A tunnel fire intelligent decision support method, device and medium

Abstract

This invention relates to a method, apparatus, and medium for intelligent decision support in tunnel fires. The method includes the following steps: constructing an ontology layer comprising a conceptual knowledge graph, a factual knowledge graph, and a normative knowledge graph; acquiring multi-source heterogeneous data related to tunnel fires, extracting structured triples and textual semantics; using a graph database as the core storage medium, instantiating structured triples into nodes, relations, and attributes based on the constructed ontology layer, and mounting textual semantics as attributes or nodes into the corresponding knowledge graph of the ontology layer; acquiring a natural language question input by the user, identifying the core intent, converting the natural language question into a standard graph query statement for a specific knowledge graph level based on the type of the core intent, retrieving it in the corresponding knowledge graph, generating and outputting a fire decision support answer based on the retrieval results. Compared with existing technologies, this invention has the advantages of significantly improving the efficiency and scientific rigor of tunnel disaster prevention design and emergency decision-making.

Description

Technical Field

[0001] This invention relates to the fields of disaster prevention and mitigation in tunnel engineering and artificial intelligence technology, and in particular to a method, device and medium for intelligent decision support in tunnel fires. Background Technology

[0002] With the continuous expansion of global tunnel construction, tunnel fires have become a frequent and socially impactful disaster. Due to the enclosed and unique environment of tunnels, fires often result in severe casualties and property damage. Effective fire prevention, emergency response, and post-disaster recovery rely heavily on a comprehensive and structured knowledge system.

[0003] However, existing knowledge on tunnel fire safety suffers from significant fragmentation, widely scattered across heterogeneous data from multiple sources, including academic papers, technical manuals, standards and regulations, and accident investigation reports. This fragmentation makes it difficult to achieve real-time retrieval and application of critical information in actual engineering design and emergency response scenarios. Traditional methods relying on manual review of standards and documents are inefficient and prone to errors, severely hindering the efficiency of scientific decision-making and emergency management.

[0004] Furthermore, while large language models demonstrate powerful capabilities in natural language understanding and generation, they often exhibit a "hallucination" phenomenon when dealing with specialized problems in vertical domains, such as tunnel fires, generating factual errors or outdated knowledge. Traditional domain-specific knowledge graphs, while possessing high accuracy and interpretability, have limitations in natural language interaction and semantic reasoning capabilities.

[0005] Chinese patent CN120542919A discloses a method for identifying building safety risks using a large language model-assisted knowledge graph. This method integrates unstructured text, structured tables, and technical specification data, uses GPT-4 for triple extraction, and combines RAG retrieval enhancement and self-refinement mechanisms to construct a knowledge graph. Features are integrated through Node2Vec graph embedding and attention mechanisms, and model parameters are optimized using Bayesian optimization and feature selection algorithms to achieve automated feature selection and model tuning. However, this method is not suitable for tunnel fire scenarios. Furthermore, its technical approach leads to ambiguous system output, opaque decision-making logic, and a knowledge organization format unfavorable for rule-based reasoning. It also suffers from low retrieval efficiency and cannot meet the urgent need for real-time, accurate, interpretable, and specification-based intelligent decision support in complex emergency scenarios such as tunnel fires. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies in the field of tunnel fires, such as fragmented knowledge, low retrieval efficiency, and insufficient specialization of general large models. It provides a method, device, and medium for intelligent decision support in tunnel fires. By constructing a multi-dimensional domain knowledge graph and combining retrieval enhancement generation technology with a large language model, it achieves rapid and accurate retrieval and intelligent reasoning of professional knowledge.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for intelligent decision support in tunnel fires includes the following steps: Step S1: Determine the domain and scope of fire-related ontology construction, identify important concepts and terms, define generalization relationships between concepts to achieve attribute inheritance, define aggregation relationships to represent physical composition, set data type constraints and value ranges for attributes, and perform logical consistency verification through an inference engine, thereby constructing an ontology layer that includes a concept knowledge graph, a fact knowledge graph, and a normative knowledge graph. Step S2: Obtain multi-source heterogeneous data related to tunnel fires, including unstructured text data and other types of data; based on the unstructured text data, use a hybrid extraction strategy combining statistics and large language models to extract structured triples and construct a structured corpus; based on other types of data, use large language models to extract text semantics. Step S3: Using a graph database as the core storage medium, based on the ontology layer constructed in step S1, the structured triples are instantiated as nodes, relations and attributes, and the extracted text semantics are mounted as attributes or nodes into the corresponding knowledge graph of the ontology layer. Step S4: Obtain the natural language question input by the user, identify the core intent, and based on the type of the core intent, convert the natural language question into a standard graph query statement for a specific knowledge graph level, perform a search in the corresponding knowledge graph, and generate and output a fire decision support answer based on the search results.

[0008] The construction of the conceptual knowledge graph is specifically as follows: a top-down modeling method is adopted, with tunnel fire as the root node. The classification hierarchy of highway and railway tunnel fires is constructed using generalization relations. The physical composition structure of tunnel ventilation, lighting and sprinkler subsystems is constructed using aggregation relations. Attribute value relations are used to assign specific attributes such as service life and maximum flow rate to entities, forming a multi-dimensional conceptual semantic network.

[0009] The construction of the aforementioned factual knowledge graph specifically involves: constructing subgraphs in three dimensions—pre-disaster, during-disaster, and post-disaster—based on the logic of the entire lifecycle. The pre-disaster static environment subgraph maps the geometric parameters of the tunnel structure and the topological location of the equipment into static entity nodes and associated edges. The during-disaster dynamic evolution subgraph establishes time slice nodes or dynamic attribute sequences to record the spatiotemporal evolution data of heat release rate, smoke spread rate, and temperature field distribution. The post-disaster response and assessment subgraph constructs a subject-responsibility-action relationship model to describe the emergency collaboration logic and constructs a cause-event-effect causal inference model to link the fire initiation cause with the disaster consequences.

[0010] The construction of the specified knowledge graph is as follows: through semantic parsing, regulations in unstructured text data are reconstructed into three topological structures: parameter tables, computational theories, and logical judgments. Specifically, the parameter table structure constructs a three-layer graph structure of specification documents, table objects, and data entries, connecting data entries with row-family relationships and storing parameter values ​​as attributes of entry nodes. The computational theory structure constructs a directed graph with computational nodes at its core, defining input relationship edges from parameter nodes to computational nodes and output relationship edges from computational nodes to result parameter nodes, and storing parsable mathematical formulas in the computational nodes. The logical judgment structure constructs a subject-related constraint structure, connecting engineering entity nodes and logical rule nodes using constraint relationships, and storing comparison operators, thresholds, or existence identifiers based on the type attributes of quantitative or existence constraints in the rule nodes.

[0011] The hybrid extraction strategy specifically includes: Corpus Acquisition and Core Concept Classification Extraction: A basic corpus is constructed by crawling multi-source unstructured text using Python tools. The word frequency weights are calculated using the TF-IDF algorithm. Based on this, the extracted high-frequency core terms are mapped to five entity categories: tunnel civil engineering structure, electromechanical facilities and equipment, fire environmental characteristics, emergency control measures, and regulatory standard parameters, thus completing the preliminary domain vocabulary filtering and classification. Joint entity-relation extraction based on instruction fine-tuning: Construct a hybrid instruction dataset containing simple entity queries and complex multi-hop reasoning, and perform supervised fine-tuning of the large language model to adapt it to the professional context of tunnel fires; use the fine-tuned large language model in conjunction with manual verification to identify and extract entity, attribute and relation triplet from unstructured text. Constructing a multidimensional semantic association framework: For accident report texts, a cause-event-effect causal inference model and a subject-responsibility-action relationship model are used for extraction, specifically covering the deep semantic relationships between basic entities, fire causes, environmental evolution parameters, and disaster consequences.

[0012] Step S4 specifically includes: Step S41: Use a large language model to extract features from the natural language questions input by the user. By identifying keyword types and entity attributes, determine the core intent. If it contains definitions, meanings, or similar explanatory words and the subject is a general term, it is identified as a concept definition class. If it contains specific project names, time and location features, and causes, consequences, or other related retrospective words, it is identified as a historical fact class. If it contains numerical parameters and words related to calculation, solution, or other related operational intents, it is identified as a standard calculation class. Step S42: Based on the type of the identified core intent, the natural language is converted into a standard graph query statement for a specific graph level using a corresponding method, and the standard graph query statement is used to perform a retrieval in the corresponding knowledge graph. Step S43: Assemble the retrieved concept definitions, historical triple paths, calculation results, and logical verification status into a structured evidence chain; input the prompt engineering template containing role settings, original questions, retrieved evidence, and restriction instructions, along with the structured evidence chain, into the large language model, force the model to generate a response based on the structured evidence, and label the knowledge source of the answer to obtain the fire decision support answer.

[0013] The process of converting natural language into standard graph query statements for a specific graph level, and then retrieving information from the corresponding knowledge graph based on these standard graph query statements, specifically includes: For concept definition problems, multi-hop queries are performed in the concept knowledge graph to retrieve common attributes of parent classes along generalization relationships and subsystem structures along aggregation relationships to supplement the contextual semantics of terms. For questions related to historical facts, the time, location, and accident entity nodes are anchored in the factual knowledge graph to retrieve the disaster evolution sequence and emergency records of the responsible parties and their responsibilities. For standardized calculation problems, parameter nodes and calculation logic nodes are located in the standardized knowledge graph, the input relation edges are traversed to obtain variable values, the mathematical formula strings stored in the nodes are parsed for dynamic reasoning, or compliance logic judgments are performed based on constraint relationships.

[0014] The process of constructing an ontology layer that includes a conceptual knowledge graph, a factual knowledge graph, and a normative knowledge graph also includes performing a cross-domain reusability assessment of existing ontology: retrieving and analyzing existing mature ontology in relevant domains, verifying its applicability in the specific closed space scenario of the tunnel through semantic mapping, and screening reusable class and relation structures to reduce construction redundancy and improve the standardization of the knowledge system.

[0015] A tunnel fire intelligent decision support device includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described therein.

[0016] A storage medium having a program stored thereon, which, when executed, implements the method.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention solves the problem of scattered and disorganized knowledge about tunnel fires by constructing a graph covering three types: concepts, facts, and norms. It not only realizes the semantic integration and dynamic expansion of multi-source heterogeneous data, but also gives the knowledge base a clear logical structure and internal consistency. At the same time, the graph database-based storage mode supports the dynamic expansion and updating of knowledge. When new regulations are issued or new cases occur, they can be incorporated into the corresponding graph level as new nodes and relationships, ensuring the timeliness and completeness of the knowledge system.

[0018] (2) This invention uses a large language model as a tool for information extraction and semantic understanding, and its output is anchored to a structured knowledge graph. When a user asks a question, retrieval and reasoning are performed in the knowledge graph. The large language model is mainly used for the natural language organization of the final answer. This approach fundamentally constrains the answer to the scope of existing and verified knowledge, effectively suppressing the "illusion" problem of the large language model and ensuring the accuracy of the question-answering system when it involves standardized parameters, calculation formulas, and historical cases. In addition, any decision suggestion can be traced back to a specific node in the graph, forming an auditable and verifiable decision-making link, which greatly improves the credibility and reliability of the system.

[0019] (3) This invention not only supports simple knowledge retrieval, but also assists in complex engineering calculations and compliance reviews by standardizing the computational logic and logical judgment rules in the knowledge graph, significantly improving the efficiency and scientific nature of tunnel disaster prevention design and emergency decision-making. By standardizing the domain logic encoded in the knowledge graph, rule-based proactive reasoning is realized, which can provide in-depth decision support with both scientific basis and operability for tunnel disaster prevention design, emergency plan formulation and real-time emergency command, significantly improving the efficiency and scientific nature of the decision-making process. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the construction of the tunnel fire body layer in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hierarchical network of the concept knowledge graph in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the pre-disaster stage classification of the fact knowledge graph in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the disaster-prone stage classification of the fact knowledge graph in an embodiment of the present invention; Figure 6This is a schematic diagram illustrating the post-disaster stage classification of the fact knowledge graph in an embodiment of the present invention; Figure 7 This is a flowchart illustrating the natural language question answering process in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0022] This embodiment provides an intelligent decision support method for tunnel fires, such as... Figure 1 As shown, it includes the following steps: Step S1: Determine the domain and scope of fire-related ontology construction, identify important concepts and terms, define generalization relationships between concepts to achieve attribute inheritance, define aggregation relationships to represent physical composition, set data type constraints and value ranges for attributes, and perform logical consistency verification through inference engine, thereby constructing an ontology layer containing concept knowledge graph, fact knowledge graph, and normative knowledge graph.

[0023] like Figure 2 As shown, the specific steps include: Step S11: Define the scope and business boundaries of the domain ontology. Systematically collect heterogeneous data from multiple sources, including academic publications, expert experience databases, and industry standards. Through requirements analysis, determine the capability questions that the ontology needs to answer, thereby clarifying the core concept set that the tunnel fire knowledge graph needs to cover, such as structure, facilities, environment, control, and regulations.

[0024] Step S12: Analyze the feasibility of reusing existing ontologies. Retrieve and analyze existing mature ontologies in general building fire, disaster risk assessment, and other related fields. Verify their applicability in the specific enclosed space scenario of a tunnel through semantic mapping. Select reusable classes and relational structures to reduce construction redundancy and improve the standardization of the knowledge system.

[0025] Step S13 establishes a core concept set encompassing tunnel civil engineering structure, electromechanical facilities and equipment, fire environment characteristics, emergency control measures, and regulatory and standard parameters. A corpus is constructed by crawling relevant papers on tunnel fires from a database using the Python Selenium tool, ensuring the crawling process is legal and compliant. The TF-IDF algorithm is used to calculate word frequency weights and extract keywords from the corpus. Based on this, the extracted high-frequency core concepts are divided into five entity categories. The first category is tunnel civil engineering structure, including entities such as the tunnel body, vehicular cross passages, pedestrian cross passages, and fire-resistant coating layers. The second category is electromechanical facilities and equipment, including entities such as jet fans, smoke exhaust fans, heat-sensing optical fibers, water sprinkler heads, and evacuation signs. The third category is fire environment characteristics, including entities such as heat release rate, smoke spread rate, visibility, CO concentration, and ceiling temperature. The fourth category is emergency control measures, including entities such as traffic control, evacuation guidance, fan forward and reverse rotation, and smoke exhaust valve opening. The fifth category is regulatory and standard parameters, including physical parameters such as design critical wind speed, mandatory evacuation time, and fire resistance limit requirements.

[0026] Step S14: Construct the semantic topology and hierarchical system between concepts. Establish standardized semantic relationships between defined class sets, focusing on defining the generalization relationship IsA that describes the parent-child inheritance logic, so as to realize the automatic inheritance of common attributes of the parent class by the subclass. At the same time, describe the aggregation relationship PartOf of the whole-part logic to characterize the physical composition of the tunnel complex system, thereby forming a multi-dimensional and deductive domain knowledge classification tree.

[0027] Step S15: Define the attribute dimensions and data constraints of the class. Define internal attributes for each specific class, including physical parameters such as power and fire resistance rating, status parameters such as on / off, and management parameters such as maintenance cycle. Simultaneously, formally define the data types of attribute values ​​at the ontology schema layer, such as floating-point, boolean, and enumeration types, along with corresponding value range constraints, providing a metadata foundation for subsequent data integrity verification.

[0028] Step S16: Ontology instantiation and logical consistency reasoning verification. Specific entities are populated into the constructed ontology framework to generate instances, and a logic check is performed using an inference engine to automatically identify and correct circular references, attribute conflicts, or logical contradictions in class definitions, ensuring the self-consistency of the knowledge base.

[0029] like Figure 3As shown, this embodiment employs a top-down ontology modeling approach to construct a conceptual knowledge graph. At the outset, three core semantic relationships are established: the first is the generalization relationship IsA (inheritance), used to describe the parent-child hierarchy between concepts. The second is the aggregation relationship PartOf (composition), used to describe the structural relationship between the whole and its parts. The third is the attribute value relationship PropertyValue, used to describe the association between entities and specific numerical values ​​or features. Based on these relationship definitions, the specific construction steps are as follows: Step A1: Establish the top-level concept and construct the inheritance hierarchy. Set tunnel fire as the root node, and use the generalization relation IsA to derive subclasses for highway tunnel fire and railway tunnel fire, thus establishing the basic classification hierarchy; Step A2: Construct entity associations based on composition relationships. For the concept of highway tunnel equipment and facilities, for example, using the aggregation relationship PartOf, connect it with the ventilation system, lighting system, and sprinkler system, thus clarifying the physical composition structure of the system; Step A3: Refine the subsystem structure. For example, under the ventilation system node, the PartOf relationship is used to connect the fan and smoke exhaust duct sub-components; under the sprinkler system node, the PartOf relationship is used to connect the fire hydrant and sprinkler valve sub-components, thus achieving a deep deconstruction of complex equipment and facilities. Step A4: Define attribute dimensions to complete the graph instantiation. Specific parameters are assigned to entities through PropertyValue relationships. For example, a lifespan attribute is associated with a fire hydrant entity, and a maximum flow rate attribute is associated with a sprinkler valve entity, ultimately forming a multi-dimensional conceptual semantic network.

[0030] like Figures 4-6 As shown, in order to achieve accurate characterization and tracing of the entire life cycle of tunnel fires, this embodiment constructs a tunnel fire factual knowledge graph containing three temporal dimensions based on the full life cycle logic of prevention, evolution, and response: Pre-disaster static environment sub-diagram: such as Figure 4 As shown, the tunnel civil structure and electromechanical facilities are mapped as entity nodes in the graph, and their static attributes and topological locations are defined. Specifically, geometric parameters such as tunnel length, width, and cross-section are used as attribute data of the structural entities, and the installation locations and system connections of equipment such as fire alarm detectors and ventilation and smoke exhaust fans are established as associated edges between equipment entities, thereby forming the initial boundary conditions for the occurrence of a fire. Disaster dynamic evolution subgraph: such as Figure 5As shown, the focus is on the structured storage of the spatiotemporal evolution characteristics of fire scene environmental parameters. Time slice nodes or dynamic attribute sequences are established to record key physical quantities in the fire development process, specifically covering heat release rate curves, maximum temperature field distribution, smoke spread velocity vector, and visibility attenuation, to achieve dynamic tracking of the fire's physical state. Post-disaster response and assessment sub-diagram: as shown Figure 6 As shown, an emergency response workflow and accident causal chain involving multiple stakeholders are modeled. On one hand, a stakeholder-responsibility-action relationship model is constructed to clarify the collaborative logic among tunnel management, traffic police, and fire and rescue forces in the emergency plan. The implementation process of the stakeholder-responsibility-action relationship model includes: Step B1: Pre-set the emergency response entity and define the set of responsibilities for each entity; Step B2: Based on the emergency plan text and accident report text, jointly extract the main entities, responsible entities, and action entities, and perform synonym unification and unique identification; Step B3: Generate directed relation edges according to the triple <subject, undertake, responsibility>, <responsibility, trigger, action> and <subject, execution, action>, and attach attributes such as trigger conditions, priority, resource requirements and execution status to the action nodes; Step B4: Write the nodes, relationships and attributes into the graph database to form a queryable collaborative business flow network, and implement responsibility missing detection, action conflict detection and cross-entity handover link tracking based on rule constraints.

[0031] On the other hand, a cause-event-effect causal inference model is constructed, connecting fire-causing factors such as traffic accidents and electrical faults with disaster consequences such as casualty statistics and property damage assessments through semantic relationships, supporting post-disaster source tracing analysis. The implementation process of the causal inference model includes: Step C1: Define the ontology type and relationship type of the causal node, event node and consequence node. The relationship type includes at least the following: cause, trigger, aggravate, inhibit and evidence support. Step C2: Extract the causes of fire (including traffic accidents, electrical faults, etc.), key events (including fire, power outage, smoke exhaust equipment failure, evacuation activation, etc.) and disaster consequences (including casualty statistics, property damage, road closure duration, etc.) from accident reports, monitoring logs and handling records, and assign timestamps, spatial locations and severity attributes to each node; Step C3: Generate causal edges between cause, event, and consequence based on the causal rule base and assign confidence weights, wherein the confidence weights are updated according to the evidence support relationship; Step C4: After inputting observational evidence after the disaster, perform forward extrapolation and backward tracing to output the most likely causal chain path and corresponding consequence assessment results.

[0032] To enable unstructured regulatory and standard texts to be directly invoked and reasoned by computer programs, this embodiment employs semantic parsing technology to reconstruct them into three specific topological structures: parameter table structure, computational theory structure, and logical judgment structure.

[0033] The parameter table structure constructs a three-layer graph structure: specification document, table object, and data entry. Data entries are connected by row-level relationships, and parameter values ​​are stored as attributes of entry nodes. Taking a vehicle heat release rate table from a specification as an example, this embodiment uses a three-layer graph structure for storage. The first layer establishes a specification document node as the root node, used to represent the specification's metadata. The second layer establishes table object nodes, connected to the specification document node through inclusion relationships, used to locate the specific table source. The third layer establishes data entry nodes, connected to the table object node through row-level relationships, with each row of data corresponding to an independent data entry node. The vehicle model name is used as the search key for the entry node, and the corresponding heat release rate (MW) value is stored as the attribute value of that node. This structure allows the system to quickly locate and query specific parameters corresponding to different vehicle models by traversing the specification and table nodes.

[0034] The computational theory-based structure constructs a directed graph centered on computational nodes. It defines input relationship edges from parameter nodes to computational nodes, and output relationship edges from computational nodes to result parameter nodes. The computational nodes store parsable mathematical formulas. Taking the calculation of the required air volume for diluting smoke in a specific regulation as an example, the structure is constructed as follows: First, two types of nodes are defined: computational nodes representing the computational logic and parameter nodes representing physical quantities. Second, two types of directed edges are defined. The first type is input relationship edges, from parameter nodes representing the design smoke concentration and smoke emission to computational nodes. The second type is output relationship edges, from computational nodes to parameter nodes representing the required air volume. Finally, node attributes are defined: computational nodes store mathematical formula strings that can be parsed by Python and the applicable operating conditions; parameter nodes store the standard SI units, value range, and symbol identifiers of the physical quantities. This structure enables the system to automatically obtain variable values ​​and perform dynamic calculations by traversing the input edges.

[0035] A logical judgment-based structure constructs a main body-related constraint structure, using the `hasConstraint` constraint relationship to connect engineering entity nodes and logical rule nodes. Based on the type attribute of quantitative or existence constraints in the rule nodes, comparison operators, thresholds, or existence indicators are stored. First, a constraint entity node is created, representing the engineering entity to be reviewed, such as a pedestrian crosswalk. Next, a logical rule node is created, storing the specific review criteria. The two are connected by a `hasConstraint` constraint edge. To distinguish between the two constraint types, a `type` attribute is set in the logical rule node. When the `type` attribute value is a quantitative constraint, the node stores comparison operators and threshold attributes, such as greater than or equal to 1.2 meters. When the `type` attribute value is an existence constraint, the node stores existence indicators and associated component attributes, such as the requirement to include fire-resistant lining. This structure allows the graph database to directly distinguish and perform different types of compliance checks through attribute filtering.

[0036] Step S2: Obtain multi-source heterogeneous data related to tunnel fires. The multi-source heterogeneous data includes unstructured text data (papers, accident reports, standards, etc.) and other types of data (images, videos, and other unstructured multimedia data, sensor data). Based on the unstructured text data, a hybrid extraction strategy combining statistics and large language models is used to extract structured triples and construct a structured corpus. Based on the other types of data, large language models are used to extract text semantics.

[0037] In this embodiment, the hybrid extraction strategy first acquires the corpus and extracts core concepts through classification. Then, based on the output terminology list and entity category labeling system, it performs joint entity relation extraction with fine-tuning based on instructions. Finally, for accident report texts, a multi-dimensional semantic association framework is further constructed based on the joint extraction results to supplement deep relationships. The implementation process of the hybrid extraction strategy specifically includes: Corpus Acquisition and Core Concept Classification Extraction: A basic corpus is constructed by crawling multi-source unstructured text using Python tools. The word frequency weights are calculated using the TF-IDF algorithm. Based on this, the extracted high-frequency core terms are mapped to five entity categories: tunnel civil engineering structure, electromechanical facilities and equipment, fire environmental characteristics, emergency control measures, and regulatory standard parameters, thus completing the preliminary domain vocabulary filtering and classification. Joint entity-relation extraction based on instruction fine-tuning: Construct a hybrid instruction dataset containing simple entity queries and complex multi-hop reasoning, and perform supervised fine-tuning of the large language model to adapt it to the professional context of tunnel fires; use the fine-tuned large language model in conjunction with manual verification to identify and extract entity, attribute and relation triplet from unstructured text. Constructing a multidimensional semantic association framework: For accident report texts, a cause-event-effect causal inference model and a subject-responsibility-action relationship model are used for extraction, specifically covering the deep semantic relationships between basic entities, fire causes, environmental evolution parameters, and disaster consequences.

[0038] Step S3: Using a graph database as the core storage medium, based on the ontology layer constructed in step S1, the structured triples are instantiated as nodes, relations, and attributes. The extracted text semantics are used as attributes or nodes and mounted into the corresponding knowledge graph of the ontology layer. An access link pointing to the original file storage address is established, thereby realizing the semantic association management between multimodal data and knowledge graph entities.

[0039] Step S4: Obtain the natural language question input by the user, identify the core intent, and based on the type of the core intent, convert the natural language question into a standard graph query statement for a specific knowledge graph level, perform a search in the corresponding knowledge graph, and generate and output a fire decision support answer based on the search results.

[0040] Step S4 specifically includes: Step S41: Use a large language model to extract features from the natural language questions input by the user, and determine the core intent by identifying keyword types and entity attributes.

[0041] For a sequence of natural language text input by a user, preprocessing is first performed, including stop word removal and text standardization, to form a query request to be parsed. Then, the semantic reasoning capabilities of a large language model are used to extract features and classify the query request, identifying the user's core query intent. Based on the intent type, a corresponding routing strategy is generated to determine which specific level in the concept, fact, or normative knowledge graph to activate. Specifically, if the preprocessed text contains definitions, meanings, or similar explanatory terms, and the subject is a general term rather than a specific item name, it is identified as a concept definition. The system then automatically locates the concept knowledge graph and generates a query to retrieve the definition of the term and its subordinate relationship to the subject. If it contains specific item names, time and location characteristics, and causes, consequences, or other related retrospective terms, it is identified as a historical fact. The system then automatically locates the fact knowledge graph and generates a Cypher statement to accurately extract the attributes of the accident entity. If it contains numerical parameters and words related to calculation, solution, or other related computational intents, it is identified as a normative computation. The system then automatically locates the normative knowledge graph and activates computational theory nodes within it, performing dynamic computation based on the extracted parameters.

[0042] Step S42: Based on the type of the identified core intent, the natural language is converted into a standard graph query statement for a specific graph level using a corresponding method, and the standard graph query statement is used to perform a retrieval in the corresponding knowledge graph.

[0043] For concept definition problems, multi-hop queries are performed in the concept knowledge graph to retrieve the common attributes of the parent class along the generalization relation IsA and the subsystem structure along the aggregation relation PartOf, in order to supplement the contextual semantics of the terms. For questions related to historical facts, the time, location, and accident entity nodes are anchored in the factual knowledge graph to retrieve the disaster evolution sequence and emergency records of the responsible parties and their responsibilities. For standardized computation problems, parameter nodes and computation logic nodes are located in the standardized knowledge graph, variable values ​​are obtained by traversing the input relation edges, and mathematical formula strings stored in the nodes are parsed for dynamic reasoning, or compliance logic judgments are performed based on the hasConstraint relation.

[0044] Furthermore, for complex decision-making requests involving multi-source data correlation analysis, such as compliance reviews and contingency plan assessments, cross-data graph joint retrieval logic is configured to achieve dynamic correlation of heterogeneous knowledge. Specifically, this includes the following steps: Step D1: Scene Parameter Extraction and Mapping. Based on the entity identifiers in the user request, retrieve and extract the current fire scene status parameters from the fact knowledge graph, including environmental data such as temperature and smoke concentration, as well as facility status such as fan operation status and evacuation route occupancy status. Step D2: Rule Logic Invocation and Matching. The extracted scenario parameters are used as input conditions. The corresponding logical judgment nodes are located and activated in the standardized knowledge graph, invoking preset conditional constraint rules and compliance thresholds. Step D3: Execution of Intra-Graph Logic Verification. The parameters and rules are compared within the graph database engine, directly generating and returning Boolean compliance verification results or numerical deviation indicators via the graph query language.

[0045] Step S43: Assemble the retrieved concept definitions, historical triple paths, calculation results, and logical verification status into a structured evidence chain; input the prompt engineering template containing role settings, original questions, retrieved evidence, and restriction instructions, along with the structured evidence chain, into the large language model, force the model to generate a response based on the structured evidence, and label the knowledge source of the answer to obtain the fire decision support answer.

[0046] To achieve the aforementioned logical reasoning and high-precision question answering, this embodiment constructs a retrieval-enhanced generation system architecture specific to the field of tunnel fires, such as... Figure 7 As shown, the workflow of natural language question answering includes: Step E1: Construct a domain-adaptive instruction fine-tuning training process. Construct a hybrid instruction dataset containing simple entity queries and complex multi-hop inference. Utilize the Unsloth optimization framework in conjunction with low-rank adaptation techniques to efficiently fine-tune the parameters of the basic large language model, optimizing the model's ability to accurately convert natural language into semantic mappings of the Cypher language for graph database queries. Step E2: Deploy the link orchestration and integration environment. The LangChain framework is used as the orchestration hub, integrating and calling a localized large language model deployed on Ollam, while establishing a persistent connection with the Neo4j graph database to achieve bidirectional interaction between the inference engine and the knowledge storage layer. Step E3: Configure the dynamic prompt project template. Establish a standardized prompt dictionary, specifically including: Step E31: Graph query generation template configuration. The template embeds the current graph's schema definition, task description instructions, and few-sample reference use cases, guiding the fine-tuned model to generate query statements that conform to syntax rules and are logically correct. Step E32: Response generation template configuration. The template is designed as a combination of the original question, search results, and role settings, guiding the model to generate the final natural language answer in the tone of a domain expert, combined with the structured query results returned by the database; Step E4: Perform full-process retrieval and enhanced reasoning. Named entity recognition technology is used to extract core entities from the user's question, driving the large language model to generate corresponding subgraph query statements. A search is performed in the Neo4j database to obtain the associated triple context. Finally, the question and context are input into the large language model, outputting a professional, accurate, and traceable decision support solution.

[0047] The above is an introduction to the method embodiments. The following describes the solution of the present invention further through device embodiments.

[0048] A tunnel fire intelligent decision support device, comprising: Ontology layer construction module: Determines the domain and scope of fire-related ontology construction, identifies important concepts and terms, defines generalization relationships between concepts to achieve attribute inheritance, defines aggregation relationships to represent physical composition, sets data type constraints and value ranges for attributes, and performs logical consistency verification through inference engine, thereby constructing an ontology layer that includes a concept knowledge graph, a fact knowledge graph, and a normative knowledge graph; Information extraction module: Acquires multi-source heterogeneous data related to tunnel fires, including unstructured text data and other types of data; Based on the unstructured text data, a hybrid extraction strategy combining statistics and large language models is used to extract structured triples and construct a structured corpus; Based on other types of data, large language models are used to extract text semantics. Knowledge graph mounting module: Using a graph database as the core storage medium, based on the ontology layer constructed in step S1, the structured triples are instantiated as nodes, relations and attributes, and the extracted text semantics are mounted as attributes or nodes to the corresponding knowledge graph of the ontology layer. Question answering module: It acquires natural language questions input by users, identifies the core intent, and converts the natural language questions into standard graph query statements for specific knowledge graph levels based on the type of core intent. It then performs a search in the corresponding knowledge graph, generates fire decision support answers based on the search results, and outputs them.

[0049] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0050] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A tunnel fire intelligent decision support method, characterized by, Includes the following steps: Step S1: Determine the domain and scope of fire-related ontology construction, identify important concepts and terms, define generalization relationships between concepts to achieve attribute inheritance, define aggregation relationships to represent physical composition, set data type constraints and value ranges for attributes, and perform logical consistency verification through an inference engine, thereby constructing an ontology layer that includes a concept knowledge graph, a fact knowledge graph, and a normative knowledge graph. Step S2: Obtain multi-source heterogeneous data related to tunnel fires. The multi-source heterogeneous data includes unstructured text data and other types of data, including sensor data and unstructured multimedia data including images and videos. Based on the unstructured text data, a hybrid extraction strategy combining statistics and large language models is used to extract structured triples and construct a structured corpus. Based on the other types of data, large language models are used to extract text semantics. Step S3: Using a graph database as the core storage medium, based on the ontology layer constructed in step S1, the structured triples are instantiated as nodes, relations and attributes, and the extracted text semantics are mounted as attributes or nodes into the corresponding knowledge graph of the ontology layer. Step S4: Obtain the natural language question input by the user, identify the core intent, and based on the type of the core intent, convert the natural language question into a standard graph query statement for a specific knowledge graph level, perform a search in the corresponding knowledge graph, and generate and output a fire decision support answer based on the search results.

2. The intelligent decision support method for tunnel fires according to claim 1, characterized in that, The construction of the conceptual knowledge graph is specifically as follows: a top-down modeling method is adopted, with tunnel fire as the root node. The classification hierarchy of highway and railway tunnel fires is constructed using generalization relations. The physical composition structure of tunnel ventilation, lighting and sprinkler subsystems is constructed using aggregation relations. Attribute value relations are used to assign specific attributes such as service life and maximum flow rate to entities, forming a multi-dimensional conceptual semantic network.

3. The tunnel fire intelligent decision support method according to claim 1, characterized in that, The construction of the aforementioned factual knowledge graph specifically involves: constructing subgraphs in three dimensions—pre-disaster, during-disaster, and post-disaster—based on the logic of the entire lifecycle. The pre-disaster static environment subgraph maps the geometric parameters of the tunnel structure and the topological location of the equipment into static entity nodes and associated edges. The during-disaster dynamic evolution subgraph establishes time slice nodes or dynamic attribute sequences to record the spatiotemporal evolution data of heat release rate, smoke spread rate, and temperature field distribution. The post-disaster response and assessment subgraph constructs a subject-responsibility-action relationship model to describe the emergency collaboration logic and constructs a cause-event-effect causal inference model to link the fire initiation cause with the disaster consequences.

4. The tunnel fire intelligent decision support method according to claim 1, characterized in that, The construction of the specified knowledge graph is as follows: through semantic parsing, regulations in unstructured text data are reconstructed into three topological structures: parameter tables, computational theories, and logical judgments. Specifically, the parameter table structure constructs a three-layer graph structure of specification documents, table objects, and data entries, connecting data entries with row-family relationships and storing parameter values ​​as attributes of entry nodes. The computational theory structure constructs a directed graph with computational nodes at its core, defining input relationship edges from parameter nodes to computational nodes and output relationship edges from computational nodes to result parameter nodes, and storing parsable mathematical formulas in the computational nodes. The logical judgment structure constructs a subject-related constraint structure, connecting engineering entity nodes and logical rule nodes using constraint relationships, and storing comparison operators, thresholds, or existence identifiers based on the type attributes of quantitative or existence constraints in the rule nodes.

5. The tunnel fire intelligent decision support method according to claim 1, wherein, The hybrid extraction strategy specifically includes: Corpus Acquisition and Core Concept Classification Extraction: A basic corpus is constructed by crawling multi-source unstructured text using Python tools. The word frequency weights are calculated using the TF-IDF algorithm. Based on this, the extracted high-frequency core terms are mapped to five entity categories: tunnel civil engineering structure, electromechanical facilities and equipment, fire environmental characteristics, emergency control measures, and regulatory standard parameters, thus completing the preliminary domain vocabulary filtering and classification. Joint entity-relation extraction based on instruction fine-tuning: Construct a hybrid instruction dataset containing simple entity queries and complex multi-hop reasoning, and perform supervised fine-tuning of the large language model to adapt it to the professional context of tunnel fires; use the fine-tuned large language model in conjunction with manual verification to identify and extract entity, attribute and relation triplet from unstructured text. Constructing a multidimensional semantic association framework: For accident report texts, a cause-event-effect causal inference model and a subject-responsibility-action relationship model are used for extraction, specifically covering the deep semantic relationships between basic entities, fire causes, environmental evolution parameters, and disaster consequences.

6. The tunnel fire intelligent decision support method according to claim 1, wherein, Step S4 specifically includes: Step S41: Use a large language model to extract features from the natural language questions input by the user. Determine the core intent by identifying keyword types and entity attributes. If it contains definition, meaning, or explanatory words and the subject is a general term, it is identified as a concept definition class. If it contains specific project names, time and location features, and retrospective words including cause and effect, it is identified as a historical fact class. If it contains numerical parameters and words containing calculation and solution intentions, it is identified as a standard calculation class. Step S42: Based on the type of the identified core intent, the natural language is converted into a standard graph query statement for a specific graph level using a corresponding method, and the standard graph query statement is used to perform a retrieval in the corresponding knowledge graph. Step S43: Assemble the retrieved concept definitions, historical triple paths, calculation results, and logical verification status into a structured evidence chain; input the prompt engineering template containing role settings, original questions, retrieved evidence, and restriction instructions, along with the structured evidence chain, into the large language model, force the model to generate a response based on the structured evidence, and label the knowledge source of the answer to obtain the fire decision support answer.

7. A tunnel fire intelligent decision support method according to claim 6, characterized in that, The process of converting natural language into standard graph query statements for a specific graph level, and then retrieving information from the corresponding knowledge graph based on these standard graph query statements, specifically includes: For concept definition problems, multi-hop queries are performed in the concept knowledge graph to retrieve common attributes of parent classes along generalization relationships and subsystem structures along aggregation relationships to supplement the contextual semantics of terms. For questions related to historical facts, the time, location, and accident entity nodes are anchored in the factual knowledge graph to retrieve the disaster evolution sequence and emergency records of the responsible parties and their responsibilities. For standardized calculation problems, parameter nodes and calculation logic nodes are located in the standardized knowledge graph, the input relation edges are traversed to obtain variable values, the mathematical formula strings stored in the nodes are parsed for dynamic reasoning, or compliance logic judgments are performed based on constraint relationships.

8. The tunnel fire intelligent decision support method of claim 1, wherein, The process of constructing an ontology layer that includes a conceptual knowledge graph, a factual knowledge graph, and a normative knowledge graph also includes performing a cross-domain reusability assessment of existing ontology: retrieving and analyzing existing mature ontology in relevant domains, verifying its applicability in the specific closed space scenario of the tunnel through semantic mapping, and screening reusable class and relation structures to reduce construction redundancy and improve the standardization of the knowledge system.

9. A tunnel fire intelligent decision support device comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

10. A storage medium having stored thereon a program, characterized by When the program is executed, it implements the method as described in any one of claims 1-8.

Images