A method and system for verifying information of safety production personnel in power enterprises based on AI Agent

Abstract

This invention relates to the fields of data processing and artificial intelligence technology, specifically to a method and system for verifying information on safety production personnel in power enterprises based on an AI Agent. The method includes: classifying and grading safety production data, and formulating differentiated protection measures; constructing a standardized safety production indicator library; encapsulating high-frequency query logic into a four-tuple tool and storing it quantitatively in a tool vector library; establishing a three-level mapping mechanism from personnel to roles to tool permissions and data row-level filtering; after user login, obtaining role permissions, vectorizing natural language questions, and semantically matching them with the tool vector library to filter candidate tools; using an AI Agent to select and plan the calling order from candidate tools, extracting parameters and filling in missing values, calling the execution tool and adding permission constraints, desensitizing the results according to permissions, and finally outputting them in the form of natural language, charts, or reports. This invention achieves intelligent safety production supervision and query.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence technology, specifically to a method and system for verifying information of safety production personnel in power enterprises based on an AI Agent. Background Technology

[0002] In the safe production operations of the power industry, personnel information verification is a crucial step in ensuring safe access to the site. Power operations involve high-risk scenarios such as high voltage, high altitude, and live energized environments. The qualifications and compliance of special operation personnel and work supervisors directly determine operational safety. By verifying the validity of special operation certificates and the authorization qualifications of work supervisors, risks such as unlicensed work, expired certificates, unauthorized operations, and impersonation can be prevented at the source. This is an essential requirement for implementing the "Production Safety Law" and the "Power Safety Work Regulations," and also an important line of defense against electric shock and equipment accidents.

[0003] To enhance safety management, AI-based intelligent query technology can accurately verify the compliance of personnel certifications and the validity of authorizations, standardize access to key positions, improve on-site inspection efficiency, help enterprises fulfill their main responsibility for safe production, and promote the digitalization and standardization of personnel management. In practical applications, users can query basic information about units, departments, positions, and personnel using natural language, such as the highest educational level of personnel. Safety production management systems typically include basic personnel and organizational information, safety training records and assessment results, certificate and qualification management, safety reward and punishment records, and occupational health examination summaries. Achieving semantic mapping from natural language to structured query language faces several technical challenges: the need for semantic disambiguation of professional terms in the field of safety production is high; semantic gaps exist in multi-source heterogeneous data; and sensitive data protection and access control must meet strict compliance requirements. At the same time, the underlying database contains a large number of custom tables and complex relationships, creating semantic differences that are difficult for non-technical users to understand.

[0004] Existing technologies have made progress in natural language understanding and structured query transformation, but in the context of safety production personnel information, further solutions are needed for issues such as domain knowledge integration, mapping of business semantics and underlying logic, and data privacy protection. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for verifying the information of safety production personnel in power enterprises based on AI Agent.

[0006] The technical solution of this invention: A method for verifying the information of safety production personnel in power enterprises based on AI Agent, comprising the following specific implementation steps: S1. Classify and grade safety production data based on a three-dimensional, five-layer, six-level strategy, and divide the data into different levels according to the importance of the data and the degree of harm caused by leakage, and formulate differentiated protection measures. S2. Using ETL tools, extract the original safety production management system data to an intermediate database, perform row-to-column conversion to reconstruct the row-based storage into a column-based wide table, perform data cleaning by filling in null values ​​and eliminating ambiguities, establish a mapping conversion between coding and business semantics, and build a standardized safety production indicator library. S3. Encapsulate the high-frequency query logic into a four-tuple tool containing name, parameters, description and SQL template. Use a four-tuple skeleton to define parameters, including indicator type, time range, organizational dimension and filtering conditions. Then, vectorize the tool description text and store it in the tool vector library. S4. Establish a three-level mapping mechanism from personnel to roles, roles to tool permissions, and data row-level filtering. Define standardized safety production roles based on user positions, configure a set of callable tools and department-level data filtering conditions for each role, and automatically add permission constraints when the tool is executed. S5. After the user logs in, load their role and permission context, convert the natural language question into a question vector through an embedding model, perform semantic similarity calculation with the tool description vector in the tool vector library, filter tools with similarity exceeding the threshold as a candidate set, and if the search fails, trigger the traditional NL2SQL fallback execution. S6. Use an AI Agent to select and plan the calling order from candidate tools, extract parameters from the question and fill in missing values, execute the tool through a function call mechanism and add permission filtering conditions, anonymize the results according to permissions, and finally output the answer in the form of natural language, visual charts or analysis reports.

[0007] Preferably, the three-dimensional five-layer six-level strategy in step S1 is as follows: The data is categorized into three dimensions: personal data, public data, and enterprise data. Data assets are divided into five levels according to business characteristics: professional field, business theme, business object, data entity, and data attribute. Based on the importance of the data and the severity of the harm caused by its leakage, the data is divided into six levels from high to low: core data, important data, and general data at levels four, three, two, and one. For core and important data, only specific authorized roles are allowed to query the data through a dedicated de-identification tool, and each access must be audited. For Level 4 sensitive personal data in general data, the default output is anonymized; For first- to third-level data in general data, regular roles are allowed to query and row-level filtering is performed by organizational dimension.

[0008] Preferably, the data cleaning in step S2, including null value completion and ambiguity elimination, includes: automatically completing null values ​​in the education field; deleting or marking automatically generated summary rows in the report; Establish a mapping and conversion between coding and business semantics, specifically mapping job codes to job names and violation codes to descriptions of violation behaviors.

[0009] Preferably, in step S3, the name of the quadruple tool is the tool's unique identifier, the parameter list adopts a quadruple skeleton of indicator type, time range, organizational dimension, and filtering conditions, the description is business semantic description text, and the SQL template is a parameterized SQL statement framework. The pre-built tool templates include basic information tools for querying specific information about a single person, statistical indicator tools for calculating statistical values ​​within a specific dimension and time range, and desensitization tools that pre-set desensitization rules for querying sensitive data based on data classification and grading standards.

[0010] Preferably, the three-level mapping mechanism in step S4 is as follows: The system searches for a user's job title in the organizational structure table based on their user ID and maps it to a standardized role. Each role is authorized to access a set of tools. Each role is bound to data filtering conditions, allowing ordinary team members to view only their department's records, while security specialists can view all company records. When the tool is invoked, a WHERE condition is automatically appended to the generated SQL statement to restrict user parameters to their privileges.

[0011] Preferably, in step S5, the semantic similarity calculation uses cosine similarity, a similarity threshold is set, tools with similarity exceeding the threshold are selected, and the top K tools are selected in descending order of similarity as a candidate tool set. If the candidate toolset is empty, it is marked as a failed search, triggering fallback execution.

[0012] Preferably, the fallback execution is as follows: load the complete database schema of the safety production indicator library, the large model directly generates SQL statements based on the user's natural language questions and database schema, runs the SQL in a sandbox environment, and if an error occurs, it is automatically corrected and retried according to the error message, with a maximum of 3 retries, and finally submits the valid SQL to the database for execution and returns the original result set.

[0013] Preferably, in step S6, the AI ​​Agent is used to select and plan the invocation order from candidate tools, extract parameters from the problem, and fill in missing values, specifically including: Large-scale model inferences user intent to select one or more tools from candidate tools and determine the invocation order; Extracting the parameter values ​​needed for tools from natural language problems; For missing parameters, implicit default values ​​are used for completion, or the user is prompted to complete the parameters through multiple rounds of dialogue, or the parameters are completed by inheriting the previous query subject from the context within the same session.

[0014] Preferably, in step S6, the results are anonymized according to permissions, specifically as follows: Based on the data filtering conditions corresponding to the user role, sensitive fields in the results are processed. When the user does not have the right to view the plaintext, the ID number and mobile phone number are output in a blurred form. When the user has the right to view all data, the plaintext is output. For statistical results, visualization tools are used to generate bar charts or line charts. Based on preset business prompts, the large model is driven to intelligently interpret the results. Alternatively, multiple tools can be called sequentially in chronological order to obtain data, and then the large model is used to generate a structured report containing a summary, data details, trend analysis, and improvement suggestions.

[0015] The technical solution of this invention: A power enterprise safety production personnel information verification system based on AI Agent, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; When the processor executes the computer program, it implements a method for verifying information of safety production personnel in power enterprises based on an AI Agent, as described in any one of claims 1 to 9.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a method and system for verifying information of safety production personnel in power enterprises based on an AI Agent. Through a three-dimensional, five-layer, six-level data classification and grading strategy, it refines the classification of safety production data and formulates differentiated protection measures. Combined with a role-based three-level mapping permission system, it achieves minimized authorization and row-level filtering for data access, effectively preventing the leakage of sensitive information. It also supports anonymized output and audit traceability, significantly improving the level of data security compliance in safety production supervision. Through ETL governance, the original row-based data is reconstructed into column-based wide tables, establishing coded semantic mapping. High-frequency query logic is encapsulated into parameterized tools and vectorized storage, enabling natural language queries to quickly match the most relevant tools, greatly reducing the difficulty of constructing complex queries and response latency, and avoiding the tedious process of repeatedly writing structured query languages. AI is introduced. As the core scheduler, the Agent can autonomously understand user intent, select appropriate toolchains, automatically extract and complete parameters, and sequentially execute multiple tools through a function call mechanism. It supports empty result optimization and automatic error retries, and has multi-turn dialogue completion capabilities, significantly improving its adaptability to complex safety production analysis needs. The system adopts a dual-path mechanism of tool-enhanced retrieval and traditional NL2SQL fallback, ensuring high recall and robustness. Finally, the system can output analysis results in the form of natural language, visual charts, or structured reports, supplemented by intelligent interpretation and improvement suggestions, reducing the technical threshold for safety production supervisors and improving decision support efficiency. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for verifying the information of safety production personnel in power enterprises based on AI Agent, as proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes a method for verifying the information of safety production personnel in power enterprises based on AI Agent, which includes the following specific implementation steps: S1. Based on a three-dimensional, five-layer, six-level strategy, the data of personnel, organization, qualifications, etc. in the safety production management system are classified and graded. Different levels are divided according to the importance of the data and the degree of harm caused by leakage, and differentiated protection measures are formulated for different levels to lay the foundation for subsequent data security access and privacy protection. S2. Using ETL tools, extract the original safety production management system data to an intermediate database, perform row-to-column conversion to reconstruct the row-based storage into a column-based wide table, perform data cleaning such as null value completion and ambiguity elimination, establish a mapping conversion between coding and business semantics, and build a standardized safety production indicator library. S3. Encapsulate the high-frequency query logic into a four-tuple tool containing name, parameters, description and SQL template. Use the four-tuple skeleton of indicator type, time range, organizational dimension and filtering conditions to define parameters, and store the tool description text in the tool vector library after vectorization to realize query generation based on template matching. S4. Establish a three-level mapping mechanism from personnel to roles, roles to tool permissions, and data row-level filtering. Define standardized safety production roles based on user positions, configure the set of tools that can be called and data filtering conditions of departments and other dimensions for each role, and automatically add permission constraints when the tool is executed. S5. After the user logs in, load their role and permission context, convert the natural language question into a question vector through an embedding model, perform semantic similarity calculation with the tool description vector in the tool vector library, filter tools with similarity exceeding the threshold as a candidate set, and if the search fails, trigger the traditional NL2SQL fallback execution. S6. The model, as the core scheduler, selects and plans the calling order from candidate tools, extracts parameters from the problem and fills in missing values, executes the tool through the function call mechanism and adds permission filtering conditions, desensitizes the results according to permissions, and finally outputs the answer in the form of natural language, visualization charts or analysis reports.

[0019] In an optional embodiment, step S1 systematically classifies and grades the on-site worker data in the safety production management system according to national and enterprise data classification and grading strategies. The specific implementation process is as follows: S11. Adopt a "three-dimensional, five-layer, six-level" overall strategy to classify and grade enterprise safety production data; "Three-dimensional": Classifying data from three dimensions: personal data, public data, and enterprise data; "Five Layers": Data assets are divided into five layers according to business characteristics: professional field (layer 1), business theme (layer 2), business object (layer 3), data entity / dataset (layer 4), and data attribute / data item (layer 5). "Level Six": Based on the importance of the data and the degree of harm caused by its leakage, the data is divided into six levels from high to low; Among them, core data (level 6) and important data (level 5) are defined at the national level; general data is further subdivided into level 4, level 3, level 2, and level 1. S12. Based on actual business operations, the data related to the query of safety production personnel information shall be processed in a hierarchical manner, and the specific rules are as follows: Table 1. Hierarchical Strategy Rules ; S13. Based on the above classification results, formulate targeted safety measures: Core / Important Data (Level 5-6): Only authorized roles (such as security directors) are allowed to query this data using dedicated de-identification tools, and each access must be audited. Sensitive personal data (Level 4): The data is de-identified by default. Plain text can only be output when the user has "full view" permission and passes the secondary verification. General data (levels 1-3): Regular role queries are allowed, but row-level filtering is required by organization (e.g., only data from this department can be viewed).

[0020] In an optional embodiment, step S2 involves constructing a standardized safety production indicator library through ETL and data governance, and the specific implementation process is as follows: S21. Use ETL tools to extract all personnel, organization, qualification, training, and violation data from the original safety production management system (such as Oracle or MySQL) to an intermediate database (such as GBase 8a MPP database); during the extraction process, perform basic structured operations such as data type conversion, null value marking, and foreign key relationship identification to break the closed nature of the original system; S22. Load the intermediate database data into the target safety production indicator database. This indicator database serves as a data mart, allowing for moderate redundancy to optimize query performance. Specific governance operations include: A1. Traditional relational databases often use row-based storage to record multi-dimensional attributes of personnel (e.g., each person has multiple certificate records), which leads to complex nesting for natural language queries. This embodiment reconstructs row-based data into column-based storage through row-column conversion. For example, the original certificate table (personnel ID, certificate type, certificate number) is transformed into: personnel ID, high voltage electrician certificate number, high-altitude erection certificate number, ...; after the transformation, querying "Zhang San's high voltage electrician certificate number" can be directly mapped to a single field without the need for related subqueries, improving efficiency by more than 5 times. A2. For null values ​​and ambiguous expressions in the original data, perform the following cleaning rules: Null value completion: For example, for a null value in the "Education" field, if the "Job Level" field displays "Senior Technician", then complete it to "Bachelor's Degree" according to business rules; otherwise, complete it to "Other". Ambiguity elimination: Delete or mark automatically generated summary rows such as "of which", "subtotal", and "total" in reports to prevent them from being incorrectly identified as personnel records during queries; A3. Establish an encoding-semantic mapping dictionary to convert the abstract code in the system into readable text; For example, the job code "ZR-01" is mapped to "Work Supervisor"; the violation code "WZ-102" is mapped to "Not Wearing a Safety Helmet". End users do not need to remember the codes when querying.

[0021] In an optional embodiment, step S3 encapsulates the high-frequency query logic into a parameterized SQL template (referred to as a "tool") and vectorizes it to achieve an efficient query mode of "template matching + parameter filling". The specific implementation process is as follows: S31. Each tool is defined as a quadruple: Tool = ⟨Name, Params, Description, SQL_Template>; Here, Name represents the tool's unique identifier, such as query_basic_info; Params represents the parameter list, using a four-tuple skeleton (metric type, time range, organizational dimension, filter condition), for example: (violation rate, 2025-01-01 to 2025-03-31, operations department, work leader = Li Si); Description represents the business semantic description text, used for subsequent vector retrieval; SQL_Template represents the parameterized SQL statement framework; S32. Based on typical safety production query scenarios, three types of tool templates are pre-set: (1) Basic information tool, used to query specific information of a single person. For example, template: query the {basic information field} of person {person name}; (2) Statistical indicator tools are used to calculate statistical values ​​within a specific dimension and time range. For example, template: Statistical analysis of {department / unit} for {indicator type, such as violation rate} during the period from {start time} to {end time}. (3) Desensitization tools: Based on data classification and grading standards, pre-set desensitization rules for sensitive data queries. For example, template: query the ID number and mobile phone number of the person {person's name}. S33. Concatenate the Name, Params, and Description of each tool into a single text string. Then through the Embedding model Generate vectors: ; Where d is the vector dimension (e.g., 768); The vectors of all tools are stored in the tool vector library for subsequent semantic retrieval.

[0022] In an optional embodiment, step S4 constructs a role-based three-level mapping permission system to ensure the legality of data access and the principle of minimization. The specific implementation process is as follows: S41. Three-level mapping of builders, roles, and permissions: For example, Personnel → Role: Look up the user's position (such as "Work Supervisor", "Work Ticket Issuer", "Work Permit Holder") in the organizational structure table based on the user ID, and map it to a standardized role; Role → Tool Permissions: Each role is authorized to access a set of tools, ToolSet(r); for example, the "Work Permit Holder" role does not have permission to access the tool for querying the "Overall Violation Rate"; Role → Row-level Data Filtering: Each role is bound to a DataFilter(r) condition; for example, ordinary team members can only see records where department = the current user's department; security specialists can see all records where company = the current company; S42. When a tool is invoked, the system automatically appends a WHERE condition to the generated SQL statement; For example, when user "Zhang San" (role: team leader, department "Operations and Maintenance Team 1") calls the stat_violation_rate tool to query the violation rate, the {dept} parameter in the original template will be automatically restricted to only "Operations and Maintenance Team 1" or its sub-department. If "Substation Maintenance Center" is entered, the system will refuse to execute.

[0023] In an optional embodiment, step S5 describes how the system achieves secure and accurate tool retrieval after a user initiates a natural language query. The specific implementation process is as follows: S51. After the user enters their account and password and the system executes the verification process, it loads the user's RoleSet, ToolSet, and DataFilter according to the mapping relationship in step S4, and initializes the session context. S52. The user inputs a natural language question Q (e.g., "Check Zhang's highest educational qualification"), and the same Embedding model as in step S3 is used. Convert it into a question vector: ; S53. Based on user role permissions, filter all de-identification tools from the ToolSet; if the user's question involves sensitive fields (matched by keywords, such as "ID card" or "telephone"), force the de-identification tools to be matched first, and do not downgrade to plaintext tools; S54. Calculate the problem vector With each tool vector in the tool vector library Cosine similarity: ; Set similarity threshold (In this embodiment, the value can be 0.3), selected The tools were selected and sorted in descending order of similarity, with the top K=5 tools chosen as the candidate tool set Candidates; If |Candidates|≥1, then proceed to step S6 (large model scheduling). If |Candidates|=0, then mark it as a failed search and proceed to step S55 (catch-all). S55. When the tool-enhanced search fails, activate the fallback solution: Load the complete database schema (DDL) of the safety production indicator library; the large model directly generates SQL statements based on the user's question Q and DDL; test the SQL in the sandbox environment, and if an error occurs, automatically correct it and retry (up to 3 times) based on the error message; submit the final valid SQL to the database for execution and return the original result set.

[0024] In an optional embodiment, step S6 is completed by the AI ​​Agent core scheduler, which handles tool invocation planning, parameter filling, exception handling, and final result generation. The specific implementation process is as follows: S61. After receiving the candidate tool set (Candidates) and the original question (Q), the large model performs the following sub-steps: B1. The large model infers user intent, selects the most suitable one or more tools from the Candidates, and determines the calling order; for combined queries (e.g., "stat the violation rate of each department last month and list the top three people with the most violations"), the plan is: first call the stat_violation_rate tool to get department statistics; then call the query_top_violators tool to get personnel details; B2. The large model extracts the parameter values ​​required by the tool from Q. For missing parameters, the following strategy is adopted: Implicit default value: If no time range is specified, the default value is "from the current year to the present" or "the last 30 days"; Multi-turn dialogue completion: If key parameters are missing (such as not providing a person's name), the agent will ask the user, "Which person's information do you want to query?"; Context inheritance: In the same session, the query subject of the previous round (such as "Operations and Maintenance Department") can be inherited by the next round without having to re-enter it; The extracted parameters are represented as a set of key-value pairs; B3. Execute each tool sequentially through the Function Call mechanism: fill in the parameters into the SQL template, append the DataFilter conditions determined in step S4, generate the final SQL and submit it to the database, and return the execution result in JSON format; B4. If the tool returns an empty result or an SQL error, the Agent initiates a feedback optimization loop: Empty result: Try fuzzy matching of parameters (e.g., use LIKE instead of =), or expand the time range (e.g., expand from "this month" to "last three months"); SQL error: Automatically correct the field names in the template based on the error type (e.g., field does not exist), and re-execute; Maximum number of retries: 3. If it still fails, a friendly message will be returned, such as "No data was found that matches the criteria. Please check your input information or contact the administrator." S62. After obtaining the query results, perform the final output processing: Based on the data masking requirements in the DataFilter corresponding to the user role, sensitive fields in the results are processed: if the user does not have permission to view plaintext, the formula of the data masking tool in step S3 is applied to blur the output; for example, the ID number is output as 150430***********926 and the mobile phone number is 138****5678; if the user has the "full view" role (such as security director), the plaintext is output. Generate different types of answers based on the query type: Simple Q&A: Output directly in natural language sentences, such as: "Zhang Moumou's highest education is a bachelor's degree"; Data visualization: For statistical results (such as departmental violation rate rankings), the Agent calls visualization tools to generate bar charts and line charts, which are then embedded in the responses. The calling method is a function call. Intelligent Interpretation: Based on pre-defined business prompts (e.g., "Monthly Security Analysis"), the large model is driven to interpret the results. Example interpretation template: "The violation rate of Operations and Maintenance Department 1 this month was 5.2%, an increase of 1.3 percentage points from last month. The main type of violation was 'work permit not carried with you' (accounting for 67%). It is recommended to strengthen on-site spot checks on the carrying of work permits." For complex analysis requests (such as "generate a safety production status report for the previous quarter"), the Agent calls multiple tools sequentially in chronological order to obtain data, and then uses a large model to generate a structured report, which includes sections such as summary, data details, trend analysis, and improvement suggestions.

[0025] Example 2: This invention proposes an AI Agent-based information verification system for safety production personnel in power enterprises, which is used to execute the AI ​​Agent-based information verification method for safety production personnel in power enterprises proposed in Example 1, and includes: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement the AI ​​Agent-based method for verifying the information of safety production personnel in power enterprises, as described in Embodiment 1 above.

[0026] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An AI Agent-based power enterprise safety production employee information verification method, characterized in that, The specific implementation steps include the following: S1. Classify and grade safety production data based on a three-dimensional, five-layer, six-level strategy, and divide the data into different levels according to the importance of the data and the degree of harm caused by leakage, and formulate differentiated protection measures. S2. Using ETL tools, extract the original safety production management system data to an intermediate database, perform row-to-column conversion to reconstruct the row-based storage into a column-based wide table, perform data cleaning by filling in null values ​​and eliminating ambiguities, establish a mapping conversion between coding and business semantics, and build a standardized safety production indicator library. S3. Encapsulate the high-frequency query logic into a four-tuple tool containing name, parameters, description and SQL template. Use a four-tuple skeleton to define parameters, including indicator type, time range, organizational dimension and filtering conditions. Then, vectorize the tool description text and store it in the tool vector library. S4. Establish a three-level mapping mechanism from personnel to roles, roles to tool permissions, and data row-level filtering. Define standardized safety production roles based on user positions, configure a set of callable tools and department-level data filtering conditions for each role, and automatically add permission constraints when the tool is executed. S5. After the user logs in, load their role and permission context, convert the natural language question into a question vector through an embedding model, perform semantic similarity calculation with the tool description vector in the tool vector library, filter tools with similarity exceeding the threshold as a candidate set, and if the search fails, trigger the traditional NL2SQL fallback execution. S6. Use an AI Agent to select and plan the calling order from candidate tools, extract parameters from the question and fill in missing values, execute the tool through a function call mechanism and add permission filtering conditions, anonymize the results according to permissions, and finally output the answer in the form of natural language, visual charts or analysis reports.

2. The AI Agent-based power enterprise safety production employee information verification method according to claim 1, characterized in that, The three-dimensional, five-layer, six-level strategy in step S1 is as follows: The data is categorized into three dimensions: personal data, public data, and enterprise data. Data assets are divided into five levels according to business characteristics: professional field, business theme, business object, data entity, and data attribute. Based on the importance of the data and the severity of the harm caused by its leakage, the data is divided into six levels from high to low: core data, important data, and general data at levels four, three, two, and one. For core and important data, only specific authorized roles are allowed to query the data through a dedicated de-identification tool, and each access must be audited. For Level 4 sensitive personal data in general data, the default output is anonymized; For first- to third-level data in general data, regular roles are allowed to query and row-level filtering is performed by organizational dimension.

3. The AI Agent-based power enterprise safety production employee information verification method of claim 2, wherein The data cleaning process in step S2, which includes filling in missing values ​​and eliminating ambiguities, includes: automatically filling in missing values ​​for the education level field; deleting or marking automatically generated summary rows in the report; Establish a mapping and conversion between coding and business semantics, specifically mapping job codes to job names and violation codes to descriptions of violation behaviors.

4. The AI Agent-based power enterprise safety production employee information verification method according to claim 3, characterized in that, In step S3, the quadruple tool has a name that is a unique identifier for the tool, a parameter list that uses a quadruple skeleton of indicator type, time range, organizational dimension, and filtering conditions, a description of business semantic text, and an SQL template that is a parameterized SQL statement framework. The pre-built tool templates include basic information tools for querying specific information about a single person, statistical indicator tools for calculating statistical values ​​within a specific dimension and time range, and desensitization tools that pre-set desensitization rules for querying sensitive data based on data classification and grading standards.

5. The AI Agent-based power enterprise safety production employee information verification method according to claim 4, characterized in that, The three-level mapping mechanism in step S4 is as follows: The system searches for a user's job title in the organizational structure table based on their user ID and maps it to a standardized role. Each role is authorized to access a set of tools. Each role is bound to data filtering conditions, allowing ordinary team members to view only their department's records, while security specialists can view all company records. When the tool is invoked, a WHERE condition is automatically appended to the generated SQL statement to restrict user parameters to their privileges.

6. The method for verifying information of safety production personnel in power enterprises based on AI Agent according to claim 5, characterized in that, In step S5, the semantic similarity calculation uses cosine similarity. A similarity threshold is set, and tools with similarity exceeding the threshold are selected. The top K tools are sorted in descending order of similarity and taken as the candidate tool set. If the candidate toolset is empty, it is marked as a failed search, triggering fallback execution.

7. A method for verifying information of safety production personnel in power enterprises based on AI Agent as described in claim 6, characterized in that, The fallback execution involves loading the complete database schema of the safety production indicator library. The large model directly generates SQL statements based on the user's natural language questions and the database schema. The SQL is then tested in a sandbox environment. If an error occurs, it is automatically corrected and retried based on the error message. The maximum number of retries is 3. Finally, the valid SQL is submitted to the database for execution and the original result set is returned.

8. The method for verifying information of safety production personnel in power enterprises based on AI Agent according to claim 7, characterized in that, Step S6 utilizes an AI Agent to select and plan the invocation order from candidate tools, extract parameters from the problem, and fill in missing values. Specifically, this includes: Large-scale model inferences user intent to select one or more tools from candidate tools and determine the invocation order; Extracting the parameter values ​​needed for tools from natural language problems; For missing parameters, implicit default values ​​are used for completion, or the user is prompted to complete the parameters through multiple rounds of dialogue, or the parameters are completed by inheriting the previous query subject from the context within the same session.

9. A method for verifying information of safety production personnel in power enterprises based on AI Agent, as described in claim 8, characterized in that, In step S6, the results are anonymized according to permissions, specifically as follows: Based on the data filtering conditions corresponding to the user role, sensitive fields in the results are processed. When the user does not have the right to view the plaintext, the ID number and mobile phone number are output in a blurred form. When the user has the right to view all data, the plaintext is output. For statistical results, visualization tools are used to generate bar charts or line charts. Based on preset business prompts, the large model is driven to intelligently interpret the results. Alternatively, multiple tools can be called sequentially in chronological order to obtain data, and then the large model is used to generate a structured report containing a summary, data details, trend analysis, and improvement suggestions.

10. A power enterprise safety production personnel information verification system based on AI Agent, characterized in that, include: Memory; processor; A computer program stored in the memory and capable of running on the processor; When the processor executes the computer program, it implements a method for verifying the information of employees in the power industry based on AIAgent as described in any one of claims 1 to 9.

Images