An agent-driven automatic hierarchical evaluation method, system, device and medium for multi-type safety production evaluation indexes

Abstract

The present application relates to the technical field of power safety production intelligent evaluation, in particular to a multi-type safety production evaluation index Agent-driven automatic grading evaluation method, system, device and medium. The evaluation indexes are divided into structured quantitative indexes and unstructured semantic analysis indexes, and a multi-entity associated knowledge network covering index entities, data source entities, grading standard entities, calculation tool entities and risk dimension entities is constructed. The Agent intelligent agent queries the knowledge network, respectively plans differentiated tool chains according to the index types and performs grading calculation to obtain initial grading results. The initial grading results are ruled by a multi-condition priority judgment structure taking the basic compliance conditions as root nodes and the step-by-step promotion conditions as child nodes to obtain the index grading evaluation results. The grading evaluation results are taken as input to calculate the scores of each risk dimension, and an early warning notification is triggered when the risk dimension score rate is lower than the preset early warning threshold.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology for power safety production, and in particular to an agent-driven automated classification and assessment method, system, equipment, and medium for multiple types of safety production assessment indicators. Background Technology

[0002] The safety production risk management system is the core institutional framework for power companies to ensure safe and stable operation. Its review work covers hundreds of verification indicators in the entire business field of power production. It is the main means to identify safety hazards and quantify risk levels. Under the traditional review model, the grading calculation of various indicators mainly relies on manual completion. Auditors need to extract data item by item from multiple business platforms and unstructured documents, and compare them item by item according to the grading standards to complete the evaluation.

[0003] At the level of automated indicator processing technology, existing research has used structured data interfaces combined with script tools to automate the calculation of some numerical indicators, but it has failed to cover semantic indicators that rely on unstructured documents. Other research has built indicator association models based on knowledge graphs to achieve structured storage of indicator definitions, but lacks the ability to coordinate and schedule with calculation tools. The introduction of Retrieval Enhanced Generation (RAG) technology provides a new approach to extracting key information from unstructured documents, but existing solutions still require manual intervention for hierarchical judgment, failing to achieve end-to-end fully automated evaluation capabilities.

[0004] The above methods have three common shortcomings: First, they lack differentiated processing solutions for the heterogeneity of data sources, and it is difficult to adapt to both structured numerical indicators and unstructured indicators that rely on multi-condition semantic judgments using a single processing path. Second, they have limited ability to analyze grading standards. For tiered grading rules with multi-level conditions, existing methods are prone to omissions of conditions and logical errors, resulting in low accuracy of evaluation results. Third, the system has poor adaptability. When new review indicators are added or grading standards are revised, the judgment logic needs to be redeveloped and the reviewers need to be retrained, resulting in high reconstruction costs. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the problem to be solved by this invention is how to address the problems of single processing path, missed judgment in multi-condition classification and determination, and poor system adaptability in the existing technology for calculating the classification of safety production assessment indicators.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an agent-driven automated hierarchical assessment method for multiple types of safety production review indicators, which includes, based on the heterogeneity of data sources, dividing the review indicators into structured quantitative indicators and unstructured semantic parsing indicators, and constructing a multi-entity association knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities and risk dimension entities. The agent queries the multi-entity association knowledge network, plans differentiated toolchains according to the type of the evaluation indicators, and performs hierarchical calculations to obtain the initial hierarchical results. The initial grading results are adjudicated using a multi-condition priority judgment structure to obtain the indicator grading evaluation results. The scores for each risk dimension are calculated using the results of the index grading assessment as input, and an early warning notification is triggered when the risk dimension score rate is lower than the preset early warning threshold.

[0008] As a preferred embodiment of the Agent-driven automated hierarchical assessment method for multiple types of safety production review indicators described in this invention, the step of dividing the review indicators into structured quantitative indicators and unstructured semantic parsing indicators includes querying the data format attributes of the review indicators in the multi-entity association knowledge network, identifying indicators with structured data source format as structured quantitative indicators, and identifying indicators with unstructured data source format as unstructured semantic parsing indicators. When there is an inconsistency between the type attribute label of the review indicator and the format of the associated data source, the review indicator will be marked as an abnormal indicator, and a manual verification process will be initiated to confirm and adjust the indicator type.

[0009] As a preferred embodiment of the Agent-driven automated hierarchical assessment method for multiple types of safety production review indicators described in this invention, the construction of a multi-entity association knowledge network includes extracting the name, hierarchical judgment rules and associated risk dimension information of each review indicator from the review standard document. A combination of fuzzy name matching and manual verification is used to establish the relationships between indicator entities and data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities; The multi-entity association knowledge network is stored in a graph database. When a new review indicator is added, the corresponding entity and its association with the data source entity, the hierarchical standard entity, and the risk dimension entity are automatically added.

[0010] As a preferred embodiment of the Agent-driven automated grading and evaluation method for multiple types of safety production review indicators described in this invention, the step of planning differentiated toolchains according to the type of the review indicator includes, for structured quantitative indicators, planning a toolchain that coordinates structured data interface calls and programming calculation tools, extracting raw data from structured data sources and then performing numerical calculations to obtain intermediate numerical calculation results. For unstructured semantic parsing metrics, a toolchain is planned to collaborate with document information extraction tools and knowledge network qualification verification. After extracting effective information from unstructured data sources and filtering out invalid qualification data, semantic judgment is performed to obtain statistical intermediate results. The agent standardizes the intermediate computation results output by the toolchain and pushes them to the multi-condition priority determination structure to trigger hierarchical adjudication.

[0011] As a preferred embodiment of the Agent-driven automated classification and evaluation method for multiple types of safety production review indicators described in this invention, the method further includes: obtaining intermediate calculation results of the review indicators, validating the intermediate calculation results, and performing differentiated correction processing for different categories of abnormal data. Based on the hierarchical judgment rules of the review indicators, a hierarchical adjudication structure is constructed with basic compliance conditions as the root node and progressively improving conditions as sub-nodes. The hierarchical adjudication structure is traversed in priority order from the root node to the child nodes. When the condition of any node is met, the corresponding classification result is output and the subsequent traversal is terminated to obtain the index classification evaluation result. Using the rating results of the aforementioned indicators as input, differentiated scoring rules are applied according to the level to which the assessment subject belongs to calculate the expected score and actual score for each risk dimension. Calculate the score rate for each risk dimension, and trigger an early warning notification to the corresponding responsible department when the score rate is lower than the preset early warning threshold.

[0012] As a preferred embodiment of the Agent-driven automated classification and assessment method for multiple types of safety production review indicators described in this invention, the differential correction processing for different categories of abnormal data includes identifying data whose qualification validity period has exceeded the limit, removing the data from the valid dataset, and then re-triggering the classification decision. Identify and mark incorrectly entered data, and initiate a manual review process; The identification tool retrieves incorrect data and then calls the corresponding calculation tool again to correct the calculation results.

[0013] As a preferred embodiment of the Agent-driven automated hierarchical assessment method for multiple types of safety production review indicators described in this invention, wherein: the differentiated scoring rules are adopted according to the level to which the assessment subject belongs, including, for the three-level unit level, calculating the actual score of the risk dimension as the sum of the weight values ​​of all compliance indicators under the risk dimension; At the provincial level, the actual score of the risk dimension is calculated as the sum of the weight values ​​of each indicator under the risk dimension after being reduced according to the proportion of non-compliant units. The more non-compliant units there are, the greater the reduction.

[0014] Secondly, embodiments of the present invention provide an agent-driven automated grading and evaluation system for multiple types of safety production review indicators, which includes an indicator classification module for classifying review indicators into structured quantitative indicators and unstructured semantic parsing indicators based on the heterogeneity of data sources. The knowledge network construction module is used to build a multi-entity relational knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities. The Agent scheduling module is used to query the multi-entity association knowledge network and plan differentiated toolchains according to the indicator types to perform hierarchical calculations and obtain initial hierarchical results. The grading adjudication module is used to adjudicate the initial grading results through a multi-condition priority judgment structure to obtain the indicator grading evaluation results. The risk warning module is used to calculate the score of each risk dimension based on the indicator classification assessment results, and to trigger a warning notification when the score rate is lower than the preset warning threshold.

[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the Agent-driven automated classification and assessment method for multiple types of safety production review indicators as described in the first aspect of the present invention.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the Agent-driven automated classification and assessment method for multiple types of safety production review indicators as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are as follows: This invention constructs a heterogeneous dual-type indicator classification system and a multi-entity association knowledge network through S100. It clearly distinguishes review indicators into structured quantitative indicators and unstructured semantic parsing indicators based on data source characteristics. Furthermore, it unifies indicator entities with data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities into a graph database management system. This fundamentally solves the problem that the traditional "one-size-fits-all" processing path cannot adapt to heterogeneous data sources. Moreover, the knowledge network supports automatic association when indicators are dynamically added, avoiding the high cost of manually reconstructing the calculation process. In S200, the agent intelligent body automatically plans a differentiated toolchain based on the knowledge network, and structured quantitative indicators are extracted using SQL. The S300 employs a numerical calculation path, while the unstructured semantic parsing indicators utilize a RAG document extraction and qualification verification path. Each path is adapted to its corresponding indicator type, eliminating extraction errors for semantic indicators inherent in single-tool solutions. The S300 uses a logical tree priority adjudication structure with basic compliance conditions as the root node, traversing nodes sequentially and terminating immediately when conditions are met. This specifically addresses the issue of missed judgments caused by disordered condition traversal order in multi-condition grading standards, improving the accuracy of grading adjudication. The S400 uses grading assessment results to drive differentiated scoring of risk dimensions and configurable threshold warnings, achieving a closed-loop process for grading calculation and risk control, thus overcoming the shortcomings of traditional reviews where grading calculation and risk warning are disconnected. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Fig. 1 A flowchart for an agent-driven automated grading and assessment method for multiple types of safety production review indicators; Fig. 2 A computer equipment diagram for an agent-driven automated grading and assessment method for multiple types of safety production review indicators; Fig. 3 A knowledge graph of five entity relationships for an agent-driven automated grading and assessment method for multiple types of safety production review indicators; Fig. 4 A flowchart for Agent-driven automated grading and assessment method for multiple types of safety production review indicators; Fig. 5 A logic tree flowchart for a multi-level judgment engine for an agent-driven automated grading and evaluation method for various types of safety production review indicators. Fig. 6A flowchart of a risk dimension assessment system for an agent-driven automated grading and assessment method for multiple types of safety production review indicators. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1 Reference Figs. 1-2 This is the first embodiment of the present invention, which provides an agent-driven automated hierarchical assessment method for multiple types of safety production review indicators, including: S100: Based on the heterogeneity of data sources, the review indicators are divided into structured quantitative indicators and unstructured semantic parsing indicators, and a multi-entity association knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities and risk dimension entities is constructed. S200: The agent queries the multi-entity association knowledge network, plans differentiated toolchains according to the type of the evaluation indicators, and performs hierarchical calculations to obtain the initial hierarchical results; S300: The initial classification results are adjudicated through a multi-condition priority judgment structure to obtain the index classification evaluation results; S400: Calculates scores for each risk dimension using the indicator grading assessment results as input, and triggers an early warning notification when the risk dimension score rate is lower than the preset early warning threshold.

[0024] It should be noted that in the existing safety production review system, the verification items and indicators have fundamental differences in three dimensions: data source, calculation basis, and evaluation standards. The data of some indicators comes from structured databases that can be directly queried, and the calculation process is a deterministic numerical operation that can yield results according to fixed formulas. The data of other indicators are scattered in multiple unstructured documents, and the evaluation standards rely on comprehensive judgment of multiple conditions, making it impossible to use a unified numerical calculation path for processing.

[0025] S100 categorizes review indicators into two types based on the heterogeneity of data sources, and uses a multi-entity association knowledge network to uniformly associate and store indicators, data sources, grading standards, calculation tools, and risk dimensions. S200, relying on this knowledge network, automatically plans and executes differentiated toolchains according to indicator type, completing structured numerical calculations and extracting unstructured document information, outputting standardized intermediate calculation results. S300 uses a logical tree structure with basic compliance conditions as the root node and progressively ascending conditions as child nodes, adjudicating node by node according to priority, thus resolving the issue of missed judgments in multi-condition grading standards. S400 uses the grading assessment results as input, calculates the score rate of each risk dimension differentiated according to the level of the assessment subject, and triggers an early warning notification to the corresponding responsible department when the score rate falls below a preset warning threshold, achieving a closed-loop process from grading calculation to risk warning.

[0026] Example 2 Reference Figs. 1-6 This is the second embodiment of the present invention.

[0027] In this embodiment, step S100, based on the heterogeneity of data sources, divides the review indicators into structured quantitative indicators and unstructured semantic parsing indicators, and constructs a multi-entity association knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities, including the following steps A1-A2: A1: The evaluation indicators are divided into structured quantitative indicators and unstructured semantic parsing indicators. This includes querying the data format attributes of evaluation indicators in a multi-entity association knowledge network, identifying indicators with structured data source format as structured quantitative indicators, and identifying indicators with unstructured data source format as unstructured semantic parsing indicators. When there is an inconsistency between the type attribute labeling of the review indicator and the format of the associated data source, the review indicator will be marked as an abnormal indicator, and a manual verification process will be initiated to confirm and adjust the indicator type.

[0028] Specifically, structured quantitative indicators are defined as verification items with clearly defined data sources, calculation bases, and evaluation standards. The data formats of their associated data sources are structured formats such as MySQL databases or application programming interfaces (APIs), and the calculation results can be directly obtained through numerical calculations. Taking the "pre-test plan completion rate" indicator as an example, its data source is the production plan database in the power grid management platform. The calculation basis is the ratio of the number of completed pre-test tasks to the planned number of pre-test tasks. The graded evaluation standards are as follows: Level 3 corresponds to a completion rate of no less than 98%, Level 4 corresponds to a completion rate of no less than 98.5%, and Level 5 corresponds to a completion rate of no less than 99%. The data source, calculation logic, and grading standards are all clearly defined, making it a structured quantitative indicator.

[0029] Unstructured semantic parsing indicators are defined as verification items whose data sources are complex, whose calculation basis is subjective, and whose evaluation criteria require comprehensive judgment based on multiple conditions. The data sources associated with these indicators are in unstructured document formats such as PDF or Word, and calculations require semantic understanding and information extraction. Taking the indicator of "number of senior auditors" as an example, its data sources include multiple unstructured qualification documents such as the "List of Senior Qualified Personnel for Safety Production Risk Management System Review" and the "List of Intermediate Qualified Reviewers for Safety Production Risk Management System." The calculation basis is the statistical count of the number of intermediate and senior auditors in these documents. The grading standard is: Level 3 is achieved when there are no fewer than 5 intermediate auditors, Level 4 when there is no fewer than 1 senior auditor, and Level 5 when there are no fewer than 2 senior auditors. This is an unstructured semantic parsing indicator that requires judgment based on multiple conditions.

[0030] When a review indicator is labeled as a structured quantitative indicator but its associated data source is in an unstructured format such as PDF or Word, or when a review indicator is labeled as an unstructured semantic parsing indicator but its data source is in a structured format such as MySQL or API, it is determined to be an abnormal indicator. The system will automatically mark the indicator as an abnormal indicator, stop the automatic calculation process, and transfer it to the manual verification and confirmation stage. After verification by the reviewers, the matching relationship between the indicator type and the data source will be manually adjusted to ensure the correct execution of subsequent toolchain planning.

[0031] A2: Construct a multi-entity related knowledge network, including extracting the names, hierarchical judgment rules and related risk dimension information of each review indicator from the review standard document; A combination of fuzzy name matching and manual verification is used to establish the relationships between indicator entities and data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities; The multi-entity association knowledge network is stored in a graph database, and when a new review indicator is added, the corresponding entity and its association with the data source entity, the hierarchical standard entity, and the risk dimension entity are automatically added.

[0032] Specifically, the ontology design of the multi-entity association knowledge network covers five core entities, with the definitions and attributes of each entity as follows. The indicator entity encompasses all review and audit indicators, with core attributes including indicator number, indicator name, indicator type (structured quantitative indicator or unstructured semantic parsing indicator), calculation logic description, association grading standard number, and association risk dimension. Taking the indicator "Number of Senior Auditors" as an example, its attribute assignments are: Number IND-089, Name "Number of Senior Auditors", Type "Unstructured Semantic Parsing Indicator", Calculation Logic "Statistics of the Number of Intermediate and Senior Auditors", Association Grading Standard Number STD-23, and Association Risk Dimension "Other Risks".

[0033] Data source entities are divided into two categories: structured data sources and unstructured data sources. The core attributes include data source number, data source name, data format, and access interface. Structured data sources, such as the power grid management platform, provide a MySQL database interface, while unstructured data sources, such as qualification list documents, are recorded in the form of file paths.

[0034] The grading standard entity is used to store all the grading judgment rules corresponding to each review indicator. The attributes include the standard number, judgment conditions and corresponding level. Taking the grading standard STD-23 for "number of senior auditors" as an example, its judgment conditions are as follows: no less than 5 intermediate auditors corresponding to level 3, no less than 1 senior auditor corresponding to level 4, and no less than 2 senior auditors corresponding to level 5.

[0035] Diamond Level Standard Entity: Used to store all diamond level judgment rules corresponding to 305 verification items. Attributes include "Standard ID, Judgment Condition, Corresponding Diamond Level". Taking the diamond level standard of "Number of Senior Auditors" as an example, the attribute value is "ID:STD-23, Judgment Condition: Intermediate ≥ 5 → 3 Diamonds, Senior ≥ 1 → 4 Diamonds, Senior ≥ 2 → 5 Diamonds, Corresponding Diamond Level: 3 Diamonds / 4 Diamonds / 5 Diamonds".

[0036] The computational tools entity includes the Retrieval Enhancement Generation Tool (RAG), the Python programming tool, and the structured data interface. Attributes include tool number, function description, and applicable indicator types. The RAG tool is suitable for unstructured semantic parsing indicators, while the Python programming tool is suitable for the numerical computation of both structured quantitative indicators and unstructured semantic parsing indicators. The risk dimension entity corresponds to eight core safety risks: power grid risk, equipment risk, personal injury risk, environmental risk, occupational health risk, information network security risk, public safety risk, and other risks. Core attributes include risk dimension number, risk dimension name, and associated warning threshold.

[0037] The construction of a multi-entity association knowledge network follows a three-stage process: data collection, entity alignment, and storage and updating. In the data collection stage, core information such as the names of each indicator, grading rules, and associated risk dimensions are extracted from the review standard documents. Simultaneously, data source information is obtained from the data platform and qualification documents. In the entity alignment stage, an edit distance algorithm is used to perform fuzzy name matching, with a matching threshold of 0.8. When the calculated similarity is not lower than 0.8, the association between the indicator entity and the corresponding data source entity, grading standard entity, and risk dimension entity is automatically established. When the similarity is lower than 0.8, manual verification is triggered, and auditors manually confirm and establish the association. Taking the "number of senior auditors" indicator as an example, the system uses the edit distance algorithm to accurately associate it with the "List of Senior Qualified Personnel for Safety Production Risk Management System Review" document, the "List of Intermediate Qualified Reviewers for Safety Production Risk Management System" document, the grading standard STD-23, and the "Other Risks" dimension.

[0038] During the storage and update phase, Neo4j graph database is used to store the multi-entity association knowledge network. Neo4j natively supports graph structure storage of nodes and edges, which can efficiently perform multi-entity association queries and adapt to the multi-hop association retrieval requirements of knowledge networks. When a new review indicator is added, the system automatically adds the corresponding indicator entity node in the graph database and automatically establishes its association relationship with the data source entity, hierarchical standard entity, tool entity and risk dimension entity according to the above entity alignment rules, without the need for manual reconstruction of the knowledge network structure.

[0039] In this embodiment, step S200 involves the Agent querying the multi-entity association knowledge network, planning differentiated toolchains based on the type of the evaluation indicators, and performing hierarchical calculations to obtain initial hierarchical results, including the following step B1: B1: Based on the type of the review indicators, plan differentiated toolchains, including, for structured quantitative indicators, plan a toolchain that coordinates structured data interface calls and programming calculation tools, extract raw data from structured data sources and then perform numerical calculations to obtain intermediate numerical calculation results; For unstructured semantic parsing metrics, a toolchain is planned to collaborate with document information extraction tools and knowledge network qualification verification. After extracting effective information from unstructured data sources and filtering out invalid qualification data, semantic judgment is performed to obtain statistical intermediate results. The agent standardizes the intermediate computation results output by the toolchain and pushes them to the multi-condition priority determination structure to trigger hierarchical adjudication.

[0040] Specifically, for structured quantitative indicators, the toolchain executes in the following order: structured data interface call, numerical calculation, and hierarchical standard matching. In the structured data interface call phase: the agent intelligent body automatically generates the corresponding SQL query statement based on the access interface attributes of the data source entity associated with the indicator, and initiates a query request to structured data sources such as the power grid management platform to extract the raw numerical data required for the indicator. Taking the "pre-test plan completion rate" indicator as an example, the system extracts two raw data points through SQL statements: the planned number of pre-test tasks completed and the actual number of pre-test tasks completed. In the numerical calculation phase: the agent intelligent body calls the Python programming tool to perform numerical calculations according to the indicator calculation logic. For ratio-type structured quantitative indicators, the following percentage calculation formula is used: rate = actual / total × 100% Where actual represents the actual completed quantity, total represents the planned total quantity, and rate represents the calculated ratio. Taking "preliminary test plan completion rate" as an example, if actual is 49 and total is 50, then rate = 49 / 50 × 100% = 98%, resulting in the numerical intermediate calculation result "preliminary test plan completion rate = 98%". In the grading standard matching stage: the agent calls the multi-entity association knowledge network to retrieve the grading standard entities associated with the indicator. It compares the numerical intermediate calculation result with the judgment conditions of each level in the grading standard entities one by one, outputting the initial grading result. Taking "preliminary test plan completion rate = 98%" as an example, according to the grading standard "level 3 corresponds to no less than 98%", the initial grading result is level 3.

[0041] For unstructured semantic parsing metrics, the toolchain executes in the following order: RAG unstructured information extraction, data format standardization, knowledge network qualification verification, and multi-condition priority determination. In the RAG information extraction stage: Retrieval-Augmented Generation (RAG) is an information extraction technique that combines document retrieval with language model generation. In this solution, the agent generates domain-specific prompts for the safety production review scenario based on the document path attributes of the data source entities associated with the metrics. Taking the "number of senior auditors" metric as an example, the prompt content is "Extract the names, number, and certificate dates of intermediate and senior auditors from the qualification certificate documents." Subsequently, a finely tuned Bidirectional Encoder-Representations from Transformers (BERT) model is called to perform information extraction from the documents. The basic architecture of the BERT fine-tuning model adopts the standard BERT encoder structure, which includes an embedding layer, a multi-layer bidirectional Transformer encoding layer, and a task-specific output layer. The embedding layer is responsible for converting the input text into word vector representations. The multi-layer bidirectional Transformer encoding layer performs deep semantic encoding on the text through self-attention. The task-specific output layer is designed for the task of extracting auditor qualification information and outputs a structured label sequence containing the auditor's name, qualification level, and certificate period.

[0042] Fine-tuning training uses historical review data and qualification document datasets in the field of safety production review as training corpus. The dataset consists of 500 unstructured documents. The learning rate is set to 1e-5 to avoid over-adjustment of pre-training parameters, and the number of iterations is 100. The optimal model parameters are determined based on the convergence of information extraction accuracy on the validation set. The model input is the original text sequence of qualification documents, and the output is the structured extraction result containing the number of auditors, qualification level, and period fields. In the data format standardization stage, the unstructured extraction results from RAG are converted to a standardized format. The number of auditors field is converted to an integer type, and the qualification period field is converted to the standard date format "YYYY-MM-DD" to ensure that the input data format for subsequent qualification verification and grading adjudication is consistent.

[0043] Knowledge Network Qualification Verification Phase: The agent invokes qualification rules within the multi-entity association knowledge network to filter auditor qualification data whose validity period exceeds the specified years. Using a qualification period not exceeding three years as a constraint, data of auditors with expired certificates are excluded from the dataset, ensuring that all data participating in the grading calculation are qualified. The agent standardizes the intermediate calculation results output from the above two toolchains according to a unified format and pushes them to the multi-condition priority determination structure to trigger subsequent grading decisions.

[0044] In this embodiment, step S300 uses a multi-condition priority determination structure to adjudicate the initial classification result and obtain the index classification evaluation result, including the following C1 step: C1: The initial classification results are adjudicated through a multi-condition priority judgment structure to obtain the index classification evaluation results.

[0045] Specifically, the multi-condition priority judgment structure employs a judgment algorithm that combines logic tree parsing with priority execution to achieve accurate adjudication for situations where there are multi-level hierarchical conditions in the review indicators. The logic tree is constructed as follows: the basic compliance conditions in the indicator grading standard entity are taken as the root node of the logic tree, and the progressively higher conditions are taken as child nodes in sequence. Each node contains three elements: the condition expression, the output result when the condition is met, and the output result when the condition is not met. The basic compliance conditions are defined as the judgment conditions that must be met to reach the minimum grading threshold, and the advanced conditions are defined as the judgment conditions that must be met to further improve the grading on the premise that the basic compliance conditions are met. The division is based on the logical priority of the corresponding conditions at each level in the grading standard entity, with the condition corresponding to the lowest level as the root node, and the child node hierarchy is constructed sequentially upwards.

[0046] Taking the "Number of Senior Auditors" indicator as an example, its logic tree uses "Number of Intermediate Auditors Judgment" as the root node. The root node condition expression is "The number of intermediate auditors is no less than 5". When the root node condition is met, the child node judgment is entered. The condition expression of child node one is "The number of senior auditors is no less than 2". If it is met, level 5 is output; if not, the process moves to child node two. The condition expression of child node two is "The number of senior auditors is no less than 1". If it is met, level 4 is output; if not, level 3 is output. If the root node condition is not met, "Not up to level 3" is output directly, and the subsequent node traversal is terminated. In the priority execution phase, the logic tree is traversed in order from the root node to the child nodes. When traversing any node, its condition expression is checked. If the condition is true, the corresponding grade result is output and the traversal is terminated immediately without executing subsequent node checks. This avoids the problem of missed checks due to missing conditions or incorrect judgment order in multi-condition grade standards. Taking the "number of senior auditors" indicator as an example, if the number of intermediate auditors is 3, the root node condition is not met, and "not up to level 3" is directly output, skipping all child node checks. If the number of intermediate auditors is 6 and the number of senior auditors is 2, the root node condition is met, and the first child node check is performed. If the first child node condition is met, level 5 is output, the traversal is terminated, and the final grade assessment result of the indicator is obtained.

[0047] In this embodiment, step S400 calculates the score for each risk dimension using the indicator grading assessment result as input, and triggers an early warning notification when the risk dimension score rate is lower than a preset early warning threshold, including the following steps D1-D3: D1: also includes obtaining intermediate calculation results of the review indicators, validating the intermediate calculation results, and performing differentiated correction processing for different categories of abnormal data. Based on the hierarchical judgment rules of the review indicators, a hierarchical adjudication structure is constructed with basic compliance conditions as the root node and progressively improving conditions as sub-nodes. Traverse the hierarchical decision structure in priority order from the root node to the child nodes. When the condition of any node is met, output the corresponding classification result and terminate the subsequent traversal to obtain the index classification evaluation result. Using the results of the indicator classification assessment as input, differentiated scoring rules are adopted according to the level to which the assessment subject belongs to calculate the expected score and actual score of each risk dimension; Calculate the score rate for each risk dimension, and trigger an early warning notification to the corresponding responsible department when the score rate is lower than the preset early warning threshold.

[0048] Specifically, in the safety production review system, all review indicators are divided into three categories based on their management importance and assigned differentiated weights. The first category consists of 60 veto indicators, covering items involving significant safety hazards or directly affecting system compliance (such as "the occurrence of a major personal safety accident"), with a weight of 0.75. The second category comprises 178 ranking indicators, which are items that can be compared horizontally among different third-level units, reflecting differences in management levels, with a weight of 0.25. The third category consists of 67 non-rankable indicators, which are items that are difficult to compare horizontally and only require assessment of compliance, with a weight of 0.15. When all review indicators are met, the full score is calculated as (60 × 0.75) + (178 × 0.25) + (67 × 0.15) = 99.55 points.

[0049] The grading results of each indicator are compared with the annual target grading values ​​issued by the provincial management. When the actual grading result of a review unit reaches or exceeds its corresponding target grading value, the unit is deemed to have met the standard for that indicator and receives a score equal to its weight value. Otherwise, it is deemed to have failed to meet the standard, and the score is 0. For the 178 ranking indicators, a ranking score is added to the score for meeting the standard. This ranking score reflects the relative management level of the review unit within the province and serves as a supplement to the score for meeting the standard, together forming the unit's total score.

[0050] D2: Perform differentiated correction processing for different categories of abnormal data, including identifying data whose qualification validity period has exceeded the limit, removing the data from the valid dataset and re-triggering the classification decision; Identify and mark incorrect data, and initiate a manual review process; The identification tool retrieves incorrect data and then calls the corresponding calculation tool again to correct the calculation results.

[0051] Specifically, the judgment and correction strategies for abnormal data categories are handled according to the causes of the abnormalities. For data whose qualification period has exceeded the limit, the system checks the auditor certificate period field extracted from the BERT model based on the qualification rules in the multi-entity association knowledge network. Auditor data with expired certificates are marked as invalid qualification data and removed from the dataset participating in the grading calculation. The updated data is then used to re-trigger the multi-condition priority judgment structure to execute the grading decision, ensuring that the grading result is derived solely from the qualification data.

[0052] For incorrectly entered data, the system identifies abnormal fields by comparing them with the original records in the data source entity. After marking the data as abnormal, it triggers a manual review process. The reviewer verifies the original data, manually corrects the entered values, and recalculates. For incorrect data from tool calls, the system identifies erroneous outputs caused by abnormal parameter passing or interface call failures during Python programming tool calls. It automatically calls the corresponding calculation tool again and re-executes the calculation with the original input data, outputting the corrected intermediate calculation results.

[0053] D3: Differentiated scoring rules are adopted according to the level to which the assessment subject belongs. For the third-level unit level, the actual score of the risk dimension is calculated as the sum of the weight values ​​of all compliance indicators under the risk dimension. At the provincial level, the actual score of the risk dimension is calculated as the sum of the weight values ​​of each indicator under the risk dimension after being reduced according to the proportion of non-compliant units. The more non-compliant units there are, the greater the reduction.

[0054] Specifically, the risk assessment covers eight core safety risks: power grid risk, equipment risk, personal safety risk, environmental risk, occupational health risk, information network security risk, public safety risk, and other risks. The scoring method for each risk dimension is as follows.

[0055] The score for each risk dimension is the sum of the weights of all review indicators under that dimension, calculated as follows: The risk dimension score should be equal to Σ (the weight values ​​of all evaluation indicators under this dimension). Taking the public safety risk dimension as an example, if this dimension includes eight evaluation indicators: completion rate of rectification of major public safety hazards related to electricity (weight 0.75), occurrence of public safety incidents related to electricity for which the person with primary responsibility has occurred (weight 0.75), establishment of risk control for operations near pipelines (weight 0.15), establishment of external linkage for handling public safety emergencies (weight 0.75), establishment of public opinion monitoring and response management (weight 0.15), lack of clear safety responsibility interfaces with various social electricity users (weight 0.15), failure to urge rectification of public safety hazards related to electricity (weight 0.25), and failure to report major public safety hazards related to electricity to the local government department as required (weight 0.25), then the score for this dimension should be 0.75+0.75+0.15+0.75+0.15+0.15+0.25+0.25=3.3 points.

[0056] Example: The public safety risk dimension includes 8 assessment indicators. The weights and scores for each indicator are shown in Table 1. The score for this dimension should be 0.75+0.75+0.15+0.75+0.15+0.15+0.25+0.25=3.3 points.

[0057] Table 1. Examples of Public Safety Risk Weights

[0058] For the third-level unit, the actual score for the risk dimension is the sum of the weights of all compliance indicators under that dimension, calculated as follows: Actual score for risk dimension (three-level unit) = ∑ (weight values ​​of all compliance indicators under this dimension) At the provincial level, focusing on the overall risk control level across the province, a weighted calculation is needed, taking into account the indicator compliance status of each tertiary unit. The calculation formula is as follows: Actual score for risk dimension (provincial company) = ∑[indicator weight value × (1-n / m)] Where m is half the number of third-level units under the provincial company. When the number of third-level units is odd, it is rounded up. For example, if there are 11 units, then m=6. n is the number of third-level units that do not meet the standard. The value of n is in the range of {0,1,...,m}. The larger n is, the lower the compliance rate of the standard in the whole province, and the greater the impact on the deduction of the actual score at the provincial company level.

[0059] The formulas for calculating the score rate of each risk dimension are as follows: Risk dimension score rate = Actual score of risk dimension / Expected score of risk dimension × 100% A higher score indicates a higher level of safety management for that risk dimension. The warning triggering rules use the risk dimension score as the core evaluation indicator. When the risk dimension score is lower than the preset warning threshold, the system pushes a warning notification to the instant messaging tool of the responsible department corresponding to the non-compliance indicator for that risk dimension. The warning threshold can be adjusted as needed. The default threshold is set to 95%. For high-priority risk dimensions such as personal safety risks, the warning threshold can be increased to 98% to further improve the warning sensitivity, thereby achieving differentiated and precise control over different categories of safety risks.

[0060] In summary, in the indicator classification stage, the clear definition and detailed judgment rules of structured quantitative indicators and unstructured semantic parsing indicators enable the identification process of both types of indicators to be completed automatically based on data format attributes. Anomaly markers and manual verification paths are set for inconsistencies between type attributes and data source formats. In the knowledge network construction stage, the clear definition of attributes for five types of entities and the three-stage construction process enable the knowledge network to support agent decision-making while achieving automatic entity alignment through the edit distance algorithm. In the toolchain execution stage, the clear documentation of the BERT fine-tuning model's architecture hierarchy, training parameters, and input / output settings ensures the reproducibility of the unstructured document information extraction process and avoids incomplete technical solutions due to insufficient model construction details. In the hierarchical adjudication stage, the specific explanation of the logic tree's node composition and priority traversal termination rules provides a clear execution path for multi-condition adjudication logic. In the risk assessment stage, the introduction of differentiated scoring rules for three-level units and the provincial aggregation layer, along with configurable early warning thresholds, allows risk warning results to adapt to the differentiated control needs of different management levels.

[0061] Example 3 The experiment used a power supply bureau's 2024 safety and security system review as a scenario, selecting 100 items from 305 verification indicators as the experimental sample. The sample selection rule was random sampling, covering all 8 risk dimensions, including 80 data-related indicators and 20 semantic indicators. Experimental data includes: Structured data: Anfeng system MySQL database (containing 5000 records such as "total number of equipment defects, number of closed loops"); Unstructured data: Business guidelines, risk management guidelines, annual inspection reports, etc. (500 documents); Comparison methods: traditional manual calculation method (a team of 5 Anfeng system auditors) and single-tool automation method (using only RAG+Python, without knowledge graph and agent scheduling).

[0062] Evaluation indicators The experiment used four dimensions for evaluation: computational efficiency, accuracy, adaptability, and stability. Computational efficiency: Total time (minutes) to complete the calculation of 100 indicators for the drilling grade; Accuracy: The percentage of indicators that are consistent with the drill grade calculation results and the manual verification results (%). Adaptability: System adjustment time (in hours) when adding 10 new indicators (including 5 data-related and 5 semantic-related indicators).

[0063] Stability: The accuracy rate fluctuation range (%) is calculated for 100 sample indicators in three consecutive calculations. A fluctuation range of ≤5% in the accuracy rate of three consecutive calculations is considered to indicate good stability.

[0064] Experimental Results and Analysis Computational efficiency comparison The computation time for the three methods is shown in Table 2. As can be seen from the table, the total computation time of the proposed method is 42 minutes, which is 92.3% shorter than the traditional manual method (580 minutes) and 66.4% shorter than the single-tool method (125 minutes). This is because: the proposed method automatically schedules the toolchain through an agent, avoiding manual data extraction and tool switching; the pre-association of the knowledge graph reduces data source matching time; and the batch extraction of RAG improves the processing efficiency of semantic metrics.

[0065] Table 2 Comparison of calculation time for the three methods

[0066] Comparison of calculation accuracy The accuracy rates of the three methods are shown in Table 3. The accuracy rate of the method presented in this paper reaches 97.5%, which is higher than that of the traditional manual method (94.375%) and the single-tool method (93.75%). The improvement in accuracy for semantic indicators is particularly significant, due to: the knowledge graph qualification verification filtering out invalid data; the multi-level drilling judgment engine avoiding missed judgments based on multiple conditions; and the anomaly correction mechanism correcting 12 abnormal data entries (6 expired auditor qualification data, 4 data entries with input errors, and 2 data entries with incorrect calculation template calls), further improving accuracy. The accuracy rate of the single-tool method is lower, primarily because it lacks knowledge graph qualification verification, an anomaly correction mechanism, and intelligent scheduling of the toolchain, leading to larger errors in semantic indicator extraction and missed judgments based on multiple conditions. The accuracy rates of the method presented in this paper for three consecutive calculations were 97.5%, 100%, and 100%, respectively, demonstrating good stability.

[0067] Table 3 Comparison of accuracy rates of the three methods

[0068] Compatibility Comparison Table 4 shows the system adjustment time for the three methods when adding 10 new indicators. The adjustment time for the method presented in this paper is only 2 hours, far less than the traditional manual method (48 hours) and the single-tool method (15 hours). This is because: knowledge graphs support dynamic addition of indicators, the agent can automatically match the toolchain, and there is no need for manual reconstruction of the calculation process; the multi-level drill-down judgment engine can directly parse the drill-down criteria of the new indicators, without the need to redevelop the judgment logic. The single-tool method has a longer adjustment time, mainly because adding new indicators requires manual configuration of the tool call logic and redevelopment of judgment rules, resulting in poor adaptability.

[0069] Table 4 Comparison of the adaptability of the three methods

[0070] Experimental results show that the proposed method outperforms traditional manual methods and single-tool automated methods in four dimensions: computational efficiency, accuracy, adaptability, and stability. It can effectively solve the pain points of existing technologies and achieve automated, accurate, and efficient calculation of 305 indicators of the safety and safety system. At the same time, it can link risk management and meet the needs of refined review of the safety and safety system.

[0071] Example 4 The above is an illustrative scheme of an agent-driven automated classification and assessment method for multiple types of safety production review indicators. It should be noted that the technical solution of this agent-driven automated classification and assessment system for multiple types of safety production review indicators belongs to the same concept as the technical solution of the agent-driven automated classification and assessment method for multiple types of safety production review indicators described above. Details not described in detail in the technical solution of the agent-driven automated classification and assessment system for multiple types of safety production review indicators in this embodiment can be found in the description of the technical solution of the agent-driven automated classification and assessment method for multiple types of safety production review indicators described above.

[0072] This embodiment also provides an agent-driven automated grading and evaluation system for multiple types of safety production review indicators, including: an indicator classification module, used to classify review indicators into structured quantitative indicators and unstructured semantic parsing indicators based on the heterogeneity of data sources; The knowledge network construction module is used to build a multi-entity relational knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities. The Agent scheduling module is used to query the multi-entity association knowledge network and plan differentiated toolchains according to the indicator type to perform hierarchical calculations and obtain the initial hierarchical results. The grading adjudication module is used to adjudicate the initial grading results through a multi-condition priority judgment structure to obtain the indicator grading evaluation results. The risk warning module is used to calculate the score of each risk dimension based on the indicator classification assessment results, and to trigger a warning notification when the score rate is lower than the preset warning threshold.

[0073] This embodiment also provides an electronic device applicable to agent-driven automated classification and assessment of multiple types of safety production review indicators, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the agent-driven automated classification and assessment method for multiple types of safety production review indicators as proposed in the above embodiment.

[0074] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the Agent-driven automated classification and assessment method for implementing multiple types of safety production review indicators as proposed in the above embodiments.

[0075] The storage medium proposed in this embodiment and the Agent-driven automated classification and assessment method for implementing multiple types of safety production review indicators proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0076] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An agent-driven automated hierarchical assessment method for multiple types of safety production review indicators, characterized in that: This includes dividing the review indicators into structured quantitative indicators and unstructured semantic parsing indicators based on the heterogeneity of data sources, and constructing a multi-entity association knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities; The agent queries the multi-entity association knowledge network, plans differentiated toolchains according to the type of the evaluation indicators, and performs hierarchical calculations to obtain the initial hierarchical results. The initial grading results are adjudicated using a multi-condition priority judgment structure to obtain the indicator grading evaluation results. The scores for each risk dimension are calculated using the results of the index grading assessment as input, and an early warning notification is triggered when the risk dimension score rate is lower than the preset early warning threshold.

2. The Agent-Driven Automated Grading and Evaluation Method for Multiple Types of Safety Production Review Indicators as described in claim 1, characterized in that: The step of dividing the review indicators into structured quantitative indicators and unstructured semantic parsing indicators includes querying the data format attributes of the review indicators in the multi-entity association knowledge network, identifying indicators with structured data source format as structured quantitative indicators, and identifying indicators with unstructured data source format as unstructured semantic parsing indicators. When there is an inconsistency between the type attribute label of the review indicator and the format of the associated data source, the review indicator will be marked as an abnormal indicator, and a manual verification process will be initiated to confirm and adjust the indicator type.

3. The Agent-Driven Automated Grading and Assessment Method for Multiple Types of Safety Production Review Indicators as described in claim 2, characterized in that: The construction of the multi-entity association knowledge network includes extracting the name, classification judgment rules and associated risk dimension information of each review indicator from the review standard document; A combination of fuzzy name matching and manual verification is used to establish the relationships between indicator entities and data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities; The multi-entity association knowledge network is stored in a graph database. When a new review indicator is added, the corresponding entity and its association with the data source entity, the hierarchical standard entity, and the risk dimension entity are automatically added.

4. The agent-driven automated hierarchical assessment method for multiple types of safety production review indicators as described in claim 3, characterized in that: The method of planning differentiated toolchains based on the type of the review indicators includes, for structured quantitative indicators, planning a toolchain that coordinates structured data interface calls and programming calculation tools, extracting raw data from structured data sources and then performing numerical calculations to obtain intermediate numerical calculation results. For unstructured semantic parsing metrics, a toolchain is planned to collaborate with document information extraction tools and knowledge network qualification verification. After extracting effective information from unstructured data sources and filtering out invalid qualification data, semantic judgment is performed to obtain statistical intermediate results. The agent standardizes the intermediate computation results output by the toolchain and pushes them to the multi-condition priority determination structure to trigger hierarchical adjudication.

5. The agent-driven automated hierarchical assessment method for multiple types of safety production review indicators as described in claim 4, characterized in that: It also includes obtaining intermediate calculation results of the review indicators, validating the intermediate calculation results, and performing differentiated correction processing for different categories of abnormal data. Based on the hierarchical judgment rules of the review indicators, a hierarchical adjudication structure is constructed with basic compliance conditions as the root node and progressively improving conditions as sub-nodes. The hierarchical adjudication structure is traversed in priority order from the root node to the child nodes. When the condition of any node is met, the corresponding classification result is output and the subsequent traversal is terminated to obtain the index classification evaluation result. Using the rating results of the aforementioned indicators as input, differentiated scoring rules are applied according to the level to which the assessment subject belongs to calculate the expected score and actual score for each risk dimension. Calculate the score rate for each risk dimension, and trigger an early warning notification to the corresponding responsible department when the score rate is lower than the preset early warning threshold.

6. The agent-driven automated hierarchical assessment method for multiple types of safety production review indicators as described in claim 5, characterized in that: Differential correction processing is performed on different categories of abnormal data. This includes identifying data whose qualification validity period has exceeded the limit, removing the data from the valid dataset, and then re-triggering the classification decision; Identify and mark incorrectly entered data, and initiate a manual review process; The identification tool retrieves incorrect data and then calls the corresponding calculation tool again to correct the calculation results.

7. The agent-driven automated hierarchical assessment method for multiple types of safety production review indicators as described in claim 6, characterized in that: The differentiated scoring rules adopted according to the level to which the assessment subject belongs include, for the third-level unit level, calculating the actual score of the risk dimension as the sum of the weight values ​​of all compliance indicators under the risk dimension; At the provincial level, the actual score of the risk dimension is calculated as the sum of the weight values ​​of each indicator under the risk dimension after being reduced according to the proportion of non-compliant units. The more non-compliant units there are, the greater the reduction.

8. An agent-driven automated grading and assessment system for multiple types of safety production review indicators, based on the agent-driven automated grading and assessment method for multiple types of safety production review indicators as described in any one of claims 1 to 7, characterized in that: It also includes an indicator classification module, which is used to classify review indicators into structured quantitative indicators and unstructured semantic parsing indicators based on the heterogeneity of data sources. The knowledge network construction module is used to build a multi-entity relational knowledge network covering indicator entities, data source entities, hierarchical standard entities, calculation tool entities, and risk dimension entities. The Agent scheduling module is used to query the multi-entity association knowledge network and plan differentiated toolchains according to the indicator types to perform hierarchical calculations and obtain initial hierarchical results. The grading adjudication module is used to adjudicate the initial grading results through a multi-condition priority judgment structure to obtain the indicator grading evaluation results. The risk warning module is used to calculate the score of each risk dimension based on the indicator classification assessment results, and to trigger a warning notification when the score rate is lower than the preset warning threshold.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the Agent-driven automated classification and assessment method for multiple types of safety production review indicators as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the Agent-driven automated classification and assessment method for multiple types of safety production review indicators as described in any one of claims 1 to 7.

Images