An AI fusion management method and system for a multi-dimensional safety production knowledge base, a storage medium and a program product

Abstract

The application provides an AI fusion management method and system of a multi-dimensional safety production knowledge base, a storage medium and a program product, relates to the technical field of safety production management, and comprises the following steps: in the case that multi-source heterogeneous safety production source data is acquired, a multi-level safety knowledge graph and a bidirectional index mapping channel are constructed by using the safety production source data, the multi-dimensional safety production knowledge base comprises the multi-level safety knowledge graph and the bidirectional index mapping channel; in response to a context awareness request triggered by a terminal, real-time running parameters of a target device and behavior intention characteristics of a user are collected; the real-time running parameters and the behavior intention characteristics are mapped by using the bidirectional index mapping channel, so as to lock a target risk source node in the multi-level safety knowledge graph; the target risk source node is subjected to multi-hop graph traversal deduction according to the topological association characteristics of the multi-level safety knowledge graph, and a target safety protection strategy is generated; and the target safety protection strategy is used to perform a contextualized knowledge pushing operation on the terminal.

Description

Technical Field

[0001] This application relates to the field of safety production management technology, and in particular to an AI-integrated management method, system, storage medium, and program product for a multi-dimensional safety production knowledge base. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, the scale of data and the complexity of business operations in industrial environments are increasing dramatically, leading to a growing demand from enterprises for real-time monitoring and refined prevention of safety production management. This is particularly true in large-scale, complex manufacturing projects, which place higher demands on the dynamic analysis of safety production risks and the integrated application of multi-dimensional knowledge.

[0003] In related technologies, centralized relational databases or distributed document management systems are commonly used for safety production knowledge management. The specific implementation process is as follows: First, source documents such as safety production regulations, operation manuals, and historical accident cases are collected using manual input or simple data collection tools. Then, the identification information from the collected source documents is extracted and stored in the centralized relational database according to a preset data table structure, or the source documents are directly stored in the distributed document management system. Next, when managers or on-site workers need to obtain safety production knowledge, they input query keywords into the centralized relational database or the distributed document management system. The centralized relational database or the distributed document management system uses a single-line text matching algorithm to perform character comparison and retrieval of the query keywords within the preset data table structure or file names. Finally, a list of regulations or individual documents containing matching characters is output as a text collection for the managers or on-site workers to consult and refer to.

[0004] However, with the above approach, since safety production knowledge management is usually based on a preset data table structure or independent source files to generate static text collection output, and the files themselves mainly carry normative clauses and some attribute identification information during the retrieval process, when it is necessary to deal with the massive amount of frequently updated heterogeneous regulations or quickly identify potential complex hazards caused by multiple factors, it is often necessary to rely on manual re-sorting of the deep causal relationships between files and passively initiate new precise queries, which leads to low risk assessment and response efficiency in the multi-dimensional management of safety production knowledge in related technologies. Summary of the Invention

[0005] This application provides an AI-integrated management method, system, storage medium, and program product for a multi-dimensional safety production knowledge base, which can be used to improve the risk assessment and response efficiency of multi-dimensional safety production knowledge management.

[0006] Firstly, this application provides an AI-integrated management method for a multi-dimensional safety production knowledge base, applied to the aforementioned AI-integrated management system. The method includes: upon acquiring multi-source heterogeneous safety production source data, constructing a multi-level safety knowledge graph and a bidirectional index mapping channel using the safety production source data; the multi-dimensional safety production knowledge base includes the multi-level safety knowledge graph and the bidirectional index mapping channel; responding to a context-aware request triggered by a terminal, collecting real-time operating parameters of the target device and user behavioral intent features in the current scenario; using the bidirectional index mapping channel to perform graph mapping on the real-time operating parameters and behavioral intent features to locate the target risk source node in the multi-level safety knowledge graph; performing multi-hop graph traversal and deduction on the target risk source node based on the topological association features of the multi-level safety knowledge graph to generate a target safety protection strategy; and using the target safety protection strategy to perform contextualized knowledge push operations to the terminal.

[0007] By adopting the above technical solution, after acquiring multi-source heterogeneous safety production source data, a multi-dimensional safety production knowledge base, including a multi-level safety knowledge graph and a bidirectional index mapping channel, is constructed using the safety production source data. This allows safety production knowledge to be organized and stored in a graph-based topological structure, and enables efficient linkage between the graph database space and the attribute database through the bidirectional index mapping channel. When a terminal triggers a context-aware request, the bidirectional index mapping channel is used to map real-time operating parameters and behavioral intent features to the graph, quickly locating the target risk source node. This effectively avoids the inefficient mode of relying on manual sorting of deep causal relationships between files in traditional methods. Furthermore, based on the topological association characteristics of the multi-level safety knowledge graph, multi-hop graph traversal and deduction are performed on the target risk source node, automatically generating the target security protection strategy and actively pushing it to the terminal, realizing the transformation from passive query to active push. This solves the technical problem of low risk assessment and response efficiency in multi-dimensional management of safety production knowledge in related technologies, achieving the technical effect of improving the risk assessment and response efficiency of multi-dimensional management of safety production knowledge.

[0008] Optionally, when multi-source heterogeneous safety production source data is obtained, a multi-level safety knowledge graph and a bidirectional index mapping channel are constructed using the safety production source data. This includes: extracting entities from the safety production source data to obtain an entity set, which consists of multiple safety entity nodes; predicting relationships in the safety production source data to obtain a set of association relationships, which represents the causal mapping relationship between multiple safety entity nodes; constructing an initial safety knowledge graph within the graph database space based on a preset three-level ontology structure, entity set, and association relationship set; configuring a corresponding initial risk transmission coefficient for each association edge in the initial safety knowledge graph based on the statistical distribution characteristics of historical accident cases to obtain a multi-level safety knowledge graph; and expanding the attributes of each safety entity node in the entity set using an attribute database associated with the graph database space to establish a bidirectional index mapping channel between the graph database space and the attribute database.

[0009] By adopting the above technical solutions, entity extraction and relationship prediction are performed on safety production source data, transforming unstructured safety production knowledge into a structured set of entities and relationships. An initial safety knowledge graph is constructed within the graph database space based on a pre-defined three-level ontology structure, organizing safety production knowledge according to the hierarchical relationships of accident type, causal factors, and protective measures. By configuring an initial risk transmission coefficient for each relationship edge, quantifiable representation of causal relationships can be achieved, supporting subsequent risk extrapolation calculations. Furthermore, by establishing a bidirectional index mapping channel between the graph database space and the attribute database, collaborative management of relationship data and attribute data is achieved, providing an efficient data access path for subsequent graph mapping and knowledge retrieval.

[0010] Optionally, a bidirectional index mapping channel is used to perform graph mapping on real-time running parameters and behavioral intent features to lock target risk source nodes in a multi-level security knowledge graph. This includes: using a hybrid feature extraction engine to extract features from real-time running parameters to obtain running feature vectors; and using a hybrid feature extraction engine to perform semantic parsing on behavioral intent features to obtain a set of scene keywords; performing full-text matching on the attribute database based on the set of scene keywords to obtain initial matching records, and using a bidirectional index mapping channel to map the initial matching records to the multi-level security knowledge graph to obtain a first candidate node set; and calculating the text similarity score between each first candidate node included in the first candidate node set and the set of scene keywords; using a bidirectional index mapping channel to obtain the attribute representation vector corresponding to each security entity node in the multi-level security knowledge graph from the attribute database, and calculating the feature space distance between the running feature vector and the attribute representation vector; merging security entity nodes whose feature space distance is less than or equal to a preset distance threshold into a second candidate node set, and determining the running fit score of each second candidate node included in the second candidate node set based on the feature space distance; and determining the target risk source node based on the text similarity score and the running fit score.

[0011] By adopting the above technical solution, a hybrid feature extraction engine is used to process real-time running parameters and behavioral intent features respectively, resulting in running feature vectors and scene keyword sets, achieving unified representation of multimodal features. Full-text matching is performed in the attribute database and mapped to a multi-level security knowledge graph using a bidirectional index mapping channel, resulting in a first candidate node set based on text similarity. Simultaneously, by calculating the feature space distance between the running feature vector and the attribute representation vector, a second candidate node set based on running fit is obtained. Finally, the target risk source node is determined by comprehensively considering both text similarity score and running fit score, achieving dual localization through semantic and numerical matching, and improving the accuracy of risk source node locking.

[0012] Optionally, the target risk source node is determined based on the text similarity score and the runtime fit score, including: performing entity type matching on the set of scene keywords to determine whether the behavioral intent features include specific equipment entities and / or process entities; if the behavioral intent features include specific equipment entities and process entities, the first fusion weight is determined to be less than the second fusion weight; if the behavioral intent features only include specific equipment entities or process entities, the first fusion weight is determined to be equal to the second fusion weight; if the behavioral intent features do not include specific equipment entities or process entities, the first fusion weight is determined to be greater than the second fusion weight; performing a union operation on the first candidate node set and the second candidate node set to obtain the target candidate node set; using the first fusion weight and the second fusion weight to perform a weighted fusion calculation on the text similarity score and runtime fit score of each target candidate node included in the target candidate node set to obtain the comprehensive risk assessment score of each target candidate node; sorting each target candidate node in descending order according to the comprehensive risk assessment score, and locking the target candidate node at the top of the sorted list as the target risk source node.

[0013] By employing the aforementioned technical solution, entity type matching is performed on the set of scenario keywords to determine whether the behavioral intent features include specific equipment entities and process flow entities. Based on this, the relative magnitudes of the first and second fusion weights are dynamically adjusted. When the behavioral intent features include specific equipment entities and process flow entities, it indicates that the user's query intent is clear. In this case, the weight of the text similarity score is reduced, and the weight of the operational fit score is increased to focus more on matching the equipment's operational status. When the behavioral intent features do not include specific equipment entities and process flow entities, it indicates that the user's query intent is relatively vague. In this case, the weight of the text similarity score is increased to focus more on semantic-level matching. Through this adaptive weight adjustment mechanism, the determination of target risk source nodes can adapt to different query scenarios, improving the flexibility and accuracy of risk source location.

[0014] Optionally, based on the topological association characteristics of the multi-level security knowledge graph, a multi-hop graph traversal deduction is performed on the target risk source node to generate a target security protection strategy. This includes: using the target risk source node as the starting node for traversal, activating a preset depth-first traversal algorithm in the multi-level security knowledge graph; using the depth-first traversal algorithm to perform a step-by-step traversal along the connection direction of the association edges in the multi-level security knowledge graph, so that when the traversal depth reaches a preset depth threshold, multiple risk deduction paths are generated from the target risk source node to the final protection measure node; obtaining the initial risk transmission coefficient of the target association edges traversed by each risk deduction path in the multiple risk deduction paths, and obtaining the path depth value of each risk deduction path; determining the depth weight factor of each risk deduction path based on the path depth value, and using the depth weight factor to apply the strategy to the target risk source node. The initial risk transmission coefficients of the target association edges are weighted to obtain the weighted risk transmission coefficients of the target association edges. The weighted risk transmission coefficients of the target association edges traversed by each risk deduction path are accumulated to obtain the cumulative transmission risk value of each risk deduction path. The cumulative transmission risk values ​​of each risk deduction path are sorted in descending order, and the risk deduction path corresponding to the first cumulative transmission risk value after sorting is determined as the target risk deduction path. The final-level protection measure node corresponding to the target risk deduction path is taken as the target protection node. The standard protection operation procedure text and safety production standard content bound to the target protection node are retrieved from the attribute database using a bidirectional index mapping channel, and the standard protection operation procedure text and safety production standard content are combined to generate the target safety protection strategy.

[0015] By adopting the above technical solution, the target risk source node is used as the starting node for traversal. A depth-first traversal algorithm is used to perform a step-by-step traversal along the connection direction of the associated edges, generating multiple risk deduction paths from the target risk source node to the final protection measure node. By obtaining the initial risk transmission coefficient and path depth value of the target associated edges traversed by each risk deduction path, a depth weighting factor is determined based on the path depth value, and the initial risk transmission coefficient is weighted. Then, the weighted risk transmission coefficients are accumulated to obtain the cumulative transmission risk value. This ensures that the longer the path and the more causal chains involved, the greater the cumulative transmission risk value, which can effectively identify complex hidden dangers caused by multiple factors. By sorting the cumulative transmission risk values ​​in descending order and selecting the risk deduction path corresponding to the first one as the target risk deduction path, the path with the highest risk transmission degree is prioritized. Finally, the standard protection operation procedure text and safety production standard content bound to the target protection node are retrieved from the attribute database and combined to generate the target safety protection strategy, realizing an automated closed loop from risk identification to protection strategy generation.

[0016] Optionally, the target security protection strategy is used to perform contextualized knowledge push operations to the terminal, including: determining whether the context-aware request includes preset keywords for specific work scenarios; if the context-aware request includes keywords for specific work scenarios, the target security protection strategy is encapsulated into a mandatory security push instruction, and the standard protection operation procedure text and safety production standard content carried in the mandatory security push instruction are pushed to the target interactive interface of the terminal through a front-end microservice gateway that establishes a communication connection with a multi-dimensional safety production knowledge base; if the context-aware request does not include keywords for specific work scenarios, the user's job qualification information and historical operation behavior records are collected, and the job qualification information and historical operation behavior records are converted into user profile representation vectors; the knowledge attribute vector corresponding to the target security protection strategy is obtained; the user profile representation vector and the knowledge attribute vector are input into the context-aware recommendation engine associated with the multi-dimensional safety production knowledge base; the dual-tower recommendation model built into the context-aware recommendation engine is called to calculate the cosine similarity between the user profile representation vector and the knowledge attribute vector; if the cosine similarity reaches a preset relevance threshold, the target security protection strategy is encapsulated into a safety production knowledge recommendation instruction, and the safety production knowledge recommendation instruction is pushed to the target interactive interface of the terminal through the front-end microservice gateway.

[0017] By adopting the above technical solution, the system determines whether the context-aware request includes preset keywords for specific work scenarios. When such keywords are included, the target security protection strategy is encapsulated as a mandatory security push instruction and directly pushed to the terminal via the front-end microservice gateway, ensuring that security information in high-risk work scenarios can be delivered to relevant personnel in a timely manner. When the specific work scenario keywords are not included, the system collects the user's job qualification information and historical operation records and converts them into user profile representation vectors to achieve vectorized representation of user characteristics. It then obtains the knowledge attribute vectors corresponding to the target security protection strategy and inputs both the user profile representation vector and the knowledge attribute vectors into the context-aware recommendation engine. A dual-tower recommendation model is used to calculate the cosine similarity between the two, achieving accurate matching between users and knowledge. When the cosine similarity reaches a preset relevance threshold, the target security protection strategy is encapsulated as a safety production knowledge recommendation instruction and pushed to the terminal via the front-end microservice gateway, realizing personalized knowledge recommendation based on user profiles. This solution balances the mandatory requirements of safety production with the personalized needs of knowledge push, ensuring timely delivery of security information in high-risk scenarios while avoiding information overload in general scenarios, thus improving the targeting and timeliness of knowledge push.

[0018] Optionally, after constructing a multi-level safety knowledge graph and a bidirectional index mapping channel using safety production source data, the method further includes: monitoring the version update status of the safety production standard content in the safety production source data according to a preset time period, and extracting the updated safety production standard content from the safety production source data when the version update status is detected as updated; locating target safety entity nodes in the multi-level safety knowledge graph that have a historical version correspondence with the updated safety production standard content using the bidirectional index mapping channel; generating new safety entity nodes corresponding to the updated safety production standard content in the multi-level safety knowledge graph, and adding the updated safety production standard content to the attribute database using the bidirectional index mapping channel to establish new safety entities. The system establishes a bidirectional index mapping relationship between nodes and updated safety production standard content. It then uses this mapping to obtain the original text content of the updated safety production standard content bound to the newly added safety entity node, performs vectorized feature extraction on the original text content to generate a target attribute representation vector corresponding to the newly added safety entity node, and updates the target attribute representation vector to the attribute database. A version evolution relationship edge is established between the target safety entity node and the newly added safety entity node. A weighted attenuation calculation is performed on the initial risk transmission coefficient configured for each relationship edge connecting the target safety entity node using a preset time attenuation factor to obtain the target risk transmission coefficient. Finally, the initial risk transmission coefficient configured for each relationship edge is replaced with the target risk transmission coefficient.

[0019] By adopting the above technical solution, the system automatically detects updates to regulations and standards by monitoring the version update status of the safety production source data corresponding to the safety production standard content according to a preset time cycle. When a version update is detected, a new safety entity node corresponding to the updated safety production standard content is generated in the multi-level safety knowledge graph, and version evolution relationship edges are established with historical version nodes to achieve traceable management of knowledge versions. By using a preset time decay factor to perform weighted decay calculation on the initial risk transmission coefficient configured for each relationship edge connecting the target safety entity node, the risk transmission impact of the old version knowledge node gradually decreases over time, ensuring the timeliness of the knowledge base and the accuracy of risk assessment, and realizing the dynamic self-updating capability of the multi-dimensional safety production knowledge base.

[0020] Secondly, embodiments of this application provide an AI fusion management system, which includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the AI ​​fusion management system to perform the methods described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on an AI fusion management system, cause the AI ​​fusion management system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on an AI fusion management system, cause the AI ​​fusion management system to perform the method described in the first aspect and any possible implementation thereof. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an AI-integrated management method for a multi-dimensional safety production knowledge base in this application embodiment;

[0024] Figure 2 This is a schematic diagram of the physical device structure of an AI fusion management system in the embodiments of this application. Detailed Implementation

[0025] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0026] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0027] This application provides an AI-integrated management method for a multi-dimensional safety production knowledge base. (See attached document.) Figure 1 , Figure 1 This is a flowchart illustrating an AI-integrated management method for a multi-dimensional safety production knowledge base, as described in this application, including the following steps:

[0028] Step S101: After obtaining multi-source heterogeneous safety production source data, construct a multi-level safety knowledge graph and a two-way index mapping channel using the safety production source data. The multi-dimensional safety production knowledge base includes a multi-level safety knowledge graph and a two-way index mapping channel.

[0029] Step S102: In response to the context-aware request triggered by the terminal, collect the real-time operating parameters of the target device and the user's behavioral intent characteristics in the current scenario;

[0030] Step S103: Use the bidirectional index mapping channel to perform graph mapping on real-time running parameters and behavioral intent features in order to lock the target risk source node in the multi-level security knowledge graph;

[0031] Step S104: Based on the topological association characteristics of the multi-level security knowledge graph, perform multi-hop graph traversal and deduction on the target risk source nodes to generate the target security protection strategy.

[0032] Step S105: Use the target security protection strategy to perform contextualized knowledge push operation to the terminal.

[0033] Among them, multi-source heterogeneous safety production source data refers to safety production-related data from different data sources and with different data formats, including but not limited to regulatory and standard data from safety production supervision platforms, production operation data from enterprise manufacturing execution systems, and safety production research literature data from academic platforms; multi-level safety knowledge graph refers to a knowledge graph with a hierarchical ontology structure built within a graph database space (i.e., Neo4j graph database), in which safety entity nodes are organized according to a preset three-level ontology structure (i.e., a three-level hierarchical structure of accident type-causative factors-protective measures); bidirectional index mapping channel refers to a bidirectional data synchronization channel established between the graph database space and the attribute database (i.e., MySQL relational database) (i.e., a bidirectional index mapping channel based on Logstash). The system includes a synchronization middleware (used to implement bidirectional index association between security entity nodes in the graph database space and attribute records in the attribute database); a context-aware request refers to a knowledge query request triggered by an end user in a specific work scenario, carrying real-time operating parameters of the target device and user behavioral intent characteristics in the current scenario; the target device refers to production equipment or facilities in the current scenario that require safety monitoring and risk assessment, including but not limited to mechanical equipment, chemical plants, and electrical equipment on industrial production lines; the target risk source node refers to a security entity node that is locked in the multi-level safety knowledge graph and is highly related to the risk of the current scenario; and the contextualized knowledge push operation refers to the operation of proactively pushing target security protection strategies to the terminal in a personalized manner based on user profiles and scenario characteristics.

[0034] In the above embodiments, the AI ​​fusion management system first acquires multi-source heterogeneous safety production source data through a knowledge acquisition module. This knowledge acquisition module includes a standardized application programming interface (API) submodule. This submodule synchronizes structured and unstructured data from the safety production supervision platform and the enterprise manufacturing execution system via an authorized data interface. The standardized API submodule supports expressive state transition protocols and simple object access protocols, and is equipped with flow control valves and data verification units. After acquiring the safety production source data, the AI ​​fusion management system processes the data using a knowledge processing module. This knowledge processing module includes a preprocessing unit, a natural language processing analysis unit, and a knowledge graph construction unit. The preprocessing unit is equipped with a data cleaning pipeline, including a hypertext markup language tag stripper, a special character filter, and a duplicate data detector. The natural language processing analysis unit integrates a pre-trained language model, and the knowledge graph construction unit adopts a top-down modeling approach. After processing by the knowledge processing module, the AI ​​fusion management system uses a knowledge storage module to construct a multi-level... The security knowledge graph and bidirectional index mapping channel are used in the knowledge storage module. The module employs a dual-database architecture: a graph database stores spatial relationships, and an attribute database stores attribute data. Both databases maintain data consistency through bidirectional synchronization middleware. When a terminal triggers a context-aware request, the AI ​​fusion management system responds by collecting real-time operating parameters of the target device and user behavioral intent features in the current scenario. Real-time operating parameters include the target device's operating status parameters and environmental monitoring parameters, while behavioral intent features include the user's input query text and operational context. The AI ​​fusion management system uses the bidirectional index mapping channel to perform graph mapping on the real-time operating parameters and behavioral intent features, identifying target risk source nodes within the multi-level security knowledge graph. Then, based on the topological association features of the multi-level security knowledge graph, it performs multi-hop graph traversal and deduction of the target risk source nodes to generate target security protection strategies. Finally, the knowledge application module performs contextualized knowledge push operations to the terminal. This knowledge application module includes a hybrid retrieval engine and a context-aware recommendation engine, providing a unified access point through a front-end microservice gateway.

[0035] Through the above steps, after acquiring multi-source heterogeneous safety production source data, a multi-dimensional safety production knowledge base, including a multi-level safety knowledge graph and a bidirectional index mapping channel, is constructed using this data. This allows safety production knowledge to be organized and stored in a graph-based topological structure, and enables efficient linkage between the graph database space and the attribute database through the bidirectional index mapping channel. When a terminal triggers a context-aware request, the bidirectional index mapping channel maps real-time operating parameters and behavioral intent features to the graph, quickly locating the target risk source node. This effectively avoids the inefficient traditional method of relying on manual analysis of deep causal relationships between files. Furthermore, based on the topological association characteristics of the multi-level safety knowledge graph, multi-hop graph traversal and deduction are performed on the target risk source node, automatically generating the target security protection strategy and proactively pushing it to the terminal, realizing a shift from passive querying to proactive pushing. This solves the technical problem of low risk assessment and response efficiency in multi-dimensional management of safety production knowledge in related technologies, achieving the technical effect of improving the risk assessment and response efficiency of multi-dimensional management of safety production knowledge.

[0036] The entity performing the above steps may be a system or device, or a controller or processor in the device or system, or a separate controller or processor, or other processing devices or processing units with similar processing functions, but is not limited to these.

[0037] In an optional embodiment, when multi-source heterogeneous safety production source data is obtained, a multi-level safety knowledge graph and a bidirectional index mapping channel are constructed using the safety production source data. This includes: extracting entities from the safety production source data to obtain an entity set, which consists of multiple safety entity nodes; predicting relationships in the safety production source data to obtain a set of association relationships, which represents the causal mapping relationships between multiple safety entity nodes; constructing an initial safety knowledge graph within the graph database space based on a preset three-level ontology structure, entity set, and association relationship set; configuring a corresponding initial risk transmission coefficient for each association edge in the initial safety knowledge graph based on the statistical distribution characteristics of historical accident cases to obtain a multi-level safety knowledge graph; and expanding the attributes of each safety entity node in the entity set using an attribute database associated with the graph database space to establish a bidirectional index mapping channel between the graph database space and the attribute database.

[0038] Among them, entity extraction refers to using natural language processing analysis units to identify and extract professional entities in the safety field from safety production source data, such as identifying professional terms in the safety production field such as explosion-proof motors, benzene leaks, and electrostatic sparks; relation prediction refers to using a classifier (i.e., the Softmax classifier) ​​to determine the probability of causal relationships between entities, thereby determining the causal mapping relationship between multiple safety entity nodes; graph database space refers to a graph database environment (i.e., the Neo4j graph database) used to store knowledge graphs, supporting the storage of entities and their relationships in the form of nodes and edges; the pre-defined three-level ontology structure refers to a three-level hierarchical ontology structure defined according to accident type-causative factors-protective measures, used to organize the hierarchical relationship of safety production knowledge; initial wind The risk transmission coefficient refers to the weight value assigned to each relational edge in a multi-level safety knowledge graph. This weight value is calculated based on the statistical distribution characteristics of historical accident cases and is used to quantify the degree of risk transmission in causal relationships. The attribute database refers to a relational database (i.e., a MySQL database) associated with the graph database space, used to store attribute data of safety entity nodes. The bidirectional index mapping channel refers to a bidirectional data synchronization channel (i.e., a bidirectional synchronization middleware based on Logstash) established between the graph database space and the attribute database. Through this channel, attribute records in the attribute database can be queried based on node identifiers in the graph database space, and corresponding nodes in the graph database space can be located based on attribute records in the attribute database space.

[0039] In the above embodiments, the specific process of entity extraction from safety production source data by the AI ​​fusion management system is as follows: The natural language processing analysis unit integrates a pre-trained language model (i.e., the BERT-Base Chinese model). This pre-trained language model includes a custom safety domain entity recognition layer with an entity recognition accuracy greater than or equal to 92%. The natural language processing analysis unit inputs the safety production source data into the pre-trained language model. The pre-trained language model performs sequence labeling on the safety production source data, identifying professional entities in the safety domain, such as identifying professional terms like benzene leakage, explosion-proof motor, and confined space operation. All identified professional entities in the safety domain are summarized into an entity set, where each professional entity in the entity set corresponds to a safety entity node in a multi-level safety knowledge graph. The AI ​​fusion management system then performs related analysis on the safety production source data. The specific process of prediction is as follows: The natural language processing analysis unit uses the Softmax classifier to make a probabilistic judgment on the causal relationship between any two safe entity nodes in the entity set. When the probability of the causal relationship is greater than the preset probability threshold, it is determined that there is a causal mapping relationship between the two safe entity nodes. All entity pairs with causal mapping relationships are summarized to form an association set. The specific process of the AI ​​fusion management system constructing the initial safety knowledge graph in the graph database space is as follows: The knowledge graph construction unit adopts a top-down modeling approach. First, it defines a three-level ontology structure of accident type-causative factors-protective measures. Then, it classifies each safe entity node in the entity set according to the three-level ontology structure. Then, it establishes association edges between the corresponding safe entity nodes according to the association set and automatically generates resource description framework triples (the format can be: <accident, The AI ​​fusion management system configures the initial risk transmission coefficient for each relation edge in the initial safety knowledge graph by: statistically analyzing the frequency of occurrence and severity of various causal relationships from historical accident cases, and calculating the initial risk transmission coefficient for each relation edge based on the statistical distribution characteristics. The initial risk transmission coefficient ranges from 0 to 1, with a larger value indicating a higher degree of risk transmission for the causal relationship. The AI ​​fusion management system establishes a bidirectional index mapping channel by: using the attribute database to extend the attributes of each safety entity node in the entity set, decomposing the attribute data (including entity name, entity type, entity description, associated standard text, etc.) of each safety entity node according to the normalization and storing it in the attribute database, and establishing a bidirectional index association between the safety entity nodes in the graph database space and the attribute records in the attribute database, and maintaining data consistency between the graph database space and the attribute database through a bidirectional synchronization middleware (i.e., a bidirectional synchronization middleware based on Logstash), thereby establishing a bidirectional index mapping channel.

[0040] In an optional embodiment, a bidirectional index mapping channel is used to perform graph mapping on real-time running parameters and behavioral intent features to locate target risk source nodes in a multi-level security knowledge graph. This includes: using a hybrid feature extraction engine to extract features from real-time running parameters to obtain running feature vectors; using the hybrid feature extraction engine to perform semantic parsing on behavioral intent features to obtain a set of scene keywords; performing full-text matching on the attribute database based on the set of scene keywords to obtain initial matching records, and using the bidirectional index mapping channel to map the initial matching records to the multi-level security knowledge graph to obtain a first candidate node set; and calculating the text similarity score between each first candidate node included in the first candidate node set and the set of scene keywords; using the bidirectional index mapping channel to obtain the attribute representation vector corresponding to each security entity node in the multi-level security knowledge graph from the attribute database, and calculating the feature space distance between the running feature vector and the attribute representation vector; merging security entity nodes whose feature space distance is less than or equal to a preset distance threshold into a second candidate node set, and determining the running fit score of each second candidate node included in the second candidate node set based on the feature space distance; and determining the target risk source node based on the text similarity score and the running fit score.

[0041] Among them, the hybrid feature extraction engine refers to a feature processing engine that integrates numerical feature extraction and semantic parsing capabilities, enabling it to extract numerical features from real-time operating parameters and perform semantic parsing of behavioral intent features; the operating feature vector refers to the numerical vector obtained after feature extraction from real-time operating parameters, used to represent the device's operating status in the current scenario; the scenario keyword set refers to the set of keywords extracted after semantic parsing of behavioral intent features, used to represent the user's query intent; full-text matching refers to the operation of full-text retrieval and matching of the scenario keyword set in the attribute database; the initial matching record refers to the attribute database record returned by the full-text matching operation; and the first candidate node set refers to the set of nodes after the initial matching records are mapped to a multi-level security knowledge graph through a bidirectional index mapping channel. The obtained candidate safe entity node set; the text similarity score refers to the text similarity score between each first candidate node in the first candidate node set and the scene keyword set; the attribute representation vector refers to the attribute feature vector stored in the attribute database corresponding to each safe entity node; the feature space distance refers to the distance metric between the running feature vector and the attribute representation vector in the feature space, which can be calculated using Euclidean distance or cosine distance; the preset distance threshold refers to the pre-set feature space distance judgment threshold; the second candidate node set refers to the candidate node set composed of safe entity nodes whose feature space distance is less than or equal to the preset distance threshold; the running fit score refers to the score calculated based on the feature space distance, which represents the degree of fit between the safe entity node and the current running state.

[0042] In the above embodiments, the specific process of the AI ​​fusion management system using the hybrid feature extraction engine to extract features from real-time running parameters is as follows: the hybrid feature extraction engine performs one-hot encoding and term frequency-inverse document frequency weighting on the numerical parameters in the real-time running parameters, and concatenates the processed features to form a running feature vector; the specific process of the AI ​​fusion management system using the hybrid feature extraction engine to perform semantic parsing on behavioral intent features is as follows: the hybrid feature extraction engine uses a word segmentation component (i.e., the HanLP word segmentation component) to segment the text content in the behavioral intent features, and extracts the keywords to form a scene keyword set; the specific process of the AI ​​fusion management system performing full-text matching in the attribute database based on the scene keyword set is as follows: the AI ​​fusion management system uses a full-text search engine (i.e., the Elasticsearch full-text search engine) to perform Boolean search (supporting AND, OR, NOT combinations) on the scene keyword set in the attribute database to obtain initial matching records; the specific process of the AI ​​fusion management system using a bidirectional index mapping channel to map the initial matching records to a multi-level security knowledge graph is as follows: the AI ​​fusion management system locates the corresponding security entity nodes in the graph database space through the bidirectional index mapping channel based on the node identifiers in the initial matching records, and summarizes the located security entity nodes to form the first A candidate node set; the specific process of the AI ​​fusion management system calculating text similarity scores is as follows: the AI ​​fusion management system calculates the text similarity between the attribute text of each first candidate node in the first candidate node set and the scene keyword set. The text similarity can be calculated using cosine similarity weighted by word frequency-inverse document frequency; the specific process of the AI ​​fusion management system calculating feature space distance is as follows: the AI ​​fusion management system obtains the attribute representation vector corresponding to each security entity node in the multi-level security knowledge graph from the attribute database through a bidirectional index mapping channel, and calculates the Euclidean distance between the running feature vector and each attribute representation vector as the feature space distance; A The specific process by which the AI ​​fusion management system determines the second candidate node set and the running fit score is as follows: The AI ​​fusion management system merges safe entity nodes whose feature space distance is less than or equal to a preset distance threshold into the second candidate node set. The preset distance threshold can be determined based on the statistical distribution of historical data, for example, taking the median of the Euclidean distance between all attribute representation vectors as the preset distance threshold; The AI ​​fusion management system calculates the running fit score based on the feature space distance. The running fit score is negatively correlated with the feature space distance. The calculation formula is: Running fit score = 1 / (1 + feature space distance). The smaller the feature space distance, the higher the running fit score.

[0043] In an optional embodiment, determining the target risk source node based on text similarity score and runtime fit score includes: performing entity type matching on the set of scene keywords to determine whether the behavioral intent feature includes specific equipment entities and / or process entities; if the behavioral intent feature includes specific equipment entities and process entities, determining that the first fusion weight is less than the second fusion weight; if the behavioral intent feature only includes specific equipment entities or process entities, determining that the first fusion weight is equal to the second fusion weight; if the behavioral intent feature does not include specific equipment entities and process entities, determining that the first fusion weight is greater than the second fusion weight; performing a union operation on the first candidate node set and the second candidate node set to obtain a target candidate node set; using the first fusion weight and the second fusion weight to perform a weighted fusion calculation on the text similarity score and runtime fit score of each target candidate node included in the target candidate node set to obtain a comprehensive risk assessment score for each target candidate node; sorting each target candidate node in descending order according to the comprehensive risk assessment score, and locking the target candidate node at the top of the sorted list as the target risk source node.

[0044] Among them, entity type matching refers to the operation of identifying the entity type of keywords in the set of scene keywords, which is used to determine whether the keywords belong to specific equipment entities or process entities; specific equipment entities refer to entities that represent specific equipment, such as explosion-proof motors, pressure vessels, safety valves, etc.; process entities refer to entities that represent specific process flows, such as confined space operations, hot work operations, high-altitude operations, etc.; the first fusion weight refers to the weight coefficient used to weight the text similarity score; the second fusion weight refers to the weight coefficient used to weight the operation fit score; the target candidate node set refers to the candidate node set obtained by performing a union processing of the first candidate node set and the second candidate node set; the comprehensive risk assessment score refers to the comprehensive score obtained by weighting and fusing the text similarity score and the operation fit score using the first fusion weight and the second fusion weight.

[0045] In the above embodiments, the specific process of the AI ​​fusion management system performing entity type matching on the scene keyword set is as follows: The AI ​​fusion management system uses a natural language processing analysis unit to identify the entity type of each keyword in the scene keyword set, determining whether each keyword belongs to a specific equipment entity or a process entity; the specific process of the AI ​​fusion management system determining the first fusion weight and the second fusion weight based on the entity type matching result is as follows: When the behavioral intent features include specific equipment entities and process entities, it indicates that the user's query intent is clear and involves specific equipment and processes. At this time, the matching of equipment operating status is more important, so the first fusion weight is determined to be less than the second fusion weight, for example, the first fusion weight is 0.3 and the second fusion weight is 0.7; when the behavioral intent features only include specific equipment entities or process entities, it indicates that the user's query intent is partially clear. At this time, text matching and operating status matching are equally important, so the first fusion weight is determined to be equal to the second fusion weight, for example, the first fusion weight and the second fusion weight are equal. The weights are all 0.5. When the behavioral intent features do not include specific equipment entities and process entities, it indicates that the user's query intent is relatively vague. At this time, semantic text matching is more important. Therefore, the first fusion weight is determined to be greater than the second fusion weight. For example, the first fusion weight is 0.7 and the second fusion weight is 0.3. The specific process of the AI ​​fusion management system to calculate the comprehensive risk assessment score is as follows: The AI ​​fusion management system performs a union operation on the first candidate node set and the second candidate node set to obtain the target candidate node set. For each target candidate node in the target candidate node set, the weighted fusion calculation is performed using the formula: Comprehensive Risk Assessment Score = First Fusion Weight × Text Similarity Score + Second Fusion Weight × Operation Fit Score. The specific process of the AI ​​fusion management system to determine the target risk source node is as follows: The AI ​​fusion management system sorts each target candidate node in the target candidate node set in descending order according to the comprehensive risk assessment score, and locks the target candidate node at the top of the sorted list as the target risk source node.

[0046] In an optional embodiment, a multi-hop graph traversal deduction is performed on the target risk source node based on the topological association characteristics of the multi-level security knowledge graph to generate a target security protection strategy. This includes: activating a preset depth-first traversal algorithm in the multi-level security knowledge graph, using the target risk source node as the starting node; performing a step-by-step traversal along the connection direction of the association edges in the multi-level security knowledge graph using the depth-first traversal algorithm, so that when the traversal depth reaches a preset depth threshold, multiple risk deduction paths are generated from the target risk source node to the final protection measure node; obtaining the initial risk transmission coefficient of the target association edges traversed by each risk deduction path in the multiple risk deduction paths, and obtaining the path depth value of each risk deduction path; determining the depth weight factor of each risk deduction path based on the path depth value, and using the depth weight factor... The initial risk transmission coefficients of the target association edges are weighted to obtain the weighted risk transmission coefficients of the target association edges. The weighted risk transmission coefficients of the target association edges traversed by each risk deduction path are accumulated to obtain the cumulative transmission risk value of each risk deduction path. The cumulative transmission risk values ​​of each risk deduction path are sorted in descending order, and the risk deduction path corresponding to the first cumulative transmission risk value after sorting is determined as the target risk deduction path. The final-level protection measure node corresponding to the target risk deduction path is taken as the target protection node. The standard protection operation procedure text and safety production standard content bound to the target protection node are retrieved from the attribute database using a bidirectional index mapping channel, and the standard protection operation procedure text and safety production standard content are combined to generate the target safety protection strategy.

[0047] In this context, the starting node refers to the initial position of the multi-hop graph traversal deduction, i.e., the target risk source node; the depth-first traversal algorithm is a graph traversal algorithm that performs a depth-first, step-by-step traversal along the connection direction of the association edges starting from the starting node; the preset depth threshold is the maximum traversal depth limit of the depth-first traversal algorithm, used to control the traversal range, for example, a preset depth threshold of 4; the final protection measure node refers to the security entity node belonging to the protection measure level in the three-level ontology structure, and is the endpoint of the risk deduction path; the risk deduction path refers to the path starting from the target risk source node, passing through several intermediate nodes, and finally reaching the final protection measure node; the target association edge refers to the association edge traversed by the risk deduction path; the cumulative transmitted risk value refers to the cumulative calculation result of the initial risk transmission coefficient of all target association edges on the risk deduction path; and the path depth... The degree value refers to the number of associated edges traversed by the risk simulation path from the starting node to the final protective measure node; the depth weight factor refers to the weight coefficient determined based on the path depth value, used to weight the initial risk transmission coefficient, and the depth weight factor is positively correlated with the path depth value; the weighted risk transmission coefficient refers to the risk transmission coefficient obtained after weighting the initial risk transmission coefficient using the depth weight factor; the target risk simulation path refers to the risk simulation path with the highest cumulative transmitted risk value; the target protective node refers to the final protective measure node corresponding to the target risk simulation path; the standard protective operation procedure text refers to the standard operation procedure document bound to the target protective node; the safety production standard content refers to the relevant standard clauses bound to the target protective node, including but not limited to national standards or industry standards, such as occupational health-related standard clauses.

[0048] In the above embodiments, the specific process of the AI ​​fusion management system activating the depth-first traversal algorithm with the target risk source node as the starting node is as follows: The AI ​​fusion management system activates the preset depth-first traversal algorithm in the graph database space with the target risk source node as the starting node; the specific process of the AI ​​fusion management system executing the step-by-step traversal to generate risk deduction paths is as follows: The depth-first traversal algorithm starts from the starting node and performs step-by-step traversal along the connection direction of the relational edges in the multi-level security knowledge graph. The traversal depth is increased by 1 for each level traversed. When the traversal depth reaches a preset depth threshold (e.g., the preset depth threshold is 4), the traversal stops, and the path from the target risk source node to the current traversal endpoint is recorded as the risk deduction path. Among them, the path whose traversal endpoint is the final protection measure node is the effective risk deduction path, and finally multiple risk deduction paths from the target risk source node to the final protection measure node are generated; the specific process of the AI ​​fusion management system calculating the cumulative transmitted risk value is as follows: For each risk deduction path, the AI ​​fusion management system obtains the initial values ​​of all target relational edges traversed by the risk deduction path. The system obtains the risk transmission coefficient and the path depth value of the risk projection path. Based on the path depth value, the AI ​​fusion management system determines the depth weight factor corresponding to the risk projection path. The depth weight factor is positively correlated with the path depth value; for example, depth weight factor = 1 + 0.2 × path depth value. The AI ​​fusion management system uses the depth weight factor to weight the initial risk transmission coefficients of each target relationship edge traversed by the risk projection path, obtaining the weighted risk transmission coefficient of the target relationship edge. The calculation formula is: Weighted Risk Transmission Coefficient_i = Initial Risk Transmission Coefficient_i × Depth Weight Factor. The AI ​​fusion management system accumulates the weighted risk transmission coefficients of the target relationship edges traversed by the risk projection path to obtain the cumulative transmission risk value of the risk projection path. The calculation formula is: Cumulative Transmission Risk Value = Σ(Weighted Risk Transmission Coefficient_i), where i is the sequence number of the target relationship edge traversed by the risk projection path. Through this calculation method, the longer the path and the more causal chains involved, the greater the cumulative transmission risk value, effectively identifying complex hidden dangers caused by the superposition of multiple factors.

[0049] The specific process by which the AI ​​fusion management system determines the target risk projection path and target protection node is as follows: The AI ​​fusion management system sorts the cumulative transmission risk values ​​of all risk projection paths in descending order, and determines the risk projection path corresponding to the first cumulative transmission risk value after sorting as the target risk projection path. This target risk projection path represents the path with the highest degree of risk transmission and needs to be protected first. The last-level protection measure node corresponding to the target risk projection path is taken as the target protection node. The specific process by which the AI ​​fusion management system generates the target safety protection strategy is as follows: The AI ​​fusion management system uses a two-way index mapping channel to retrieve the standard protection operation procedure text and safety production standard content bound to the target protection node from the attribute database, and combines the standard protection operation procedure text and safety production standard content to generate the target safety protection strategy.

[0050] In an optional embodiment, a contextualized knowledge push operation is performed on the terminal using the target security protection strategy. Specifically, this includes: determining whether the context-aware request includes preset keywords for a specific work scenario; if the context-aware request includes keywords for a specific work scenario, encapsulating the target security protection strategy into a mandatory security push instruction, and pushing the standard protection operation procedure text and safety production standard content carried in the mandatory security push instruction to the target interactive interface of the terminal through a front-end microservice gateway that establishes a communication connection with a multi-dimensional safety production knowledge base; if the context-aware request does not include keywords for a specific work scenario, collecting the user's job qualification information and historical operation behavior records, and converting the job qualification information and historical operation behavior records into a user profile representation vector; obtaining the knowledge attribute vector corresponding to the target security protection strategy; inputting the user profile representation vector and the knowledge attribute vector into a context-aware recommendation engine associated with the multi-dimensional safety production knowledge base; calling the dual-tower recommendation model built into the context-aware recommendation engine to calculate the cosine similarity between the user profile representation vector and the knowledge attribute vector; if the cosine similarity reaches a preset relevance threshold, encapsulating the target security protection strategy into a safety production knowledge recommendation instruction, and pushing the safety production knowledge recommendation instruction to the target interactive interface of the terminal through the front-end microservice gateway.

[0051] Among them, specific work scenario keywords refer to predefined keywords that characterize specific high-risk work scenarios, such as confined space operations, hot work, and work at height. When the context-aware request includes specific work scenario keywords, it indicates that the current situation is a high-risk work scenario, and a forced push needs to be executed to ensure timely delivery of safety information. The forced safety push instruction refers to the instruction data packet formed by encapsulating the target safety protection strategy when specific work scenario keywords are detected, which needs to be forcibly pushed to the terminal. Job qualification information refers to information such as the user's job type, qualification certificates, and training records. Historical operation behavior records refer to the user's historical query records, browsing records, operation records, and other behavioral data in the AI ​​fusion management system. User profile representation vector refers to the user feature vector obtained by vectorizing the job qualification information and historical operation behavior records. Knowledge attribute vector refers to the attribute feature vector corresponding to the target safety protection strategy, including the vectorized table of attributes such as knowledge type, applicable scenario, and associated equipment. The context-aware recommendation engine refers to an engine component used to implement personalized knowledge recommendations based on user profiles. This context-aware recommendation engine adopts a user-knowledge dual-tower model architecture. The dual-tower recommendation model is a deep learning recommendation model architecture, including two independent neural network towers: a user tower and a knowledge tower, which are used to encode user features and knowledge features, respectively. Cosine similarity refers to the cosine similarity between the user profile representation vector and the knowledge attribute vector, used to measure the degree of matching between the user and the knowledge. The preset relevance threshold refers to a pre-set cosine similarity judgment threshold, used to determine whether the degree of matching between the user and the knowledge meets the recommendation conditions. The safety production knowledge recommendation instruction refers to the instruction data package formed by encapsulating the target safety protection strategy for personalized recommendations when no specific work scenario keywords are detected and the cosine similarity reaches the preset relevance threshold. The front-end microservice gateway refers to a gateway component (i.e., Spring Cloud Gateway microservice gateway) that establishes a communication connection with the multi-dimensional safety production knowledge base and provides a unified access point to the outside world. The target interaction interface refers to the user interaction interface on the terminal used to display the content of the forced safety push instruction or the safety production knowledge recommendation instruction.

[0052] In the above embodiments, the specific process of the AI ​​fusion management system performing contextualized knowledge push operations is as follows: The AI ​​fusion management system first determines whether the context-aware request includes preset keywords for specific work scenarios (such as confined space operations, hot work, high-altitude operations, etc.). This determination process is achieved by keyword matching of the text content in the context-aware request. The preset keywords for specific work scenarios are stored in the system configuration database. When it is determined that the context-aware request includes keywords for specific work scenarios, the AI ​​fusion management system encapsulates the target security protection strategy into a forced security push instruction, which is then pushed through the front-end microservice gateway (i.e., Spring Cloud). The Gateway microservice gateway directly pushes the standard protection operation procedure text and safety production standard content carried in the forced safety push command to the target interactive interface of the terminal. This forced push process does not rely on user profile matching results, ensuring that safety information in high-risk operation scenarios can be delivered to relevant personnel in a timely manner. When it is determined that the context-aware request does not include keywords for a specific operation scenario, the AI ​​fusion management system executes a personalized knowledge recommendation process. Specifically, the AI ​​fusion management system obtains the user's job qualification information from the user management database, including job type, qualification certificate type, training records, etc., and at the same time obtains the user's historical operation behavior records from the behavior log database, including historical query keywords, browsed knowledge items, operation time distribution, etc. The AI ​​fusion management system uses the context-aware recommendation engine to vectorize the job qualification information and historical operation behavior records to generate user profile representation vectors. The context-aware recommendation engine performs vectorization encoding on the job qualification information and historical operation behavior records every 30 seconds based on the real-time stream computing framework (i.e., Spark). The AI ​​fusion management system updates the user profile representation vector using Streaming; it retrieves knowledge attribute data corresponding to the target security protection strategy from the attribute database, including attributes such as knowledge type, applicable scenarios, associated devices, and risk level, and vectorizes the knowledge attribute data to generate knowledge attribute vectors; the AI ​​fusion management system inputs the user profile representation vector into the user tower of the context-aware recommendation engine for feature encoding, and inputs the knowledge attribute vector into the knowledge tower of the context-aware recommendation engine for feature encoding. The context-aware recommendation engine uses the online model inference service (i.e., TensorFlow Serving) to calculate the cosine similarity between the user tower output vector and the knowledge tower output vector in real time. The formula for calculating the cosine similarity is: Cosine similarity = (User tower output vector · Knowledge tower output vector) / (|User tower output vector| × |Knowledge tower output vector|); when the cosine similarity reaches a preset relevance threshold (e.g., the preset relevance threshold is 0.75), the AI ​​fusion management system encapsulates the target security protection strategy into a safety production knowledge recommendation instruction, and pushes the standard protection operation procedure text and safety production standard content carried in the safety production knowledge recommendation instruction to the target interactive interface of the terminal through the front-end microservice gateway, realizing personalized knowledge recommendation based on user profile.

[0053] In an optional embodiment, after constructing a multi-level safety knowledge graph and a bidirectional index mapping channel using safety production source data, the method further includes: monitoring the version update status of the safety production standard content in the safety production source data according to a preset time period, and extracting the updated safety production standard content from the safety production source data when the version update status is detected as updated; locating target safety entity nodes in the multi-level safety knowledge graph that have a historical version correspondence with the updated safety production standard content using the bidirectional index mapping channel; generating new safety entity nodes corresponding to the updated safety production standard content in the multi-level safety knowledge graph, and adding the updated safety production standard content to the attribute database using the bidirectional index mapping channel to establish a new... A bidirectional index mapping relationship is established between safety entity nodes and updated safety production standard content. This bidirectional index mapping relationship is used to obtain the original text content of the updated safety production standard content bound to the newly added safety entity node. Vectorized feature extraction is then performed on the original text content to generate a target attribute representation vector corresponding to the newly added safety entity node, and this target attribute representation vector is updated in the attribute database. Version evolution relationship edges are established between the target safety entity node and the newly added safety entity node. A weighted attenuation calculation is performed on the initial risk transmission coefficients configured for each relationship edge connecting the target safety entity node using a preset time attenuation factor to obtain the target risk transmission coefficient. Finally, the initial risk transmission coefficients configured for each relationship edge are replaced with the target risk transmission coefficient.

[0054] Among them, the preset time period refers to the pre-set version update monitoring cycle, such as performing version update monitoring once per quarter; the version update status refers to the status indicator of whether the version of the safety production standard content has been updated; the updated version of the safety production standard content refers to the updated version of the safety production standard content, such as the clauses of the latest national or industry standard; the historical version correspondence refers to the correspondence between the updated version of the safety production standard content and its historical versions; the target safety entity node refers to the safety entity node in the multi-level safety knowledge graph that has a historical version correspondence with the updated version of the safety production standard content, that is, the node corresponding to the old version of the standard; and the newly added safety entity node refers to the newly generated safety entity node in the multi-level safety knowledge graph that corresponds to the updated version of the safety production standard content. The bidirectional index mapping relationship refers to the bidirectional index association established between the newly added safety entity node and the updated safety production standard content; the original text content refers to the original text data of the updated safety production standard content; the target attribute representation vector refers to the attribute representation vector generated after performing vectorized feature extraction on the original text content; the version evolution relationship edge refers to the association relationship edge established between the target safety entity node and the newly added safety entity node, representing the version evolution relationship; the preset time decay factor refers to the pre-set decay factor used to calculate the time decay of the initial risk transmission coefficient, and the value of the decay factor is between 0 and 1; the target risk transmission coefficient refers to the risk transmission coefficient obtained after weighted decay calculation of the initial risk transmission coefficient using the preset time decay factor.

[0055] In the above embodiments, the specific process by which the AI ​​fusion management system monitors the version update status and extracts the updated safety production standard content according to a preset time period is as follows: The AI ​​fusion management system monitors the version update status of the safety production source data corresponding to the safety production standard content through the knowledge acquisition module according to a preset time period (e.g., quarterly). The standardized application programming interface submodule of the knowledge acquisition module polls the data source for changes every 5 minutes. When the version update status is detected as updated, the AI ​​fusion management system extracts the updated safety production standard content (e.g., referring to the latest published national standard) from the safety production source data. The AI ​​fusion management system uses a bidirectional index mapping channel to locate the target safety entity. The specific process for node generation is as follows: The AI ​​fusion management system, based on the standard number or name of the updated safety production standard content, uses a bidirectional index mapping channel to query attribute records with historical version correspondences in the attribute database. Then, it maps the queried attribute records to a multi-level safety knowledge graph through the bidirectional index mapping channel, locating the target safety entity node that corresponds to a historical version of the updated safety production standard content. The specific process for the AI ​​fusion management system to generate new safety entity nodes and establish bidirectional index mapping relationships is as follows: The AI ​​fusion management system generates new safety entity nodes corresponding to the updated safety production standard content in the graph database space, while simultaneously utilizing the bidirectional index mapping channel... The updated safety production standard content is added to the attribute database, and a bidirectional index mapping relationship is established between the new safety entity node and the updated safety production standard content. The specific process of the AI ​​fusion management system generating the target attribute representation vector is as follows: The AI ​​fusion management system uses the bidirectional index mapping relationship to obtain the original text content of the updated safety production standard content bound to the new safety entity node, uses the natural language processing analysis unit to perform vectorized feature extraction on the original text content, generates the target attribute representation vector corresponding to the new safety entity node, and updates the target attribute representation vector to the attribute database. The specific process of the AI ​​fusion management system establishing version evolution relationship edges is as follows: AI fusion management... In the graph database space, the system establishes version evolution relationship edges between target security entity nodes and newly added security entity nodes. The relationship type of these version evolution relationship edges is version evolution, used to represent the evolutionary relationship between old and new version standards. The specific process of the AI ​​fusion management system calculating and replacing the target risk transmission coefficient is as follows: The AI ​​fusion management system uses a preset time decay factor to perform weighted attenuation calculation on the initial risk transmission coefficient configured for each association relationship edge connecting the target security entity node. The calculation formula is: Target risk transmission coefficient = Initial risk transmission coefficient × Preset time decay factor, where the preset time decay factor ranges from 0 to 1 (e.g., the preset time decay factor is 0).8) When the model evaluation metric (e.g., F1 score) drops below a preset threshold (e.g., 5%), the model needs to be retrained and the decay factor updated. The AI ​​fusion management system replaces the initial risk transmission coefficient configured for each relational edge with the target risk transmission coefficient, ensuring that the risk transmission impact of the old version of the knowledge node gradually decreases over time, thus guaranteeing the timeliness of the knowledge base.

[0056] Through the embodiments of this application, the AI ​​fusion management system transforms scattered safety production knowledge into a graph-based structure with causal relationships by constructing a multi-level safety knowledge graph and a two-way index mapping channel. When the terminal triggers a context-aware request, the two-way index mapping channel is used to quickly locate the target risk source node, and the target safety protection strategy is automatically generated through multi-hop graph traversal and deduction, realizing the transformation from passive query to active push, thereby improving the risk assessment and response efficiency of multi-dimensional management of safety production knowledge.

[0057] The AI ​​fusion management system in the embodiments of this invention is described below from the perspective of hardware processing. (See attached document.) Figure 2 , Figure 2 This is a schematic diagram of the physical device structure of an AI fusion management system in the embodiments of this application.

[0058] It should be noted that, Figure 2 The structure of the AI ​​fusion management system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0059] like Figure 2 As shown, the AI ​​fusion management system includes a central processing unit (CPU) 201, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 202 or a program loaded from storage portion 208 into random access memory (RAM) 203, such as executing the methods described in the above embodiments. The RAM 203 also stores...

[0060] The system contains various programs and data required for operation. CPU 201, ROM 202, and RAM 203 are interconnected via bus 204. Input / output (I / O) interface 205 is also connected to bus 204. The following components are connected to I / O interface 205: input section 206, including audio input devices, push-button switches, etc.; output section 207, including a liquid crystal display (LCD), audio output devices, indicator lights, etc.; storage section 208, including a hard disk, etc.; and communication section 209, including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 209 performs communication processing via a network such as the Internet. Drive 210 is also connected to I / O interface 205 as needed. Removable media 211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 210 as needed so that computer programs read from them can be installed into storage section 208 as needed.

[0061] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 209, and / or installed from removable medium 211. When the computer program is executed by central processing unit (CPU) 201, it performs the various functions defined in the present invention.

[0062] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0063] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0064] Specifically, the AI ​​fusion management system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the AI ​​fusion management method for the multi-dimensional safety production knowledge base provided in the above embodiment.

[0065] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the AI ​​fusion management system described in the above embodiments; or it may exist independently and not incorporated into the AI ​​fusion management system. The storage medium carries one or more computer programs, which, when executed by a processor of the AI ​​fusion management system, enable the AI ​​fusion management system to implement the AI ​​fusion management method for the multi-dimensional safety production knowledge base provided in the above embodiments.

[0066] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0067] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. An AI-integrated management method for a multi-dimensional safety production knowledge base, characterized in that, include: When multi-source heterogeneous safety production source data is obtained, a multi-level safety knowledge graph and a two-way index mapping channel are constructed using the safety production source data. The multi-dimensional safety production knowledge base includes the multi-level safety knowledge graph and the two-way index mapping channel. In response to context-aware requests triggered by the terminal, collect real-time operating parameters of the target device and user behavioral intent characteristics in the current scenario; The bidirectional index mapping channel is used to perform graph mapping on the real-time running parameters and the behavioral intent features in order to locate the target risk source node in the multi-level security knowledge graph; Based on the topological association features of the multi-level security knowledge graph, a multi-hop graph traversal deduction is performed on the target risk source nodes to generate a target security protection strategy; The target security protection strategy is used to perform contextualized knowledge push operation to the terminal.

2. The method according to claim 1, characterized in that, The step of acquiring multi-source heterogeneous safety production source data and constructing a multi-level safety knowledge graph and bidirectional index mapping channel using the safety production source data includes: The safety production source data is subjected to entity extraction to obtain an entity set, which consists of multiple safety entity nodes. Relationship prediction is performed on the safety production source data to obtain a set of association relationships, which represents the causal mapping relationship between the multiple safety entity nodes; Within the graph database space, an initial security knowledge graph is constructed based on a preset three-level ontology structure, the entity set, and the set of relationships. Based on the statistical distribution characteristics of historical accident cases, a corresponding initial risk transmission coefficient is configured for each relational edge in the initial safety knowledge graph to obtain a multi-level safety knowledge graph; The attributes of each secure entity node in the entity set are extended using the attribute database associated with the graph database space to establish a bidirectional index mapping channel between the graph database space and the attribute database.

3. The method according to claim 2, characterized in that, The step of using the bidirectional index mapping channel to perform graph mapping on the real-time operating parameters and the behavioral intent features to locate the target risk source node in the multi-level security knowledge graph includes: The real-time running parameters are extracted using a hybrid feature extraction engine to obtain a running feature vector, and the behavioral intent features are semantically parsed using the hybrid feature extraction engine to obtain a set of scene keywords. Based on the set of scene keywords, full text matching is performed in the attribute database to obtain initial matching records. The initial matching records are then mapped to the multi-level security knowledge graph using the bidirectional index mapping channel to obtain a first candidate node set. The text similarity score between each first candidate node in the first candidate node set and the set of scene keywords is calculated. The attribute representation vector corresponding to each security entity node in the multi-level security knowledge graph is obtained from the attribute database using the bidirectional index mapping channel, and the feature space distance between the running feature vector and the attribute representation vector is calculated. Safe entity nodes whose feature space distance is less than or equal to a preset distance threshold are grouped into a second candidate node set, and the running fit score of each second candidate node included in the second candidate node set is determined according to the feature space distance. The target risk source node is determined based on the text similarity score and the runtime fit score.

4. The method according to claim 3, characterized in that, The step of determining the target risk source node based on the text similarity score and the runtime fit score includes: Perform entity type matching on the set of scenario keywords to determine whether the behavioral intent features include specific equipment entities and / or process entities; If it is determined that the behavioral intent features include the specific equipment entity and the process entity, then the first fusion weight is determined to be less than the second fusion weight. If it is determined that the behavioral intent feature only includes the specific equipment entity or the process entity, the first fusion weight is determined to be equal to the second fusion weight. If it is determined that the behavioral intent features do not include the specific equipment entity and the process entity, the first fusion weight is determined to be greater than the second fusion weight. Perform a union operation on the first candidate node set and the second candidate node set to obtain the target candidate node set; The text similarity score and running fit score of each target candidate node included in the target candidate node set are weighted and fused using the first fusion weight and the second fusion weight to obtain the comprehensive risk assessment score of each target candidate node. Each target candidate node is sorted in descending order according to the comprehensive risk assessment score, and the target candidate node at the top of the sorted list is locked as the target risk source node.

5. The method according to claim 2, characterized in that, The step of performing multi-hop graph traversal deduction on the target risk source nodes based on the topological association features of the multi-level security knowledge graph to generate a target security protection strategy includes: Using the target risk source node as the starting node for traversal, a preset depth-first traversal algorithm is activated in the multi-level security knowledge graph; The depth-first traversal algorithm is used to perform a step-by-step traversal along the connection direction of the association edges in the multi-level security knowledge graph, so that when the traversal depth reaches a preset depth threshold, multiple risk deduction paths are generated from the target risk source node to the final protection measure node. Obtain the initial risk transmission coefficient of the target association edge traversed by each of the multiple risk simulation paths, and obtain the path depth value of each risk simulation path; The depth weighting factor of each risk simulation path is determined based on the path depth value, and the initial risk transmission coefficient of the target association edge is weighted using the depth weighting factor to obtain the weighted risk transmission coefficient of the target association edge. The weighted risk transmission coefficients of the target association edges traversed by each risk deduction path are summed to obtain the cumulative transmission risk value of each risk deduction path. The cumulative transmission risk values ​​of each risk projection path are sorted in descending order, and the risk projection path corresponding to the first cumulative transmission risk value after sorting is determined as the target risk projection path, and the final protection measure node corresponding to the target risk projection path is taken as the target protection node. The standard protection operation procedure text and safety production standard content bound to the target protection node are retrieved from the attribute database using the bidirectional index mapping channel, and the standard protection operation procedure text and the safety production standard content are combined to generate the target safety protection strategy.

6. The method according to claim 5, characterized in that, The step of performing contextualized knowledge push operation on the terminal using the target security protection strategy includes: Determine whether the context-aware request includes preset keywords for a specific work scenario; If the context-aware request includes the specific work scenario keywords, the target security protection strategy is encapsulated as a forced security push instruction. The standard protection operation program text and the safety production standard content carried in the forced security push instruction are pushed to the target interactive interface of the terminal through the front-end microservice gateway that establishes a communication connection with the multi-dimensional safety production knowledge base. If the context-aware request does not include the specific job scenario keyword, the user's job qualification information and historical operation behavior records are collected, and the job qualification information and historical operation behavior records are converted into user profile representation vectors. Obtain the knowledge attribute vector corresponding to the target security protection strategy; The user profile representation vector and the knowledge attribute vector are input into the context-aware recommendation engine associated with the multi-dimensional safety production knowledge base; The context-aware recommendation engine's built-in dual-tower recommendation model is invoked to calculate the cosine similarity between the user profile representation vector and the knowledge attribute vector; When the cosine similarity reaches a preset relevance threshold, the target security protection strategy is encapsulated as a safety production knowledge recommendation instruction, and the safety production knowledge recommendation instruction is pushed to the target interactive interface of the terminal through the front-end microservice gateway.

7. The method according to claim 6, characterized in that, After constructing a multi-level safety knowledge graph and a bidirectional index mapping channel using the safety production source data, the method further includes: Monitor the version update status of the safety production source data corresponding to the safety production standard content according to a preset time period, and extract the updated version of the safety production standard content from the safety production source data when the version update status is detected as updated. The bidirectional index mapping channel is used to locate target security entity nodes in the multi-level security knowledge graph that have a historical version correspondence with the updated version of the safety production standard content; In the multi-level safety knowledge graph, new safety entity nodes corresponding to the updated safety production standard content are generated, and the updated safety production standard content is added to the attribute database using the bidirectional index mapping channel, thereby establishing a bidirectional index mapping relationship between the new safety entity nodes and the updated safety production standard content. The original text content of the updated safety production standard bound to the newly added safety entity node is obtained by using the bidirectional index mapping relationship, and vectorized feature extraction is performed on the original text content to generate the target attribute representation vector corresponding to the newly added safety entity node. The target attribute representation vector is then updated to the attribute database. Establish a version evolution relationship edge between the target security entity node and the newly added security entity node; The target risk transmission coefficient is obtained by weighting and attenuating the initial risk transmission coefficient configured for each association edge connecting the target security entity node using a preset time decay factor. Replace the initial risk transmission coefficient configured for each of the associated edges with the target risk transmission coefficient.

8. An AI fusion management system, characterized in that, The AI ​​fusion management system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the AI ​​fusion management system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the AI ​​fusion management system, the AI ​​fusion management system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the AI ​​fusion management system, the AI ​​fusion management system performs the method as described in any one of claims 1-7.

Images