Accident emergency decision-making method of knowledge graph enhanced large model in chemical field

Abstract

The invention discloses an accident emergency decision-making method of a chemical industry field knowledge graph enhanced large model. The method comprises the following steps: collecting historical accident case data, scene data and emergency field knowledge data of a chemical industry park in real time; constructing a chemical industrial park accident'scene-response 'domain ontology by using a seven-step method according to the collected chemical industrial park historical accident case data, scene data and emergency domain knowledge data; based on a chemical industry park accident scene-coping domain ontology, constructing a knowledge graph by adopting a graph database processing method, a natural language processing method and a large model extraction triple method; the reasoning ability of a large model is enhanced through entity relation information in a knowledge graph, and a reliable auxiliary emergency decision is quickly generated according to accident scene description. According to the invention, the large model is enhanced through the chemical industrial park emergency field knowledge graph, the accident auxiliary emergency decision is generated, and the accident emergency response efficiency of the chemical industrial park is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital emergency management in chemical parks, and specifically relates to an accident emergency decision-making method based on a large model enhanced by a knowledge graph in the chemical industry. Background Art

[0002] Chemical parks have increasingly dense distribution of chemical plants and complex production processes involving a large number of hazardous chemicals. Once an accident occurs, it may produce a cascade effect, leading to domino effect accidents such as fire, explosion, and toxic gas leakage, causing serious consequences such as large number of casualties, property losses, etc.

[0003] For complex accident scenarios, efficient and reliable emergency decision-making is crucial; the lack of key accident emergency scenario information will lead to wrong emergency decisions, thereby exacerbating the evolution of the accident scenario; the current emergency decision-making field includes two emergency decision-making models: "prediction-response" and "scenario-response" models; the "prediction-response" model relies on the emergency management department's prediction and preparation before the accident occurs, and speculates on the potential impact of the accident based on experience, and formulates corresponding emergency plans; however, due to the subjective cognitive limitations of the emergency plan preparers and the unpredictability and complexity of the accident scenario development, the actual accident scenario is quite different from the scenario preset in the emergency plan, which is difficult to meet the actual emergency response requirements. The "scenario-response" model considers various complex accident scenarios derived from the occurrence to the termination of an accident by identifying important emergency scenario information characteristics such as various monitoring data, the nature and location of hazardous chemicals, the type of accident and its development trend. On this basis, emergency decision-makers can use chemical park emergency plans, hazardous chemical disposal plans, etc. as emergency guidance materials, combine expert experience to make effective emergency decisions and prepare in advance, which can effectively reduce the uncertainty in the accident emergency response process. However, due to the huge amount of information elements involved in chemical park accident scenarios, relying solely on expert experience and emergency guidance documents may lead to a decrease in the reliability of emergency response decisions.

[0004] Therefore, a method for generating auxiliary emergency decision-making for chemical park accidents based on a knowledge graph-enhanced large model is currently provided, which can solve the technical problem that traditional preset emergency plans and expert experience are unable to quickly formulate reliable emergency decisions based on complex and real-time evolving accident scenarios.

[0005] The existing chemical accident emergency decision-making method based on knowledge graph (CN 116681305 A: A knowledge graph-based emergency decision-making method for emergencies) can construct an emergency knowledge graph based on historical data of emergencies and emergency decisions, but lacks the integration of emergency domain knowledge documents and cannot fully consider the comprehensive requirements of emergency decision-making. The above technology has the limitations of insufficient dynamic adaptability, inability to integrate real-time data streams, and lack of comprehensive and reliable domain knowledge support for decision generation. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for generating auxiliary emergency decision-making for chemical park accidents based on a knowledge graph enhanced large model. By constructing a comprehensive chemical emergency field knowledge graph, covering chemical park accident scenarios, historical accident cases and emergency response field knowledge, it can quickly generate reliable auxiliary emergency decision-making methods for complex and real-time evolving accident scenarios in chemical parks, thereby improving the efficiency, flexibility and reliability of emergency decision-making for chemical park accident scenarios.

[0007] The present invention is achieved through at least one of the following technical solutions.

[0008] A method for emergency decision-making based on a knowledge graph-enhanced large model in the chemical industry, comprising the following steps:

[0009] Real-time collection of historical accident case data, scenario data, and emergency knowledge data in chemical parks;

[0010] The seven-step method is used to construct the chemical park accident scenario-response domain ontology based on the collected historical accident case data, scenario data and emergency domain knowledge data of the chemical park;

[0011] Based on the chemical park accident scenario-response domain ontology, a knowledge graph for chemical emergency response is constructed using graph database processing, natural language processing, and large-scale model triple extraction methods.

[0012] The entity relationship information in the knowledge graph of chemical emergency is used to enhance the reasoning ability of the large model, and reliable auxiliary emergency decisions are quickly generated based on the accident scenario description.

[0013] Furthermore, the historical accident case data of the chemical park is obtained through the accident investigation report texts of domestic and foreign chemical parks; the scenario data is obtained with the help of various monitoring systems deployed within the chemical park; and the emergency field knowledge data is obtained from the emergency field knowledge texts of current regulations and standards.

[0014] Furthermore, the seven-step method for constructing a chemical park accident scenario-response domain ontology includes the following steps:

[0015] Step 1: Determine the domain of ontology construction. The chemical accident scenario emergency response decision-making domain is selected as the target domain of ontology construction, focusing on the emergency decision-making needs of chemical park accidents.

[0016] Step 2: Reuse the existing disaster ontology and emergency plan ontology;

[0017] Step 3: Determine the professional terms that describe the chemical park scenario and emergency response decision-making domain knowledge;

[0018] Step 4: Define categories and subcategories: Based on disaster system theory, the categories related to accident scenarios and emergency response decision-making domain knowledge are divided into four categories: disaster-prone environment, hazard-causing factors, disaster-bearing bodies, and emergency response decisions, and further subcategorized;

[0019] Step 5: Define attributes and relationships. Attributes include object attributes and data attributes.

[0020] Step 6: Define the domain, range, and data type of attributes and relationships, and determine the value range and data type of each attribute and relationship;

[0021] Step 7: Supplement instance data: Based on the categories, attributes, and relationships, combined with actual chemical park accident cases and emergency response data, fill in specific instance data for each category and its subcategories, attributes, and relationships to enrich the content of the ontology.

[0022] Furthermore, the object attributes describe the semantic associations between classes, including semantic attributes, temporal attributes, and spatial attributes. Semantic attributes include causal relationships and inclusion relationships. Temporal attributes are used to describe the temporal order of the evolution of accident scenarios in chemical parks, including the parallel, transformation, derivative, and coupled evolution relationships of accident scenarios. Spatial attributes include various relationships describing the location of the accident scenario, emergency resources, and the direction and orientation of emergency evacuation and refuge sites.

[0023] The data attributes describe the characteristics of the class.

[0024] Furthermore, the graph database processing method includes dividing the chemical park's historical accident case data, scenario data, and emergency field knowledge data into structured and unstructured data based on the data structure, selecting a neo4j graph database, and specifying the relationship between the structured data using the neo4j graph database and cypher statements to form knowledge triples describing the chemical park scenario;

[0025] The natural language processing method includes using a named entity recognition method to process unstructured data of historical accident case data of a chemical park, extracting knowledge triples describing the historical accident case data of the chemical park, and identifying various entities in the historical accident case data of the chemical park;

[0026] The method for extracting triples from the large model is to call the application programming interface API of the Deepseek large model, extract corresponding entities and relationships in the unstructured data emergency domain data based on the entities and relationships in the constructed chemical park accident scenario-response domain ontology framework, and obtain knowledge triples describing the emergency decision-making of the chemical park accident scenario;

[0027] The chemical emergency field knowledge graph is formed by aggregating the knowledge triples describing the chemical park scenario, the knowledge triples describing the historical accident cases of the chemical park, and the knowledge triples describing the emergency decision-making of the chemical park accident scenario.

[0028] Furthermore, the structured data refers to data with a fixed format and clear fields, which exists in the form of tables, databases or standardized templates and can be directly mapped to nodes and attributes of the knowledge graph. Structured data includes scenario data; the unstructured data refers to free text without a fixed format. Unstructured data includes historical accident case data of chemical parks and emergency field knowledge data.

[0029] Furthermore, the named entity recognition method is to manually label specific entity categories in historical accident cases in chemical parks, including accident type, accident cause, accident area, accident scenario, accident consequence, and accident loss entity type, using the BIOE labeling method, and perform deep learning training on the labeled data with the BERT-BiLSTM-CRF model, and finally extract knowledge triples describing historical accident cases in chemical parks.

[0030] Furthermore, the BIOE marking method includes B representing the beginning of a specific entity category word in a historical accident case of a chemical park, I representing the middle position of a specific entity category word in a historical accident case of a chemical park, O representing a specific entity category word in a historical accident case of a non-chemical park, and E representing the end of a specific entity category word in a historical accident case of a chemical park.

[0031] Furthermore, the entity relationship information in the chemical emergency knowledge graph is used to enhance the reasoning ability of the large model, and reliable auxiliary emergency decision-making is quickly generated according to the accident scenario description, specifically including:

[0032] Calling the application programming interface (API) of the Deepseek large model to collect the entities of the chemical park accident scenario-response domain ontology and the knowledge graph involved in the accident scenario description, thereby achieving efficient recognition and structured representation of key elements in the accident scenario description text;

[0033] The chemical park accident scenario-response domain knowledge graph is divided into knowledge graph communities with different themes according to the meaning of entities; the API of the Deepseek big model is called to generate a summary for each knowledge graph community, and the summary content includes key entities, important relationships and emergency decision-making models within the community.

[0034] When encountering a sudden accident scenario, emergency decision makers input the accident scenario description and extract key scenario elements as entity prompt words; the big model searches in knowledge graph communities of different topics based on the entity prompt words, and each graph community generates corresponding community answers based on the entity prompt words. Finally, the big model integrates the content of the answers from each graph community to form a global answer and generate reliable auxiliary emergency decisions.

[0035] A computer device of the present invention comprises: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, it implements the chemical park accident auxiliary emergency decision-making generation method based on the knowledge graph enhanced large model as described in any one of claims 1 to 9.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The accident emergency decision-making method of the chemical industry knowledge graph enhanced large model of the present invention can provide a chemical park accident "scenario-response" domain ontology for chemical park accident scenarios, which is conducive to the sharing and reuse of emergency decision-making knowledge, thereby improving the efficiency of emergency decision-making in chemical park accident scenarios; and providing a chemical park emergency field knowledge graph is conducive to enhancing the reasoning ability of the large model based on the rich entity information in the graph, which is conducive to improving the efficiency and reliability of chemical park accident scenario emergency decision-making, and significantly improving the efficiency of chemical park accident emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 It is a flow chart of an accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model disclosed in an embodiment of the present invention.

[0040] Figure 2 It is a flow chart of the seven-step method disclosed in the embodiment of the present invention for constructing the "scenario-response" domain ontology method for chemical park accidents.

[0041] Figure 3It is a flowchart of the structured and unstructured domain knowledge text processing method disclosed in the embodiment of the present invention.

[0042] Figure 4 The present invention is a flowchart of a method for generating emergency decision-making for accident scenarios in a chemical park disclosed in an example.

[0043] Figure 5 This is the ontology structure diagram of the “scenario-response” domain of chemical park accidents disclosed in the example of the present invention. DETAILED DESCRIPTION

[0044] In order to enable persons skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0045] The present invention discloses an accident emergency decision-making method for a large model enhanced by a knowledge graph in the chemical industry. The method can provide a chemical park accident "scenario-response" domain ontology for chemical park accident scenarios, which is conducive to sharing and reusing emergency decision-making knowledge, thereby improving the efficiency of emergency decision-making in chemical park accident scenarios; and provides a chemical park emergency field knowledge graph, which is conducive to enhancing the reasoning ability of the large model based on the rich entity information in the graph, and is conducive to improving the efficiency and reliability of generating emergency decisions for chemical park accident scenarios, and significantly improving the efficiency of chemical park accident emergency response; a detailed description is given below.

[0046] like Figure 1 As shown, an accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to an embodiment of the present invention includes the following steps:

[0047] 101. Real-time collection of historical accident case data, scenario data and emergency knowledge data in chemical parks.

[0048] In an embodiment of the present invention, the optional real-time collection of historical accident case data, scenario data, and emergency domain knowledge data of a chemical park may include:

[0049] Historical chemical park accident case data: This data is collected through accident investigation reports from chemical parks at home and abroad, such as from the official website of the Ministry of Emergency Management, annual reports from chemical park management departments, and case analysis reports published by professional chemical safety research institutions. These reports contain detailed information on the accident process, cause analysis, emergency response measures, and ultimate accident consequences, providing a rich historical data foundation for building the knowledge graph.

[0050] Scenario data: This data is obtained through various monitoring systems deployed within the chemical park. It covers disaster scenario monitoring data (such as real-time monitoring data uploaded by equipment such as fire detectors, combustible gas alarms, and toxic gas detectors), equipment and facility operation data (including parameters such as pressure, temperature, and liquid level of pressure vessels, and the operating status of pumps), hazardous chemical storage data (such as the liquid level, temperature, and pressure of each storage tank, as well as the inventory of hazardous chemicals), various sensor monitoring data (such as wind direction, wind speed, temperature, and humidity monitored by weather stations), and a hazardous chemical catalog (which details the names, physical and chemical properties, hazardous characteristics, and storage locations of various hazardous chemicals stored and used within the park). This data is stored in a structured format in the chemical park's management information system, providing a basis for real-time monitoring of the park's operating status and potential risks.

[0051] Emergency knowledge data: This data is obtained from emergency knowledge texts, such as current regulations and standards (such as the "Regulations on the Safety Management of Hazardous Chemicals" and the "Chemical Enterprise Emergency Rescue Management Specifications" issued by the State Administration of Work Safety, as well as relevant international chemical safety standards such as ISO 17776 "Process Industry Safety Management System"), chemical park emergency plans (including comprehensive emergency plans, special emergency plans, and on-site disposal plans, covering accident prevention, emergency organization, emergency response procedures, emergency resource guarantees, and other aspects), and on-site disposal plans (detailed disposal measures formulated for specific accident types and on-site conditions, such as fire fighting plans, leak plugging plans, and personnel evacuation plans). These text materials provide authoritative guidance and standards for emergency decision-making and are an important source of knowledge for constructing knowledge graphs.

[0052] This can provide comprehensive emergency domain knowledge for constructing the domain ontology and knowledge graph of the "scenario-response" domain of chemical park accidents and generating auxiliary emergency decision-making.

[0053] 102. The seven-step method is used to construct the chemical park accident "scenario-response" domain ontology based on the collected historical accident case data, scenario data and emergency knowledge data of the chemical park.

[0054] In one embodiment of the present invention, optionally, a seven-step method is used to construct a chemical park accident "scenario-response" domain ontology based on the collected historical accident case data, scenario data and emergency domain knowledge data of the chemical park.

[0055] In one embodiment of the present invention, Figure 2 As shown in the figure, the seven-step method is used to construct the chemical park accident "scenario-response" domain ontology, which includes the following steps:

[0056] Step 1. Determine the domain of ontology construction: Clearly define the chemical accident scenario emergency response decision-making domain as the target domain for ontology construction, focus on the emergency decision-making needs of chemical park accidents, and provide a clear direction for subsequent ontology construction work.

[0057] Step 2: Reuse existing disaster and emergency plan ontologies: Refer to existing disaster ontologies (e.g., natural disaster ontologies, technological disaster ontologies, etc.) and emergency plan ontologies (e.g., chemical park emergency plan ontologies, hazardous chemical accident emergency plan ontologies, etc.), draw on mature concepts, terms, and relationships, avoid duplication, and improve the efficiency and quality of ontology construction. For example, when describing a fire accident, you can refer to the definition, classification, and characteristics of fire in the natural disaster ontologies; in terms of emergency response processes, you can refer to the emergency response level classification and emergency handling procedures in the emergency plan ontologies.

[0058] Step 3: Determine professional terminology: Experts in the field of chemical park emergency response decision-making will determine professional terminology to describe chemical park scenarios and emergency response decision-making knowledge based on the characteristics and evolution of accident scenarios, combined with emergency knowledge such as chemical park emergency plans and on-site disposal plans. For example, "disaster-prone environment" is used to describe the macro-environmental factors that may lead to accidents, including natural disasters (earthquakes, tsunamis, etc.) and technological disasters (fires, explosions, etc.); "hazard factors" are used to describe the specific risk factors that cause accidents, such as the physical and chemical properties of various hazardous chemicals (flash point, boiling point, explosion limit, etc.); and "hazard-bearing objects" are used to describe the objects affected by the accident, including various equipment, facilities, and personnel within the chemical park.

[0059] Step 4. Define categories and their subcategories: Based on the theory of disaster systems, the categories related to accident scenarios and emergency response decision-making domain knowledge are divided into four categories: disaster-prone environment, disaster-causing factors, disaster-bearing bodies, and emergency response decisions, and further subdivided into subcategories. For example, the disaster-prone environment subcategory includes the description of accident scenario characteristics of natural disasters such as earthquakes and tsunamis, as well as technical disasters such as fires and explosions, as well as the domain knowledge of the transformation and coupling of accident disaster scenarios; the disaster-causing factors subcategory includes the description of the physical and chemical properties of various hazardous chemicals; the disaster-bearing body subcategory includes the knowledge description of monitoring indicators of various equipment, storage tanks, warehouses, etc. within the chemical park; the emergency response decision-making category is divided according to the "Guidelines for the Preparation of Emergency Plans for Production Safety Accidents in Production and Operation Units" (GB / T29639-2023), including accident scenario monitoring and early warning subcategories, event emergency response subcategories, emergency organization responsibilities and tasks subcategories, accident emergency disposal subcategories, emergency resource scheduling subcategories and other emergency domain knowledge subcategories, such as Figure 5 shown.

[0060] Step 5. Define attributes and relationships: Attributes include object attributes and data attributes. Object attributes describe the semantic associations between classes and include semantic, temporal, and spatial attributes. Semantic attributes include causal relationships (such as the causal relationship between the cause and outcome of an accident) and inclusion relationships (such as the inclusion relationship that a certain piece of equipment belongs to a certain workshop). Temporal attributes describe the temporal order of the evolution of accident scenarios in a chemical park, including evolutionary relationships such as parallelism (e.g., a fire and explosion occurring simultaneously), transformation (e.g., a fire triggering an explosion), derivation (e.g., an explosion leading to a toxic gas leak), and coupling (e.g., the interaction between fire and explosion exacerbating the hazard of an accident). Spatial attributes include relationships describing the location of the accident scenario (e.g., the specific device or area where the accident occurred), emergency resources (e.g., the location of emergency resources such as fire hydrants and fire extinguishers), and the direction and orientation of emergency evacuation and shelters. Data attributes describe the characteristics of a class. For example, data attributes describing the characteristics of a natural disaster scenario in a chemical park include disaster type (e.g., earthquake, tsunami, typhoon), disaster occurrence time (specific date and time), and disaster intensity (e.g., earthquake magnitude, tsunami wave height, typhoon wind force).

[0061] As shown in Table 1 below, Table 1 shows the main relationships in the “scenario-response” domain ontology of chemical park accidents.

[0062] Table 1 Main relationships in the “scenario-response” domain ontology of chemical park accidents.

[0063]

[0064]

[0065] Step 6. Define the domain, range, and data type of attributes and relationships: Clearly define the value range, data type, and other characteristics of various attributes and relationships. For example, the value range of the disaster type attribute can be a predefined set of disaster types (earthquake, tsunami, fire, explosion, etc.), with a string data type; the value range of the disaster occurrence time attribute can be in date and time format, with a date and time data type; the value range of the accident scenario parallel relationship can be Boolean (yes or no), with a Boolean data type, etc. By clarifying these definitions, we ensure the accuracy and consistency of the data in the ontology, facilitating subsequent knowledge graph construction and querying.

[0066] As shown in Table 2 below, Table 2 shows the domain, scope, and data type of the attributes and relationships of the “scenario-response” domain ontology for chemical park accidents.

[0067] Table 2 Main attributes and data types of the “scenario-response” domain ontology for chemical park accidents.

[0068]

[0069]

[0070] Step 7: Supplement instance data: After defining categories, attributes, and relationships, and drawing on actual chemical park accident cases and emergency response data, specific instance data is populated for each category and its subcategories, attributes, and relationships to enrich the ontology's content. For example, in the "Disaster-Prone Environment" category, specific information about an earthquake accident can be populated, including the time, location, magnitude, and extent of damage caused. In the "Hazard Factor" category, data on the physical and chemical properties of a hazardous chemical, such as flash point, boiling point, and explosion limit, can be populated. In the "Hazard-Bearing Object" category, operating parameters and monitoring data for a key piece of equipment within the chemical park, such as pressure, temperature, and liquid level, can be populated. In the "Emergency Response Decision" category, information on the emergency response measures for a particular accident can be populated, such as the level of emergency plan activated, emergency measures implemented, and emergency resources mobilized. This instance data provides concrete knowledge units for the construction of the knowledge graph, enabling it to better reflect the actual characteristics of chemical park accidents and the emergency response process.

[0071] 103. Based on the "scenario-response" domain ontology of chemical park accidents, a knowledge graph is constructed using graph database processing, natural language processing and large model triple extraction methods.

[0072] To construct the knowledge graph, it is necessary to clean the historical accident case data, scenario data and emergency knowledge data of the chemical park, and chemical experts determine the features and attributes used for accident emergency decision-making. In one embodiment of the present invention, optionally, based on the "scenario-response" domain ontology framework of chemical park accidents, a graph database processing method is used for structured domain knowledge texts, and natural language processing and large model triple extraction methods are used for unstructured domain knowledge texts to construct a knowledge graph. In this embodiment of the present invention, Figure 3 As shown, Figure 3 The following is a flowchart of the data processing method for structured and unstructured fields. The specific method is as follows:

[0073] 1. Processing structured domain knowledge data: For structured domain knowledge texts such as scenario data, a graph database processing method is used.

[0074] The graph database processing method includes dividing the historical accident case data, scenario data and emergency field knowledge data of the chemical park into structured and unstructured data according to the data structure, selecting the neo4j graph database, and writing cypher statements according to the neo4j graph database running syntax to specify the relationship between the structured data, thereby forming knowledge triples that describe the chemical park scenario.

[0075] Among them, the structured data refers to data with a fixed format and clear fields, usually in the form of tables, databases or standardized templates, which can be directly mapped to nodes and attributes of the knowledge graph, including scenario data; the unstructured data refers to free text without a fixed format, including historical accident case data of chemical parks and emergency field knowledge data.

[0076] Graph databases store data in a graph structure, enabling efficient representation and querying of complex relationships between entities. Entities in scenario data (such as equipment, facilities, and hazardous chemicals) are used as nodes, and relationships between entities (such as ownership, adjacency, and influence) as edges, to construct a contextual knowledge graph for chemical parks. For example, the storage and query capabilities of a graph database can quickly access information on various entities and their relationships within a chemical park's real-time context, providing real-time data support for emergency decision-making.

[0077] 2. Processing unstructured domain knowledge data: Processing unstructured domain knowledge data through natural language processing methods. Natural language processing methods include using named entity recognition methods to process unstructured data of historical accident case data in chemical parks and identify various entities in the historical accident case data of chemical parks.

[0078] For unstructured domain knowledge text, such as historical accident case data from chemical parks, we manually annotated the characteristic entity categories in these historical accident cases using the BIOE annotation method, including accident type, accident cause, accident location, accident scenario, accident consequences, and accident losses. We then used the BERT-BiLSTM-CRF model for deep learning training on the annotated data, ultimately extracting knowledge triples describing the historical accident cases in chemical parks ((He W, Xu Y, Yu Q. BERT-BiLSTM-CRF Chinese Resume Named Entity Recognition Combining Attention Mechanisms[C] / / Proceedings of the 4th International Conference on Artificial Intelligence and Computer Engineering. 2023: 542-547.)). For example, we integrated accident cases of the same type identified in multiple accident investigation reports to form a knowledge graph containing multiple accident instances, where each accident instance serves as a node and the similarity and causal relationships between accident instances serve as edges.

[0079] The BIOE marking method includes B representing the beginning of a specific entity category word in a historical accident case of a chemical park, I representing the middle position of a specific entity category word in a historical accident case of a chemical park, O representing a specific entity category word in a historical accident case of a non-chemical park, and E representing the end of a specific entity category word in a historical accident case of a chemical park.

[0080] For unstructured domain knowledge texts such as emergency domain knowledge data, a method of extracting triples using a large model is used. First, the emergency domain knowledge text is segmented and divided into fragmented texts to facilitate large-scale model processing. For example, an emergency plan text is divided into chapters, clauses, etc. to form multiple text fragments. Then, the emergency entity categories and relationships in the "scenario-response" domain ontology of chemical park accidents are used as prompt words to enhance the semantic understanding and reasoning capabilities of the large model. For example, when processing the emergency response measures text in the emergency plan, the large model is prompted to pay attention to entity categories such as "emergency response level", "emergency disposal measures", and "emergency resource scheduling" and relationships such as "start", "take", and "mobilize". Next, the Deepseek large model API is called to extract a large number of "entity-relationship-entity" triples from the unstructured emergency domain knowledge text, capturing the entities, relationships, and key descriptions in large-scale text information. For example, we extracted the "emergency organization department - responsibility - responsibility" triple from the organizational responsibilities section of the chemical park's overall emergency plan; and the "accident type - measures taken - disposal measures" triple from the emergency response measures section of the chemical park's special emergency plan. Finally, we supplemented and expanded the extracted triples according to the structure of the chemical park accident "scenario-response" domain ontology to construct a knowledge graph for the chemical park emergency domain.

[0081] The API of Deepseek's large model is called for its ability to effectively process long Chinese texts, and to achieve effective word segmentation and processing of historical accident case data, scenario data, and emergency field knowledge data of the chemical park involved in chemical accident emergency decision-making.

[0082] 104. Enhance the reasoning ability of large models by using the entity relationship information in the knowledge graph of chemical emergency field, and quickly generate reliable auxiliary emergency decisions based on the description of accident scenarios.

[0083] In one embodiment of the present invention, optionally, Figure 4 As shown in the figure, the entity relationship information in the knowledge graph is used to enhance the reasoning ability of the large model, and reliable auxiliary emergency decision-making is quickly generated based on the accident scenario description. The specific method is as follows:

[0084] 1. Knowledge Graph Community Division: The knowledge graph for the "scenario-response" domain of chemical park accidents is divided into different thematic knowledge graph communities based on the meaning of entities. For example, it can be divided into the accident scenario community (containing information such as accident type, accident cause, and accident scenario evolution), the emergency resource community (containing information such as emergency supplies, emergency equipment, and emergency personnel), and the emergency response measures community (containing information such as emergency disposal measures and emergency rescue plans). The entities and relationships within each community are relatively concentrated, facilitating targeted processing and analysis by the large model.

[0085] 2. Community summary generation: The Deepseek large model API is called to generate a summary for each knowledge graph community. The summary content elements include key entities, important relationships, and common emergency decision-making patterns within the community. For example, the summary of the accident scenario community may include common accident types in the park (such as fire, explosion, leakage, etc.), typical accident causes (such as equipment failure, operational errors, natural disasters, etc.), and common evolution laws of accident scenarios (such as fire causing explosion, explosion causing toxic gas leakage, etc.); the summary of the emergency resource community may include the storage status of various emergency materials in the park, the distribution location and operating status of emergency equipment, the professional skills and division of responsibilities of emergency personnel, etc.; the summary of the emergency response measures community may include common emergency disposal measures for different types of accidents (such as fire extinguishing methods, leak plugging measures, etc.), the principles and processes for formulating emergency rescue plans, etc.

[0086] 3. Community Search and Answer Generation: When encountering an unexpected accident scenario, emergency decision-makers input a description of the accident scenario into the system, which then extracts key scenario elements as entity prompts. The large-scale model searches knowledge graph communities across different topics based on the entity prompts. Each graph community generates corresponding community answers based on the scenario element prompts. For example, if the accident scenario description mentions key elements such as fire, chemical plant, leak, hazardous materials, and hazardous equipment and facilities, the large-scale model searches the accident scenario community for accident cases and scenario evolution information related to fires and leaks, searches the emergency resource community for information on emergency supplies and equipment near chemical plants, and searches the emergency response measures community for emergency response measures and rescue plans for fires and leaks. Finally, the large-scale model integrates the responses from each graph community to form a global answer, rapidly generating reliable auxiliary emergency decisions. For example, the generated emergency decision may include the corresponding emergency plan level to be activated, the fire extinguishing and leak control measures to be taken, the emergency resources mobilized (such as fire trucks, fire extinguishers, and leak control tools), and the evacuation and rescue plan.

[0087] Through the above method, the present invention can make full use of the entity relationship information in the knowledge graph, enhance the reasoning ability of the large model, quickly generate reliable auxiliary emergency decision-making based on the accident scenario description, and improve the efficiency of chemical park accident emergency response.

[0088] Finally, it should be noted that the method for generating auxiliary emergency decision-making for chemical park accidents based on a knowledge graph-enhanced large model disclosed in the embodiment of the present invention is only a preferred embodiment of the present invention, and is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, ordinary technicians in this field should understand that it is still possible to modify the technical solutions recorded in the aforementioned embodiments, or to make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for emergency decision-making based on a knowledge graph-enhanced large model in the chemical industry, characterized in that: The following steps are involved: Real-time collection of historical accident case data, scenario data, and emergency knowledge data in chemical parks; The seven-step method is used to construct the chemical park accident scenario-response domain ontology based on the collected historical accident case data, scenario data and emergency domain knowledge data of the chemical park; Based on the chemical park accident scenario-response domain ontology, a knowledge graph for chemical emergency response is constructed using graph database processing, natural language processing, and large-scale model triple extraction methods. The entity relationship information in the knowledge graph of chemical emergency is used to enhance the reasoning ability of the large model, and reliable auxiliary emergency decisions are quickly generated based on the accident scenario description.

2. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 1 is characterized in that: The historical accident case data of the chemical park is obtained through the accident investigation report texts of domestic and foreign chemical parks; the scenario data is obtained with the help of various monitoring systems deployed within the chemical park; and the emergency field knowledge data is obtained from the emergency field knowledge texts of current laws and regulations.

3. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 1 is characterized in that: The seven-step method for constructing the chemical park accident scenario-response domain ontology includes the following steps: Step 1: Determine the domain of ontology construction. The chemical accident scenario emergency response decision-making domain is selected as the target domain of ontology construction, focusing on the emergency decision-making needs of chemical park accidents. Step 2: Reuse the existing disaster ontology and emergency plan ontology; Step 3: Determine the professional terms that describe the chemical park scenario and emergency response decision-making domain knowledge; Step 4: Define categories and subcategories: Based on disaster system theory, the categories related to accident scenarios and emergency response decision-making domain knowledge are divided into four categories: disaster-prone environment, hazard-causing factors, disaster-bearing bodies, and emergency response decisions, and further subcategorized; Step 5: Define attributes and relationships. Attributes include object attributes and data attributes. Step 6: Define the domain, range, and data type of attributes and relationships, and determine the value range and data type of each attribute and relationship; Step 7: Supplement instance data: Based on the categories, attributes, and relationships, combined with actual chemical park accident cases and emergency response data, fill in specific instance data for each category and its subcategories, attributes, and relationships to enrich the content of the ontology.

4. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 3 is characterized in that: The object attributes describe the semantic associations between classes, including semantic attributes, temporal attributes, and spatial attributes. Semantic attributes include causal relationships and inclusion relationships. Temporal attributes are used to describe the temporal order of the evolution of accident scenarios in the chemical park, including the parallel, transformation, derivative, and coupled evolution relationships of accident scenarios. Spatial attributes include various relationships that describe the location of the accident scenario, emergency resources, and the direction and orientation of emergency evacuation and refuge sites. The data attributes describe the characteristics of the class.

5. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 1 is characterized in that: The graph database processing method includes dividing the chemical park's historical accident case data, scenario data, and emergency knowledge data into structured and unstructured data based on the data structure, selecting a neo4j graph database, and specifying the relationship between the structured data using the neo4j graph database and cypher statements to form knowledge triples describing the chemical park scenario; The natural language processing method includes using a named entity recognition method to process unstructured data of historical accident case data of a chemical park, extracting knowledge triples describing the historical accident case data of the chemical park, and identifying various entities in the historical accident case data of the chemical park; The method for extracting triples from the large model is to call the application programming interface API of the Deepseek large model, extract corresponding entities and relationships in the unstructured data emergency domain data based on the entities and relationships in the constructed chemical park accident scenario-response domain ontology framework, and obtain knowledge triples describing the emergency decision-making of the chemical park accident scenario; The chemical emergency field knowledge graph is formed by aggregating the knowledge triples describing the chemical park scenario, the knowledge triples describing the historical accident cases of the chemical park, and the knowledge triples describing the emergency decision-making of the chemical park accident scenario.

6. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 5 is characterized in that: The structured data refers to data with a fixed format and clear fields, which exists in the form of tables, databases or standardized templates and can be directly mapped to nodes and attributes of the knowledge graph. Structured data includes scenario data; the unstructured data refers to free text without a fixed format. Unstructured data includes historical accident case data of chemical parks and emergency field knowledge data.

7. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 5 is characterized in that: The named entity recognition method is to manually label specific entity categories in historical accident cases in chemical parks, including accident type, accident cause, accident area, accident scenario, accident consequence, and accident loss entity type, using the BIOE labeling method, and perform deep learning training on the labeled data with the BERT-BiLSTM-CRF model, and finally extract knowledge triples describing historical accident cases in chemical parks.

8. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 5 is characterized in that: The BIOE marking method includes B representing the beginning of a specific entity category word in a historical accident case of a chemical park, I representing the middle position of a specific entity category word in a historical accident case of a chemical park, O representing a specific entity category word in a historical accident case of a non-chemical park, and E representing the end of a specific entity category word in a historical accident case of a chemical park.

9. The accident emergency decision-making method based on a chemical industry knowledge graph-enhanced large model according to claim 1 is characterized in that: The above method enhances the reasoning ability of the large model by using the entity relationship information in the chemical emergency knowledge graph, and quickly generates reliable auxiliary emergency decision-making based on the accident scenario description, specifically including: Calling the application programming interface (API) of the Deepseek large model to collect the entities of the chemical park accident scenario-response domain ontology and the knowledge graph involved in the accident scenario description, thereby achieving efficient recognition and structured representation of key elements in the accident scenario description text; The chemical park accident scenario-response domain knowledge graph is divided into knowledge graph communities with different themes according to the meaning of entities; the API of the Deepseek big model is called to generate a summary for each knowledge graph community, and the summary content includes key entities, important relationships and emergency decision-making models within the community. When encountering a sudden accident scenario, emergency decision makers input the accident scenario description and extract key scenario elements as entity prompt words; the big model searches in knowledge graph communities of different topics based on the entity prompt words, and each graph community generates corresponding community answers based on the entity prompt words. Finally, the big model integrates the content of the answers from each graph community to form a global answer and generate reliable auxiliary emergency decisions.

10. A computer device, characterized in that: include: A memory and a processor and a computer program stored in the memory, when the computer program is executed on the processor, implements the chemical park accident auxiliary emergency decision-making generation method based on the knowledge graph enhanced large model as described in any one of claims 1 to 9.

Images