrated multiple agents can achieve accurate parsing of cross-domain data association through task cascading and distributed collaboration, thereby overcoming the shortcomings of a single model that cannot take into account both intent recognition and dynamic prediction. Ultimately, this significantly improves the processing capability of complex composite intents.

[0045] As an optional implementation of step S110 above, for example, the electronic device can receive the question text input by the user from the user terminal device. Suppose the question text is "Query the number of flights at a certain city's airport during the 2024 Spring Festival travel rush, and the year-on-year increase compared to 2023 and 2019." This is a complex intent with a vague time range, lacking a specific start and end date for the "Spring Festival travel rush." ​​It needs to be parsed in conjunction with business rules. For example, if the Civil Aviation Administration stipulates that the Spring Festival travel rush is from January 20th to February 20th each year, then "2024 Spring Festival travel rush" can be parsed as January 20th to February 20th, 2024. The processing result of the airport data corresponding to the above question text can be represented in JSON format as: {"agent":"TimeSpaceAlignment","status":"success","output":{"time_ranges": {"2024 Spring Festival Travel Rush":["2024-01-26","2024-03-05"],"2023 Year-on-Year":["2023-01-07","2023-02-15"],"2019 Baseline":["2019-01-21","2019-03-01"]},"holiday_rules_used":["China Spring Festival Travel Rush Standard Date Mapping Table"]}}

[0046] As an optional implementation of step S120 above, for example, the overall control planning agent in a multi-agent collaborative processing system can call the intent recognition agent, so that the intent recognition agent can perform intent recognition on the question text through rule matching and intent recognition model to obtain the target intent category. Specifically, the above implementation of performing intent recognition on the question text through rule matching and intent recognition model may include: obtaining a regular expression in the intent recognition agent that represents a preset intent category rule, and using the regular expression to match the question text to obtain a matching result, determining whether the matching result matches the preset intent category, and if the matching result matches the preset intent category, then determining the preset intent category as the target intent category. Correspondingly, if the matching result does not match the preset intent category, then obtaining the built-in prompt words in the intent recognition agent, and filling the built-in prompt words and the question text into the prompt template to obtain the prompt engineering text, and then inputting the prompt engineering text into the intent recognition model in the intent recognition agent, so that the intent recognition model in the intent recognition agent can perform intent recognition on the question text to obtain the target intent category and the confidence value of the target intent category.

[0047] As an optional implementation of step S120 above, after the intent recognition agent in the multi-agent cooperative processing system performs intent recognition on the question text, it may further include:

[0048] Step S121: Obtain the confidence value of the target intent category.

[0049] An example implementation of step S121 above is as follows: It is understood that after inputting the prompt engineering text into the intent recognition model in the intent recognition agent, in addition to obtaining the target intent category, a confidence value for the target intent category can also be obtained. The aforementioned target intent category can be represented as... The confidence value of this target intent category can be represented as .

[0050] Step S122: If the confidence value of the target intent category is greater than or equal to the first confidence threshold, then the target intent category is directly determined as the final intent category.

[0051] An example implementation of step S122 above is as follows: if the confidence value of the target intent category is greater than or equal to the first confidence threshold, i.e. Then, the intent identification agent directly returns the target intent category to the overall control planning agent, enabling the overall control planning agent to orchestrate multiple agents in the multi-agent collaborative processing system based on the target intent category. This represents the confidence value for the aforementioned target intent category. This represents the first confidence threshold, which can be set according to specific circumstances, such as 90%, 95%, 98%, or 99%.

[0052] Step S123: If the confidence value of the target intent category is less than the first confidence threshold, then the preset language model is called to perform intent recognition on the question text to obtain the candidate intent category and the confidence value of the candidate intent category. The model parameters of the preset language model are less than the model parameters of the post-trained language model used in the intent recognition agent.

[0053] The pre-set language model refers to a language model whose model parameters are smaller than those of the post-trained language model used in the intent recognition agent. Specifically, it can be a Bidirectional Encoder Representation from Transformers (BERT) model or a BGE (BAAI General Embedding) model, etc.

[0054] An example implementation of step S123 above is as follows: if the confidence value of the target intent category is less than the first confidence threshold, i.e. If the text is not found to be true, then the second-round verification step begins, which involves calling the BERT or BGE model to perform intent recognition on the question text, obtaining candidate intent categories and their confidence values. This represents the confidence value for the aforementioned target intent category. Representing the first confidence threshold, the above candidate intent categories can be represented as: The confidence value of this candidate intent category can be expressed as: Understandably, the parameter size of the aforementioned pre-trained language model is smaller than that of the post-trained language model. Therefore, when the error correction mechanism is triggered, the system can prioritize using the pre-trained model, which has lower computational costs, to identify the candidate intent category of the problematic text, rather than relying on a large model for repeated computation. This lightweight collaborative mode reduces hardware resource consumption (such as GPU computing power), while shortening response time and avoiding redundant computation, enabling the system to maintain efficient operation even in high-concurrency scenarios.

[0055] Optionally, the aforementioned intent recognition agent can also include the self-recognition result ( , ) and the results of recognition by the preset language model ( , These are used in conjunction with model fusion error correction; the specific model fusion error correction methods are described below. Among them, The above intent identification agent identifies the target intent category itself. This represents the confidence value for the aforementioned target intent category. This indicates the candidate intent categories identified by the aforementioned preset language model. This represents the confidence value for the above candidate intent categories.

[0056] Step S124: If the candidate intent category is the same as the target intent category, then the candidate intent category is directly determined as the final target intent category.

[0057] An example implementation of step S124 above is as follows: if the candidate intent category is the same as the target intent category, i.e. == If the candidate intent category is directly determined as the final target intent category, it can be represented as: .in, The above intent identification agent identifies the target intent category itself. This indicates the candidate intent categories identified by the aforementioned preset language model. This indicates the target intent category that will ultimately be returned to the master planning agent.

[0058] Step S125: If the candidate intent category is different from the target intent category, and the confidence values ​​of the candidate intent category and the target intent category meet the preset error correction conditions, then the target intent category is corrected through the self-reflection mechanism of the intent recognition agent. The preset error correction conditions include: the confidence values ​​of the candidate intent category and the target intent category are both less than the second confidence threshold, or the difference between the confidence values ​​of the candidate intent category and the target intent category is less than the third confidence threshold.

[0059] Understandably, the aforementioned preset error correction conditions (such as confidence difference thresholds) can be flexibly configured to allow the system to adapt to the accuracy requirements of different business scenarios. For example, in airport data scenarios, if a user's query involves highly sensitive indicators (such as "flight cancellation compensation policy"), the second threshold can be lowered for strict error correction; while in low-risk scenarios (such as "flight schedule query"), the threshold can be relaxed to prioritize response speed.

[0060] An example implementation of step S125 above is: determining that the candidate intent category is different from the target intent category (i.e. Furthermore, the confidence values ​​of the candidate intent category and the target intent category satisfy a preset error correction condition, which includes: both the confidence values ​​of the candidate intent category and the target intent category are less than a second confidence threshold (i.e., ...). Alternatively, the difference between the confidence value of the candidate intent category and the confidence value of the target intent category is less than the third confidence threshold (i.e., If the candidate intent category differs from the target intent category, and the confidence values ​​of both the candidate and target intent categories meet preset error correction conditions, then the target intent category is corrected through the self-reflection mechanism of the intent recognition agent. If this self-reflection mechanism can directly determine the result, then it returns... If the situation is uncertain, an auxiliary prompt can be generated. This prompt is used to collect user feedback, such as "Do you want to perform future data prediction or current data query?" Based on user feedback or predefined strategies, the intent category is ultimately determined. .in, The above intent identification agent identifies the target intent category itself. This represents the confidence value for the aforementioned target intent category. This indicates the candidate intent categories identified by the aforementioned preset language model. This represents the confidence value for the above candidate intent categories. This represents the second confidence threshold, which can be set according to specific circumstances, such as 60%, 70%, 80%, or 85%. This represents the third confidence threshold, which can be set according to specific circumstances, such as 10%, 20%, or 25%.

[0061] Understandably, the above solution utilizes the collaborative constraint of dual confidence thresholds (the first threshold triggers error correction, and the second / third threshold determines the error correction condition). This allows the system to proactively correct errors in low-confidence scenarios, rather than directly outputting unreliable results. For example, if the target intent category is "historical data comparison," but the candidate category generated by the preset model is "real-time data query," and both have confidence levels below the second threshold, the system will correct itself to a more reasonable category through a self-reflection mechanism. This design effectively avoids misclassification by the intent recognition agent due to edge cases (such as a user asking "data from the same period in 2019"), thus improving the system's robustness.

[0062] Please see Figure 3The illustrated embodiment of this application provides a flowchart of the overall control planning agent orchestrating and invoking multiple agents. Optionally, after the intent recognition agent in the multi-agent collaborative processing system identifies the target intent category from the question text, the intent recognition agent can return the target intent category to the overall control planning agent. The overall control planning agent can then orchestrate and sequentially invoke multiple agents in the multi-agent collaborative processing system according to the following process. The target intent category may include: open-ended questions, structured data queries, data prediction, or domain knowledge question answering, etc. Furthermore, the execution status of each sub-agent can be monitored. After each sub-task is completed, the sub-agent returns the execution result and review report to the overall control planning agent. Based on the report status, the overall control planning agent decides whether to continue to the next stage, instruct the corresponding agent to retry, plan to try a new task path, or trigger a circuit breaker strategy, etc., to ensure the stability, security, and controllability of the entire system under high concurrency and complex scenarios.

[0063] As an optional implementation of step S130 above, for example, the overall control planning agent in the multi-agent collaborative processing system dynamically arranges multiple agents in the multi-agent collaborative processing system according to the target intent category. Specifically, it can arrange spatiotemporal alignment agents, domain knowledge agents, data selection agents, code generation agents, trusted computing agents, visualization agents, data cleaning agents, feature engineering agents, and modeling agents to obtain multiple agents after arrangement.

[0064] As an optional implementation of step S140 above, the target intent category may include: open-ended questions (also known as open-ended question-and-answer), and the multiple agents may include: open-ended question-and-answer agents; before processing the airport data based on the question text through the orchestrated multiple agents, it may further include:

[0065] Step S141a: The overall control planning agent calls the large language model LLM in the open question answering agent so that the large language model LLM can reason about the airport data based on the question text and obtain the answer text.

[0066] An example implementation of step S141a above is as follows: Suppose the question text is "Please introduce West Lake in a certain city." Clearly, this is an open-ended question. In this case, the overall control planning agent can call the large language model (LLM) in the open-ended question-answering agent, allowing the LLM to infer the airport data based on the question text. The inference process may include searching the web content for the question text to obtain structured intermediate text and answer text. Here, the structured intermediate text includes, for example, the title, body text, example paragraphs, and summary text.

[0067] Step S141b: The answer text is summarized by the overall control planning agent to obtain the processing results of the airport data.

[0068] For example, step S141b can be implemented by using a master control planning agent to perform operations such as removing duplicate content, completing transitions, and merging summaries on the structured intermediate text and answer text to obtain the processing result of the airport data. Optionally, the master control planning agent can also insert titles, lists, bolding, and citations into the processing result text of the airport data according to the requirements described in the question text, thereby ultimately generating a structured and stylistically consistent processing result text of the airport data.

[0069] As an optional implementation of step S140 above, the multiple agents mentioned above may include: a spatiotemporal alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, and a trusted computing agent. The implementation method described above, which processes airport data based on the question text using multiple orchestrated agents, may include:

[0070] Step S142a: Map the relative time alignment in the question text to standard date and time using the spatiotemporal alignment agent.

[0071] An example implementation of step S142a above is as follows: If the target intent category of the question text is structured data query or domain knowledge question answering, then the relative time alignment in the question text can be mapped to standard date and time. Suppose the question text is "What is the flight delay rate of a certain international airport during the Spring Festival travel rush? Please compare with the data for the same period in 2019." Since the question text does not explicitly state the specific start and end dates of the "Spring Festival travel rush," the spatiotemporal alignment agent can call a standard calendar library, a holiday rule knowledge base, and relative time reasoning tools to resolve "Spring Festival travel rush" as January 20th to February 20th, 2025. The aforementioned holiday rule knowledge base can contain specific business rule parsing, such as the Civil Aviation Administration's regulation that the Spring Festival travel rush is from January 20th to February 20th each year.

[0072] Step S142b: Map the relative spatial alignment in the question text to standard regional locations using the spatiotemporal alignment agent.

[0073] For example, if the target intent category of the above question text is structured data query or domain knowledge question answering, then the spatiotemporal alignment agent can be used to map the relative spatial alignment in the question text to standard regional locations. Suppose the question text is "What is the average flight delay rate in East China? Please compare with the data from the same period in 2019." Since the question text does not explicitly specify the region of "East China," the spatiotemporal alignment agent can call the standard administrative region database, standard geographic location interface, and relative distance reasoning tools to resolve "East China" into provinces and cities such as Jiangsu, Zhejiang, and Shanghai.

[0074] Step S142c: The domain knowledge agent retrieves domain knowledge from the vertical domain knowledge base corresponding to the airport data based on the question text.

[0075] An example implementation of step S142c above is as follows: A domain knowledge agent retrieves domain knowledge from the vertical domain knowledge base corresponding to the airport data based on the question text. The unstructured data can be stored in vector format in the vertical domain knowledge base, which may include background knowledge and business rules, indicator calculation formulas, and definitions of terms related to the airport's vertical domain. It is understood that the domain knowledge agent can use built-in tools to call a vector model (such as a BGE model) in real time to vectorize the user-input question text, obtain a question vector, and calculate the similarity between the question vector and knowledge fragments in the vertical domain knowledge base that are greater than a preset threshold. Then, the domain knowledge agent can integrate and filter these knowledge fragments to obtain integrated and filtered knowledge blocks, and return these integrated and filtered knowledge blocks to the overall control and planning agent.

[0076] The integrated and filtered knowledge blocks mentioned above can be represented in JSON format as follows: {"agent":"DomainKnowledge","status":"success","output":{"recall_knowledge" :["Flight Statistical Scope: Flights that have completed takeoffs and landings (excluding cancellations)","Year-on-Year Formula: (Current Period Value - Base Period Value) / Base Period Value × 100%","Extreme Weather Impact Correction Coefficient Table"], "calculation_formulas":[{"formula_name":"Year-on-Year Increase Calculation","expression":"(Current Period Value - Base Period Value) / Base Period Value * 100%","example":"2024 Increase Compared to 2023 = (2024_total - 2023_total) / 2023_total * 100","constraints":["Base Period Value ≠ 0"]}]}}.

[0077] Step S142d: Select the Agent based on the problem text to determine the list of table names and fields in the target data warehouse.

[0078] An example implementation of step S142d above is as follows: The data selection agent automatically locates and determines the list of table names and fields in the target data warehouse based on the problem text, thereby obtaining the list of table names and fields in the target data warehouse. Then, the data selection agent can also separately return the list of table names and fields in the target data warehouse to the master planning agent, so that the master planning agent can continue to execute or modify the orchestration plan among multiple agents. The table names and field list in the target data warehouse mentioned above can be represented in JSON format as follows: {"agent":"DataSelector","status":"success","output":{"target_tables":[{"db":"airport_warehouse","table":"fact_flight_daily","fields":["flight_date","flight_count","airport_code"],"filter":"airport_code='HGH' AND flight_date BETWEEN{start} AND {end}"}],"query_template":"SELECT flight_date, SUM(flight_count)AS total_flights FROM fact_flight_daily WHERE {conditions} GROUP BY flight_date"}}.

[0079] Step S142e: Generate an SQL query statement by generating the agent through code, based on standard date and time, standard regional location, domain knowledge and / or a list of table names and fields in the target data warehouse.

[0080] An example implementation of step S142e above is as follows: An agent generates an SQL query statement based on standard date and time, standard location, domain knowledge, and / or a list of table names and fields in the target data warehouse. This SQL query statement matches the intent of the user's input question text and is capable of extracting data from the target data warehouse. The generated SQL query statements are as follows: sql_2024 = query_template.format(conditions = "flight_date BETWEEN '2024-01-26' AND '2024-03-05'"); sql_2023 = query_template.format(conditions = "flight_date BETWEEN '2023-01-07' AND '2023-02-15'"); sql_2019 = query_template.format(conditions ="flight_date BETWEEN '2019-01-21' AND '2019-03-01'").

[0081] Step S142f: Execute an SQL query statement in the target data warehouse using the trusted computing agent to obtain the query results corresponding to the problem text.

[0082] For example, step S142f described above can be implemented as follows: The trusted computing agent performs static security audits, execution plan evaluations, and logical verifications on the SQL query statement to obtain security performance verification results. If the security performance verification result is successful, the SQL query statement is executed in the target data warehouse. After execution, the query results corresponding to the problematic text are obtained. If the security performance verification result is unsuccessful, the trusted computing agent returns error information to the overall control planning agent, allowing the overall control planning agent to adjust the plan, decide on retrying, or abort based on the error information.

[0083] As an optional implementation of step S140 above, the target intent category may further include: structured data query, and the multiple agents may further include: visualization agent. After obtaining the query results corresponding to the question text, it may further include:

[0084] Step S143a: The visualization agent generates a visualization script based on the visualization description in the question text and the query results. The visualization script is used to visualize the query results during execution.

[0085] For example, the above step S143a can be implemented by generating a visualization script based on the visualization description in the question text and the query results through a visualization agent, and then visualizing the query results when the visualization script is executed by the visualization agent.

[0086] Step S143b: Perform static security review, feasibility assessment review and / or logical verification review on the visualization script through the trusted computing agent to obtain the visualization review results.

[0087] For example, the implementation of step S143b above involves performing static security review, feasibility assessment review, and / or logical verification review on the visualization script using a trusted computing agent. The obtained visualization review result can be represented in JSON format as: {"agent":"TrustedComputing", "status":"verified","output":{"security_check":{"sql_injection_risk":"pass","syntax_validation":"pass"},"logic_validation":{"date_range_length":[40,40,40],"calculation_consistency":"pass"}}}.

[0088] Step S143c: If the visualization review result is passed, send or execute the visualization script.

[0089] For example, if the visualization review result is passed, and the value of the status field in the visualization review result obtained above is verified, then the visualization review result is passed. At this time, the electronic device can send or execute the visualization script to the target device to obtain the execution result of the visualization script. The execution result of the above visualization script can be represented in JSON format as follows: {"agent":"DataExecutor","status":"success","output": {"2024_total":35800,"2023_total":28755,"2019_total":32492,"2024_growth_2023":{"value":24.5,"formula":"(35800-28755) / 28755*100","formula_source": "DomainKnowledge / calculation_formulas[0]"},"2024_growth_2019":{"value":10.18,"formula":"(35800-32492) / 32492*100","formula_source": "DomainKnowledge / calculation_formulas[0]"}}}.

[0090] As an optional implementation of step S140 above, the target intent category may further include: data prediction, and the multiple agents may further include: trusted computing agent, data cleaning agent, and feature engineering agent. The implementation method of processing airport data based on the question text using multiple orchestrated agents may further include:

[0091] Step S144a: Clean the airport data using the data cleaning agent to obtain cleaned airport data.

[0092] For example, the implementation of step S144a above involves cleaning airport data using a data cleaning agent, such as automatically identifying and correcting missing values, outliers, and duplicates in the original data, effectively eliminating noise interference, and ultimately obtaining cleaned airport data.

[0093] Step S144b: The feature engineering agent generates a feature processing script based on the feature descriptions in the problem text and the cleaned airport data. The feature processing script is used to perform feature derivation, feature filtering and / or feature encoding on the airport data.

[0094] For example, the implementation of step S144b above is as follows: the feature engineering agent generates a feature processing script based on the feature description in the problem text and the cleaned airport data. The feature engineering agent then uses the feature processing script to perform a three-layer review operation on the airport data, namely feature derivation, feature filtering, and / or feature encoding, to obtain the feature processing script after the three-layer review operation.

[0095] Step S144c: Using a trusted computing agent, perform static security review, feasibility assessment review, and / or logical verification review on the feature processing script to obtain the feature review results.

[0096] An example implementation of step S144c above is as follows: A trusted computing agent performs static security review, feasibility assessment review, and / or logical verification review on the feature processing script after the three-layer review operation to obtain the feature review results. Optionally, after obtaining the feature review results, a sample feature set data (usually the first N rows of data) from the feature review results can be returned to the overall control planning agent so that the overall control planning agent can make further plans or adjustments.

[0097] Step S144d: If the feature review result is passed, execute the feature processing script to obtain the data features of the airport data, and use the machine learning model determined based on the question text to predict the data features of the airport data to obtain the processing result of the airport data. The machine learning model is trained using historical airport data.

[0098] An example implementation of step S144d above is as follows: If the feature review result is passed, a feature processing script is executed to obtain the data features of the airport data. Then, using a machine learning model determined based on the question text, the data features of the airport data are predicted to obtain the processing result of the airport data. The machine learning model is trained using historical airport data. The aforementioned machine learning model can be selected from time-series models, regression models, or ensemble models, etc., based on the question text. Optionally, the machine learning model determined based on the question text can be stored in a preset designated path so that when the overall control planning Agent needs to load, it can load and use the machine learning model from the preset designated path, avoiding repeated tasks such as modeling and training the machine learning model.

[0099] As an optional implementation of step S140 above, the target intent category may further include: domain knowledge question answering, and the multiple agents may further include: spatiotemporal alignment agent and domain knowledge agent. The implementation method described above, which processes airport data based on question text using multiple orchestrated agents, may include:

[0100] Step S145a: Map the relative time alignment in the question text to standard date and time using the spatiotemporal alignment agent.

[0101] An example implementation of step S145a above is as follows: If the target intent category of the question text is structured data query or domain knowledge question answering, then the relative time alignment in the question text can be mapped to standard date and time. For instance, if the question text is "What is the flight delay rate at a certain international airport during the Spring Festival travel rush? Please compare with the data from the same period in 2019," since the question text does not explicitly specify the start and end dates of the "Spring Festival travel rush," the spatiotemporal alignment agent can call a standard calendar library, a holiday rule knowledge base, and relative time reasoning tools to resolve "Spring Festival travel rush" as January 20th to February 20th, 2025. The aforementioned holiday rule knowledge base can contain specific business rule parsing, such as the Civil Aviation Administration's regulation that the Spring Festival travel rush is from January 20th to February 20th each year.

[0102] Step S145b: Map the relative spatial alignment in the question text to standard regional locations using a spatiotemporal alignment agent.

[0103] For example, if the target intent category of the above question text is structured data query or domain knowledge question answering, then the spatiotemporal alignment agent can be used to map the relative spatial alignment in the question text to standard regional locations. Suppose the question text is "What is the average flight delay rate in East China? Please compare with the data from the same period in 2019." Since the question text does not explicitly specify the region of "East China," the spatiotemporal alignment agent can call the standard administrative region database, standard geographic location interface, and relative distance reasoning tools to resolve "East China" into provinces and cities such as Jiangsu, Zhejiang, and Shanghai.

[0104] Step S145c: Retrieve multiple knowledge fragments of airport data from a preset domain knowledge base using a domain knowledge agent. The preset domain knowledge base is constructed by converting preset unstructured data into content vectors.

[0105] An example implementation of step S145c above is as follows: A domain knowledge agent retrieves multiple knowledge fragments of airport data from a preset domain knowledge base. This preset domain knowledge base is pre-constructed using a vector model (such as a BGE model) to preprocess preset unstructured data, transforming it into content vectors. The preset unstructured data may include internal documents, maintenance history, formulas, rules, policies, and other unstructured data from the preset airport vertical domain. It is understood that the domain knowledge agent can use built-in tools to call the vector model (such as a BGE model) in real time to vectorize the user-input question text, obtain a question vector, and calculate the similarity between the question vector and knowledge fragments in the vertical domain knowledge base that are greater than a preset threshold. Then, the domain knowledge agent can filter these knowledge fragments to obtain filtered knowledge blocks and return these filtered knowledge blocks to the overall control and planning agent.

[0106] Step S145d: Aggregate multiple knowledge fragments of airport data through the overall control planning agent to obtain the processing results of airport data.

[0107] An example implementation of step S145d above is as follows: the overall control planning agent aggregates and formats multiple knowledge fragments of airport data to obtain a structured airport data processing result. Optionally, if the overall control planning agent needs to supplement or add information to the structured airport data processing result, it can re-scheduling the multiple agents mentioned above to retrieve, supplement, or add information.

[0108] Optionally, a timeout circuit breaker mechanism can be set for any of the multiple agents in the multi-agent collaborative processing system. For example, if the execution time of any of the orchestrated agents exceeds a preset maximum time limit (e.g., 90 seconds), that is, if the overall control planning agent does not receive a return result from any of the multiple agents within the preset maximum time limit, the overall control planning agent will trigger the timeout circuit breaker mechanism. For example, the task executed by the agent can be marked as "timeout failure". Then, the overall control planning agent can choose to optimize retry, downgrade, re-plan the task, or skip the execution stage of the task according to the task priority or preset strategy.

[0109] Optionally, if an execution error occurs during the execution of the visualization script for static security review, feasibility assessment review, and / or logical verification review by the aforementioned trusted computing agent or other data processing agent, the aforementioned overall control planning agent can also record the type and detailed information of this execution error. Furthermore, the overall control planning agent can determine whether the agent corresponding to the execution error is still within the retry count range. If it is, the agent is rescheduled, and an optimized prompt is used to re-call it; or, multiple agents in the multi-agent collaborative processing system are re-planned and orchestrated according to the target intent category. Conversely, if the agent corresponding to the execution error exceeds the retry count range, it is downgraded according to a preset degradation strategy. For example, after calling the agent, it may only return the original data, or a prompt may be displayed indicating that manual intervention is needed in the agent's operation.

[0110] Optionally, during the monitoring of the execution of each sub-agent, the execution status, rollback logs, and context information of each sub-agent are written to the monitoring system. This facilitates bad case identification, improvement, and optimization by staff within the monitoring system. Understandably, if an agent is to be added to a multi-agent collaborative processing system, the agent's information can first be registered in the registry of the central control planning agent. This allows the central control planning agent to invoke registered agents and receive data returned by registered agents.

[0111] Optionally, the scripts mentioned above (such as visualization scripts or feature processing scripts) can be used to generate metadata containing the review results, and the metadata of the review results can be stored in the audit log. The audit log can be stored in an immutable manner, such as storing the audit log in a blockchain, to ensure that the audit trail of each audit log is traceable.

[0112] Please see Figure 4 The diagram shown is a structural schematic of an airport data intelligent processing device provided in an embodiment of this application; this application provides an airport data intelligent processing device 200, including:

[0113] The question text acquisition module 210 is used to acquire question texts related to airport data.

[0114] The intent category recognition module 220 is used to perform intent recognition on the question text through the intent recognition agent in the multi-agent collaborative processing system to obtain the target intent category.

[0115] The intelligent collaborative orchestration module 230 is used to orchestrate multiple agents in the multi-agent collaborative processing system according to the target intent category through the overall control planning agent in the multi-agent collaborative processing system, and obtain multiple orchestrated agents.

[0116] The airport data processing module 240 is used to process airport data based on the question text through multiple orchestrated agents to obtain the processing results of the airport data.

[0117] As an optional embodiment of the above-mentioned device, the airport data intelligent processing device further includes:

[0118] The target confidence value acquisition module is used to obtain the confidence value of the target intent category.

[0119] The candidate confidence value acquisition module is used to call a preset language model to perform intent recognition on the question text if the confidence value of the target intent category is less than the first confidence threshold, thereby obtaining the candidate intent category and the confidence value of the candidate intent category. The model parameters of the preset language model are less than the model parameters of the post-trained language model used in the intent recognition agent.

[0120] The target intent category correction module is used to correct the target intent category through the self-reflection mechanism of the intent recognition agent if the candidate intent category is different from the target intent category and the confidence value of the candidate intent category and the confidence value of the target intent category meet the preset correction conditions. The preset correction conditions include: the confidence value of the candidate intent category and the confidence value of the target intent category are both less than the second confidence threshold, or the difference between the confidence value of the candidate intent category and the confidence value of the target intent category is less than the third confidence threshold.

[0121] As an optional implementation of the above-mentioned device, the target intent category includes: open-ended questions; the multiple agents include: open-ended question-answering agents; the airport data intelligent processing device further includes:

[0122] The LLM model invocation module is used to invoke the large language model (LLM) in the open question-answering agent through the overall control planning agent, so that the large language model (LLM) can infer the airport data based on the question text and obtain the answer text.

[0123] The answer text aggregation module is used to aggregate answer texts through the overall control planning agent to obtain the processing results of airport data.

[0124] As an optional implementation of the above-mentioned device, the multiple agents include: a spatiotemporal alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, and a trusted computing agent; the airport data processing module includes:

[0125] The Time Alignment Mapping submodule is used to map the relative time alignment in the question text to standard date and time using the Spatiotemporal Alignment Agent.

[0126] The Spatial Alignment Mapping submodule is used to map the relative spatial alignment in the question text to standard regional locations using a spatiotemporal alignment agent.

[0127] The domain knowledge retrieval submodule is used to retrieve domain knowledge from the vertical domain knowledge base corresponding to airport data based on the question text using the domain knowledge agent.

[0128] The database table field determination submodule is used by the data selection agent to determine the list of table names and fields in the target data warehouse based on the question text.

[0129] The query statement generation submodule is used to generate SQL query statements by generating an agent based on standard date and time, standard regional location, domain knowledge, and / or a list of table names and fields in the target data warehouse.

[0130] The query execution submodule is used to execute SQL query statements in the target data warehouse through the trusted computing agent to obtain the query results corresponding to the problem text.

[0131] As an optional implementation of the above-mentioned device, the target intent category includes: structured data query; the multiple agents also include: a visualization agent; the airport data processing module further includes:

[0132] The Visual Script Generation submodule is used to generate a visual script based on the visual description in the question text and the query results through the visual agent. The visual script is used to visualize the query results during execution.

[0133] The Visual Script Review Submodule is used to perform static security reviews, feasibility assessment reviews, and / or logical verification reviews on visual scripts through a trusted computing agent, and obtain visual review results.

[0134] The script sending and execution submodule is used to send or execute the visualization script if the visualization review result is passed.

[0135] As an optional implementation of the above-mentioned device, the target intent category includes: data prediction; the multiple agents also include: a data cleaning agent and a feature engineering agent; the airport data processing module further includes:

[0136] The airport data cleaning submodule is used to clean airport data using a data cleaning agent to obtain cleaned airport data.

[0137] The feature script generation submodule is used to generate feature processing scripts based on the feature descriptions in the problem text and the cleaned airport data through the feature engineering agent. The feature processing scripts are used to perform feature derivation, feature filtering and / or feature encoding on the airport data.

[0138] The Feature Script Review Submodule is used to perform static security review, feasibility assessment review, and / or logical verification review on feature processing scripts through a trusted computing agent to obtain feature review results.

[0139] The feature script execution submodule is used to execute the feature processing script if the feature review result is passed, obtain the data features of the airport data, and predict the data features of the airport data by using a machine learning model determined based on the question text to obtain the processing result of the airport data. The machine learning model is trained using historical airport data.

[0140] As an optional implementation of the above-mentioned device, the target intent category includes: domain knowledge question answering; multiple agents include: spatiotemporal alignment agent and domain knowledge agent; the airport data processing module includes:

[0141] The Time Alignment Mapping submodule is used to map the relative time alignment in the question text to standard date and time using the Spatiotemporal Alignment Agent.

[0142] The Spatial Alignment Mapping submodule is used to map the relative spatial alignment in the question text to standard regional locations using a spatiotemporal alignment agent.

[0143] The knowledge fragment retrieval submodule is used to retrieve multiple knowledge fragments of airport data from a preset domain knowledge base using a domain knowledge agent. The preset domain knowledge base is constructed by converting preset unstructured data into content vectors.

[0144] The knowledge fragment aggregation submodule is used to aggregate multiple knowledge fragments of airport data through the overall control planning agent to obtain the processing results of airport data.

[0145] It should be understood that this device corresponds to the above-described intelligent airport data processing method embodiment and is capable of performing the various steps involved in the above method embodiment. The specific functions of this device can be found in the description above, and detailed descriptions are appropriately omitted here. The device includes at least one software functional module that can be stored in memory or embedded in the device's operating system (OS) in the form of software or firmware.

[0146] Please see Figure 5 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. An electronic device 300 provided in this application includes a processor 310 and a memory 320. The memory 320 stores machine-readable instructions executable by the processor 310. When the machine-readable instructions are executed by the processor 310, the method described above is performed.

[0147] This application embodiment also provides a computer-readable storage medium 330, on which a computer program is stored. This computer program is executed by a processor 310 to perform the methods described above. The computer-readable storage medium 330 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0148] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above.

[0149] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0150] It should be understood that the disclosed apparatus and methods can also be implemented in other ways, as provided in the embodiments of this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending primarily on the functions involved.

[0151] Furthermore, the functional modules of each embodiment in this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the reference to terms such as "one embodiment," "some embodiments," "example," "specific example," "some examples," etc., means that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0152] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.

Claims

1. An airport data intelligent processing method, characterized in that, The method comprises: acquiring question text for airport data; performing intent recognition on the question text by an intent recognition agent in a multi-agent collaborative processing system to obtain a target intent category; scheduling a plurality of agents in the multi-agent collaborative processing system according to the target intent category by a general control planning agent in the multi-agent collaborative processing system to obtain the plurality of agents after scheduling; processing the airport data according to the question text by the plurality of agents after scheduling to obtain a processing result of the airport data; wherein after the intent recognition agent in the multi-agent collaborative processing system performs intent recognition on the question text, a confidence value of the target intent category is further acquired; if the confidence value of the target intent category is less than a first confidence threshold, a preset language model is called to perform intent recognition on the question text to obtain a candidate intent category and a confidence value of the candidate intent category, and model parameters of the preset language model are less than model parameters of a post-training language model used in the intent recognition agent; if the candidate intent category is different from the target intent category and the confidence value of the candidate intent category and the confidence value of the target intent category satisfy a preset error correction condition, the target intent category is corrected by a self-reflection mechanism of the intent recognition agent, and the preset error correction condition includes that the confidence value of the candidate intent category and the confidence value of the target intent category are both less than a second confidence threshold, or a difference value between the confidence value of the candidate intent category and the confidence value of the target intent category is less than a third confidence threshold.

2. The method of claim 1, wherein, The target intent category includes an open question, and the plurality of agents includes an open question and answer agent; before the plurality of agents after scheduling processes the airport data according to the question text, the method further comprises: calling a large language model LLM in the open question and answer agent by the general control planning agent to enable the large language model LLM to reason the airport data according to the question text to obtain an answer text; summarizing the answer text by the general control planning agent to obtain the processing result of the airport data.

3. The method of claim 1, wherein, The plurality of agents includes a space-time alignment agent, a domain knowledge agent, a data selection agent, a code generation agent, and a trusted computing agent; and the processing of the airport data by the plurality of agents after scheduling according to the question text comprises: mapping relative time in the question text to a standard date and time by the space-time alignment agent; mapping relative space in the question text to a standard regional location by the space-time alignment agent; retrieving domain knowledge from a vertical domain knowledge base corresponding to the airport data according to the question text by the domain knowledge agent; determining, by the data selection Agent, a table name and a field list in a target data warehouse according to the question text; generating, by the code generation Agent, an SQL query statement according to the standard date and time, the standard area and place, the domain knowledge, and / or the table name and the field list in the target data warehouse; executing, by the trusted computing Agent, the SQL query statement in the target data warehouse to obtain a query result corresponding to the question text.

4. The method of claim 3, wherein, The target intent category includes a structured data query, and the plurality of Agents further includes a visualization Agent; after the query result corresponding to the question text is obtained, the method further includes: generating, by the visualization Agent, a visualization script according to a visualization description in the question text and the query result, the visualization script being used to visualize the query result when executed; performing, by the trusted computing Agent, static security review, feasibility evaluation review, and / or logic verification review on the visualization script to obtain a visualization review result; if the visualization review result passes, sending or executing the visualization script.

5. The method of claim 3, wherein, The target intent category includes data prediction, and the plurality of Agents further includes a data cleaning Agent and a feature engineering Agent; the processing of the airport data according to the question text by the plurality of Agents arranged in sequence further includes: performing, by the data cleaning Agent, data cleaning on the airport data to obtain cleaned airport data; generating, by the feature engineering Agent, a feature processing script according to a feature description in the question text and the cleaned airport data, the feature processing script being used to perform feature derivation, feature screening, and / or feature coding on the airport data; performing, by the trusted computing Agent, static security review, feasibility evaluation review, and / or logic verification review on the feature processing script to obtain a feature review result; if the feature review result passes, executing the feature processing script to obtain data features of the airport data, and using a machine learning model determined according to the question text to predict the data features of the airport data to obtain a processing result of the airport data, the machine learning model being obtained by training on airport historical data.

6. The method of claim 1, wherein, The target intent category includes domain knowledge question answering, and the plurality of Agents includes a spatio-temporal alignment Agent and a domain knowledge Agent; the processing of the airport data according to the question text by the plurality of Agents arranged in sequence includes: aligning, by the spatio-temporal alignment Agent, a relative time in the question text to a standard date and time; aligning, by the spatio-temporal alignment Agent, a relative space in the question text to a standard area and place; retrieving, by the domain knowledge Agent, a plurality of knowledge segments of the airport data in a preset domain knowledge base, the preset domain knowledge base being constructed by converting preset unstructured data into content vectors; The total control planning agent aggregates the multiple knowledge pieces of the airport data to obtain a processing result of the airport data.

7. An airport data intelligent processing device, characterized by, The method comprises the following steps: an airport data question text acquisition module is configured to acquire a question text for airport data; an intent category identification module is configured to identify an intent of the question text by an intent identification agent in a multi-agent collaborative processing system to obtain a target intent category; an intelligent collaborative arrangement module is configured to arrange multiple agents in the multi-agent collaborative processing system according to the target intent category by a total control planning agent in the multi-agent collaborative processing system to obtain the arranged multiple agents; an airport data processing module is configured to process the airport data according to the question text by the arranged multiple agents to obtain a processing result of the airport data; wherein, after the intent identification agent in the multi-agent collaborative processing system identifies the intent of the question text, the method further comprises the following steps: acquiring a confidence value of the target intent category; if the confidence value of the target intent category is less than a first confidence threshold, calling a preset language model to identify the intent of the question text to obtain a candidate intent category and a confidence value of the candidate intent category, the model parameters of the preset language model being less than the model parameters of a post-training language model used in the intent identification agent; if the candidate intent category is different from the target intent category and the confidence value of the candidate intent category and the confidence value of the target intent category satisfy a preset error correction condition, correcting the target intent category by a self-reflection mechanism of the intent identification agent, the preset error correction condition comprising: the confidence value of the candidate intent category and the confidence value of the target intent category are both less than a second confidence threshold, or the difference between the confidence value of the candidate intent category and the confidence value of the target intent category is less than a third confidence threshold.

8. An electronic device, comprising: The method comprises the following steps: a processor and a memory, the memory storing machine-readable instructions executable by the processor, the machine-readable instructions being executed by the processor to perform the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program is executed by the processor to perform the method of any one of claims 1 to 6.

10. A computer program product, characterised in that, The computer program or computer instructions are executed by the processor to perform the method of any one of claims 1 to 6. The computer program or computer instructions are executed by the processor to perform the method of any one of claims 1 to 6.

Images