Electrical fire scene information map construction method and device, electronic equipment and medium

Abstract

The application discloses an electrical fire scene information graph construction method, device, electronic equipment and medium related to the technical field of electrical fire scene information graph construction methods, and the method comprises the steps of: based on the pre-defined electrical fire domain ontology constraint, using a large language model to construct a local subgraph corresponding to each investigation evidence file; based on the semantic consistency of the entity nodes in the local subgraph and the context consistency on the graph structure, and in combination with the abstract text, the entity nodes associated across the subgraphs are identified, an associated dataset is formed, the associated dataset is merged with different local subgraphs, and a global fusion graph is generated; based on the global fusion graph, the entity relationship structure associated across multiple local subgraphs is identified, and a hot relationship cluster is determined; and according to the hot relationship cluster and the global fusion graph, an information graph for the electrical fire investigation scene is constructed. The method can convert multi-source investigation evidence into a semantically consistent knowledge structure, and improve the efficiency of fact restoration and correlation analysis in the electrical fire investigation process.

Description

Technical Field

[0001] This invention relates to the technical field of electrical fire scene information map construction methods, and in particular to an electrical fire scene information map construction method, device, electronic device and medium. Background Technology

[0002] Electrical fires, as a significant type of disaster threatening public safety and the lives and property of people and property, typically involve the coupling and evolution of multiple factors, resulting in complex, dynamic, and fragmented information. After a fire, relevant information is often scattered across multiple heterogeneous sources, including accident investigation reports, on-site inspection records, monitoring and log data, meteorological information, building structure records, lists of combustible materials, and witness testimonies. These different sources exhibit significant differences in information organization, granularity of expression, and use of technical terminology. This information not only contains numerous technical terms but also implicitly contains complex causal relationships, temporal evolutionary relationships, and regulatory constraints. While crucial for fire cause analysis, accident liability determination, and the formulation of prevention and control measures, it also significantly increases the difficulty of information understanding and comprehensive analysis.

[0003] The processing and utilization of electrical fire accident information primarily relies on manual analysis or expert systems based on experience-based rules. Accident analysis typically involves manually reviewing multiple source documents, extracting key elements, and tracing the accident's evolution chain. This approach is highly dependent on the investigators' professional experience and subjective judgment, resulting in low processing efficiency, difficulty meeting timeliness requirements in complex accident scenarios, and a tendency to miss information and fail to adequately identify correlations when dealing with large-scale, multi-dimensional data. Furthermore, the manually processed analysis results are often stored in unstructured or semi-structured formats, making it difficult to form a unified and reusable knowledge representation, and hindering the systematic inheritance and utilization of historical accident experience in subsequent investigations.

[0004] With the development of artificial intelligence technology, especially the improved capabilities of large language models in text understanding, information extraction, and summarization, these technologies have begun to be applied to text analysis tasks across multiple domains. However, general natural language processing techniques and large language models typically lack the ability to deeply model the proprietary conceptual systems and accident evolution logic of the fire investigation field. When dealing with information in fire scenarios that exhibits strong causal, temporal, and physical constraints, they still struggle to accurately express the intrinsic relationships between entities. Furthermore, large language models are limited by context length and attention allocation mechanisms when handling long contexts and multi-document scenarios, easily disrupting the consistency and integrity of cross-document information. This results in analysis results that fail to meet the requirements of electrical fire investigations for structured expression, logical rigor, and the completeness of the evidence chain. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the purpose of this invention is to propose a method, apparatus, electronic device, and medium for constructing an electrical fire scene information map, so as to achieve automated modeling and unified expression of entities and their relationships in multi-source electrical fire investigation information. To achieve the above objectives, a first aspect of the present invention proposes a method for constructing an electrical fire scene information map, comprising: Based on predefined ontology constraints in the field of electrical fires, a large language model is used to perform semantic understanding and structured extraction on multiple electrical fire investigation evidence documents, and a local subgraph corresponding to each of the investigation evidence documents is constructed. The local subgraph includes entity nodes, entity attributes describing the characteristics of the entity nodes, relation edges representing the interaction logic between entities, and summary text generated by combining the factual information of the entity nodes or relation edges. Based on the semantic consistency of entity nodes in the local subgraphs and the contextual consistency of the graph structure, and combined with the summary text, entity nodes that are related across subgraphs are identified to form a related dataset. The related dataset is then merged with different local subgraphs to generate a global fusion graph representing the relationship between multiple investigation evidence documents. Based on the graph structure features, entity semantic features, and relation connection patterns in the global fusion graph, the entity relation structure that is related across multiple local subgraphs is identified, and hotspot relation clusters are determined. Based on the hotspot relationship clusters and the global fusion graph, an information graph for electrical fire investigation scenarios is constructed.

[0006] In addition, the method of the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, based on the semantic consistency of entity nodes in the local subgraph and the contextual consistency of the graph structure, and in conjunction with the summary text, entity nodes that are related across subgraphs are identified to form an association dataset, including: For each entity node in the local subgraph, the standardized name, entity attributes, and corresponding summary text of the entity node are extracted to construct a triple for a structured description of the entity node. The triples of the entity node are transformed into the semantic feature vectors of the entity node; For any two candidate entity node pairs consisting of entity nodes from different local subgraphs, the semantic consistency and graph structure context consistency of the candidate entity node pairs are calculated based on the semantic feature vectors, and the associated dataset is determined based on the calculation results.

[0007] According to one embodiment of the present invention, the process for determining the semantic consistency of entity nodes includes: Based on the semantic feature vectors corresponding to the candidate entity node pairs, calculate the cosine similarity of the candidate entity node pairs; The cosine similarity is weighted and corrected using the attribute adjacency coefficient to obtain the final semantic similarity; wherein, the attribute adjacency coefficient represents the semantic affinity between the candidate entity nodes and their respective attributes; If the semantic similarity is greater than or equal to the first preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

[0008] According to an embodiment of the present invention, the process for determining the contextual consistency of a graph structure includes: For each entity node in the candidate entity node pair, extract the neighborhood relationship set of the entity node in its respective local subgraph; the neighborhood relationship set includes the neighboring entity nodes of the entity node, and the relationship type between the entity node and its neighboring entity nodes; Semantic vectorization representation is performed on the adjacent entity nodes involved in the neighborhood relationship set to obtain the adjacent entity semantic embedding vector used to characterize the semantic features of the neighborhood relationship; The context similarity of the candidate entity node pair is calculated based on the semantic embedding vector of the adjacent entity corresponding to each entity node in the candidate entity node pair. If the context similarity is greater than or equal to the second preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

[0009] According to one embodiment of the present invention, the method further includes: For entity node pairs that are related in the associated dataset, the summary text of the entity node pairs in the local subgraph is fused, and a large language model is called to generate summary text of cross-subgraph association; The generated summary text is combined with the corresponding entity nodes and their relationships to form a structured relational unit for characterizing cross-subgraph relationships.

[0010] According to one embodiment of the present invention, the determination of hotspot relationship clusters includes at least one of the following: Based on the neighborhood connectivity of entity nodes in the global fusion graph, the neighborhood interaction strength of each entity node is statistically analyzed, and the entity nodes whose neighborhood interaction strength reaches the third preset threshold and their corresponding neighborhood sets are determined as topological density hotspot clusters. The attribute dimensions of each entity node are weighted based on a preset attribute weight function, the neighborhood semantic density of each node is calculated, and the entity nodes whose neighborhood semantic density reaches the fourth preset threshold and their corresponding neighborhood sets are identified as semantic density hotspot clusters. Relationship edges that meet the preset association type are selected from the global fusion graph, and entity relationship paths that meet the preset path conditions are searched in the inference subgraph constructed based on the relationship edges. The entity relationship paths are determined as hotspot relationship clusters of the inference chain.

[0011] According to one embodiment of the present invention, constructing an information graph for electrical fire investigation scenarios based on the hotspot relationship clusters and the global fusion graph includes: Obtain the hotspot type corresponding to the hotspot relationship cluster, as well as the entity nodes, relationship edges, and entity relationship path information contained therein; The large language model is invoked to generate hotspot information keywords and summary text describing the hotspot relationship clusters based on the hotspot type and the structural features and semantic content of the hotspot relationship clusters. The hot topic information keywords and summary text generated by the hot topic relationship clusters are structured and organized to construct a hot topic summary index structure; The hotspot summary index structure and the global fusion graph are used as the output information graph for electrical fire investigation scenarios.

[0012] To achieve the above objectives, a second aspect of the present invention provides an electrical fire scene information map construction device, comprising: The local subgraph construction module is used to construct a local subgraph corresponding to each of the obtained electrical fire investigation evidence documents by performing semantic understanding and structured extraction on multiple obtained electrical fire investigation evidence documents based on predefined electrical fire domain ontology constraints and using a large language model. The local subgraph includes entity nodes, entity attributes describing the characteristics of the entity nodes, relation edges representing the interaction logic between entities, and summary text generated by combining the factual information of the entity nodes or relation edges. The cross-subgraph association identification module is used to identify entity nodes that are associated across subgraphs based on the semantic consistency of entity nodes in the local subgraphs and the contextual consistency of the graph structure, and in combination with the summary text, to form an association dataset, and to merge the association dataset with different local subgraphs to generate a global fusion graph representing the association relationship of multiple investigation evidence documents. The hotspot relationship cluster determination module is used to identify the entity relationship structure that is related across multiple local subgraphs and determine the hotspot relationship clusters based on the graph structure features, entity semantic features and relationship connection patterns in the global fusion graph. The information graph construction module is used to construct an information graph for electrical fire investigation scenarios based on the hotspot relationship clusters and the global fusion graph.

[0013] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for constructing an electrical fire scene information map.

[0014] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed, implements the steps of the above-described method for constructing an electrical fire scene information map.

[0015] The electrical fire scene information graph construction method, apparatus, electronic device, and medium of this invention introduce a semantic understanding and structured extraction mechanism based on domain ontology constraints in electrical fire investigation scenarios. This maps key information scattered across multiple investigation evidence files into a unified local subgraph representation. Furthermore, based on entity semantic consistency and graph structure context consistency, it achieves cross-subgraph entity association and fusion, thereby constructing a global fusion graph that reflects the inherent connections between multiple sources of evidence. On this basis, through analysis and clustering of entity relationship structures in the fusion graph, it can identify hotspot relationship clusters with investigative value, making key relationships in complex accident information more concentrated and prominent. The resulting information graph integrates entities, relationships, and their semantic summaries in a structured manner, helping to improve the efficiency of understanding accident elements and their associated logic and the systematic nature of overall analysis during electrical fire investigations. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for constructing an electrical fire scene information map in one embodiment; Figure 2 This is a schematic diagram illustrating the definition of the domain map ontology for an electrical fire scenario in one embodiment; Figure 3 This is a schematic diagram of a partial sub-graph output for an electrical fire scenario in one embodiment; Figure 4 This is a schematic diagram of the process for forming an associated dataset in one embodiment; Figure 5 This is a schematic diagram of the semantic consistency determination process for entity nodes in one embodiment; Figure 6 This is a schematic diagram of semantic-based entity consistency recognition in one embodiment; Figure 7 This is a schematic diagram of the process for determining the contextual consistency of a graph structure in one embodiment; Figure 8 This is a schematic diagram of graph-based context consistency identification in one embodiment; Figure 9 This is a flowchart illustrating the process of generating a cross-subgraph association summary in one embodiment; Figure 10 This is a schematic diagram of a cross-subgraph association summary in one embodiment; Figure 11 This is a schematic diagram illustrating the construction of hotspot relationship clusters with inference chain step size in one embodiment; Figure 12 This is a flowchart illustrating the process of constructing an information graph for an electrical fire investigation scenario in one embodiment. Figure 13 This is a structural block diagram of an electrical fire scene information map construction device in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] The implementation details of the technical solutions of the embodiments of the present invention are described in detail below.

[0019] In one embodiment, such as Figure 1 The diagram illustrates a method for constructing an information map of an electrical fire scene. This method may include the following steps: Step S101: Based on predefined ontology constraints in the electrical fire domain, a large language model is used to perform semantic understanding and structured extraction on the multiple electrical fire investigation evidence documents obtained, and a local subgraph corresponding to each investigation evidence document is constructed.

[0020] The construction of local subgraphs uses electrical fire investigation evidence documents as the core input object, and comprehensively introduces the electrical fire domain graph ontology, electrical fire entity and relationship type system, and summary generation specifications for fire investigation facts in the process, so as to realize the automatic conversion from unstructured evidence text to structured semantic subgraphs.

[0021] In the field of electrical fire investigations, ontology constraints are used to uniformly define the entity attribute system and the types of relationships between entities involved in electrical fire investigations. This allows investigative evidence documents from different sources and in different forms to be interpreted and expressed within the same semantic framework. In practical applications, such as... Figure 2The diagram illustrates the definition of the domain ontology for an electrical fire scenario. A domain ontology is pre-constructed for electrical fire investigation scenarios to constrain subsequent semantic understanding and information extraction processes. Based on guiding theories in areas such as electrical fire investigation standards, electrical fire events can be decomposed into elements to construct an entity attribute system for the electrical fire information graph. This attribute system categorizes all entities involved in electrical fire investigations into nine attribute dimensions: personnel or units, equipment, combustible materials, building components, environmental conditions, management behavior, event process, regulations and standards, and consequences.

[0022] Each attribute has a clear semantic boundary and scope of application. All fire-related entity nodes can be classified into one and only one attribute. The specific definitions of the attributes are shown in Table 1.

[0023] Table 1

[0024] This attribute system ensures that each entity extracted from the investigation evidence files has a unique and clear attribute attribution, thus avoiding entity type confusion at the semantic level.

[0025] Based on this, such as Figure 2 As shown, an entity and relation type system is constructed for electrical fire investigation scenarios, serving as the semantic skeleton for knowledge graph construction. This system consists of two parts: entity specifications and relation specifications.

[0026] Regarding entity specifications, entities are the basic nodes of the graph, representing objects or concepts with independent semantics in an electrical fire scenario. Each entity node is required to include at least a standardized name after terminology normalization and its corresponding element attribute category (strictly limited to one of E01 to E09 in Table 1 above), thereby ensuring a consistent representation of semantically equivalent entities from different sources of investigative evidence at the graph level. For example, the entity "kitchen distribution box" belongs to E02 (electronic equipment). The standardized name is determined through synonym merging, professional terminology mapping, and domain-specific thesaurus constraints to eliminate entity ambiguities caused by differences in expression in different evidence materials; for example, "air switch" is standardized as "circuit breaker".

[0027] In terms of relation specification, relations are used to describe specific semantic associations between two entities that conform to the evolutionary patterns of electrical fires. During the local subgraph construction phase, undirected relations are used to represent the associations between entities, prioritizing the establishment of semantic connections without presupposing causal directions or order of action. This adapts to the characteristics of investigations where the facts are not yet fully closed in the early stages of multi-source evidence fusion. Through relation specification, various relation types are predefined, including factual associations, spatial associations, logical associations, temporal associations, and co-occurrence associations. These are used to describe structural configuration relationships, spatial location binding relationships, abstract connections formed by professional reasoning, temporal relationships during event evolution, and weak associations formed solely based on evidence co-occurrence, respectively.

[0028] In this context, factual association refers to a relatively static structural or configurational relationship that existed objectively before the accident occurred. This type of relationship does not depend on the evolution of the event but reflects the objective connection between entities in terms of spatial layout, functional dependence, operational affiliation, or management binding. For example, there is a factual association between electrical wires and curtains due to spatial proximity and potential contact risk; between the on-duty electrician and the distribution box due to operational and management affiliation; and between cluttered items and fire escape routes due to obstruction of the passageway.

[0029] Spatial correlation is used to represent the binding relationship between the location of an event or the distribution of accident consequences and a specific physical space. It is mainly used to anchor the evolution of electrical faults or fires to a specific spatial location. For example, spatial correlation between a short circuit event and a cable shaft indicates the location of the fault; spatial correlation between a structural collapse and a suspended ceiling mezzanine limits the area affected by the accident; and spatial correlation between personal injury consequences and specific floors or rooms determines the location of the outcome.

[0030] Logical associations are used to represent abstract semantic connections that cannot be directly determined spatially or temporally, but require specialized knowledge, regulatory comparison, or inferential analysis to establish. These relationships are typically used to express conclusions such as causal implications, compliance judgments, liability attribution, or impact assessments. They have high semantic strength but rely on context for interpretation and verification. For example, a logical association can reflect a potential causal relationship between infrequent maintenance and electric arcing; a logical association can reflect a compliance judgment between insufficient installation spacing of distribution boxes and relevant national standard provisions; and a logical association can reflect the impact assessment results between the flame spread process and direct economic losses.

[0031] Temporal correlation is used to describe the temporal sequence, concurrent relationship, or duration relationship between sub-events during the evolution of electrical faults and fires, in order to characterize the dynamic process of accident development. For example, temporal correlation reflects the sequential or causal triggering order between electric arc and short circuit, and temporal correlation reflects the stages of fire development between smoke and flame spread.

[0032] Furthermore, to enhance the connectivity of the information graph during the initial construction phase, co-occurrence associations are introduced as supplementary weak associations. Co-occurrence associations represent situations where two entities are jointly mentioned in the same structured evidence unit and exist in the same semantic context, but cannot yet be clearly categorized as factual, spatial, logical, or temporal associations. This type of relationship does not presuppose an objective or causal connection between the entities; it is only used to retain co-occurrence information from the original evidence. For example, a witness and a flame co-occur because they both appear in the same interrogation record; thermal insulation and company procedures co-occur because they are mentioned in the same rectification notice. Subsequently, co-occurrence associations can be retained, strengthened, or removed by combining multi-source evidence cross-validation, semantic enhancement, or conflict resolution mechanisms.

[0033] After completing the unified definition of entity attribute specifications and relation type specifications, to ensure that the information extracted from the investigation evidence can be expressed in an objective, verifiable form with investigative semantics in the graph, such as... Figure 2 As shown, a summary generation specification for electrical fire investigation scenarios is further pre-constructed to constrain the semantic representation of subsequent entities and relational facts.

[0034] The abstract generation guidelines define the generation objectives, semantic scope, and expression requirements of the abstract text. The abstract text should serve the purpose of fact-finding and evidence presentation in fire accident investigations, and must not introduce subjective judgments, speculative conclusions, or causal inferences unsupported by evidence. The abstract generation process follows the principles of complete elements, standardized terminology, clear semantics, and consistent time sequence to ensure consistency in semantic granularity and expression style across abstracts generated from different evidence sources.

[0035] In terms of specific format, summary texts include two categories: entity summaries and relation summaries. Entity summaries describe the objective state of a single entity, covering its technical parameters, spatial location, structural features, management background, or static attributes related to the accident. For example, for the entity "kitchen electrical distribution box," the following summary text can be generated: This distribution box is installed on the north wall of the first-floor kitchen, contains 6 outgoing lines, has a main switch rated current of 40A, the box has no protective cover, and the interior is heavily dusty. Relationship summaries describe the objective connection or interaction between two entities, focusing on the factual role of this relationship in the evolution of an electrical fault or fire. For example, for the relationship "electrical wire (E02) — factual association — curtain (E03)," the following summary text can be generated: Before the fire, the kitchen electrical wiring was laid close to the cotton curtain, with a gap of less than 5 centimeters.

[0036] By pre-defining the summary generation specifications, the summary text subsequently generated by the large language model can remain consistent with the facts of the electrical fire investigation at the semantic level, and be comparable and integrable between different local subgraphs, laying the foundation for cross-evidence association analysis and global graph construction.

[0037] After completing the unified definition of the domain ontology specification, entity and relation type specification, and summary generation specification, we began to process the multiple electrical fire investigation evidence documents to construct local subgraphs corresponding to each investigation evidence document.

[0038] When processing each piece of investigative evidence, the data type of the evidence document is first identified, and corresponding parsing processing is performed according to its specific format. Specifically, by identifying the source of the evidence, file format, and data characteristics, different types of evidence documents are automatically matched to the corresponding data parsing process to achieve unified access and processing of multi-source evidence. For text documents, encoding standardization, paragraph reorganization, and noise filtering are performed; for scanned documents or images, optical character recognition technology is used to extract text content, and structural information such as titles and legends is preserved in conjunction with layout analysis; for audio and video materials, speech recognition technology is used to convert them into text with timestamps; for structured data such as tables or logs, field alignment and unit standardization are performed. After the above processing, standardized text content that can be used for unified semantic analysis is formed.

[0039] Subsequently, using the ontology of the electrical fire domain as a semantic constraint framework, a large language model was employed to segment standardized text content into semantic fragments. For each semantic fragment, a two-stage structured extraction process was performed, including entity node identification and undirected relation reasoning. In the entity node identification process, the large language model identified entities from the text based on the semantic definitions of nine element attributes and labeled them with their corresponding attribute categories. In the relation reasoning process, based on the identified entities and combined with pre-defined relation rules, it determined whether there were factual, spatial, logical, temporal, or co-occurrence relationships between entities and generated corresponding relation records.

[0040] After the structured extraction of entities and relationships is completed, the large language model is guided to generate corresponding summary texts for the identified entity nodes and the relationships between entities, based on the aforementioned summary generation specifications. These summary texts are used to semantically refine and supplement the objective facts reflected in the investigation evidence. The generated summary texts are used to describe the objective state of entities or the objective connections between entities, and do not contain conclusive judgments or liability determinations.

[0041] In practical applications, such as Figure 3 The diagram shows a partial subgraph output of an electrical fire scenario. The aforementioned entity recognition, relation reasoning, and summary generation processes all guide the large language model to complete the corresponding semantic analysis and text generation tasks according to predetermined rules by constructing structured prompts that include ontology constraints, task descriptions, and contextual examples. This achieves the automated conversion from unstructured text to semantically coherent and logically closed-loop knowledge graph fragments.

[0042] Finally, the entity nodes, entity attributes, relation edges, and corresponding summary texts extracted and generated from the same investigation evidence file are uniformly organized to construct a local subgraph corresponding to that investigation evidence file. In the local subgraph, entity nodes are used as the basic units, and relation edges connect related entities. The generated summaries are then appended to the corresponding entity nodes and relation edges, thereby fully expressing the electrical fire facts and their inherent semantic structure reflected in the investigation evidence. This provides basic data support for subsequent multi-source evidence fusion, global graph construction, and accident analysis and reasoning.

[0043] Step S102: Based on the semantic consistency of entity nodes in the local subgraphs and the contextual consistency of the graph structure, and combined with the summary text, identify entity nodes that are related across subgraphs, form a related dataset, and merge the related dataset with different local subgraphs to generate a global fusion graph representing the relationship between multiple investigation evidence documents.

[0044] After constructing the local subgraphs corresponding to each investigation evidence document, and while preserving the independence of each local subgraph, the entity nodes and their relationships in each local subgraph are analyzed to identify entity nodes in different local subgraphs that may point to the same real-world object or event. Specifically, by comparing the semantic features of entity nodes and their contextual relationships within their respective local subgraphs, potential associations between different subgraphs can be determined, and these cross-subgraph related entity nodes are organized into an association data set.

[0045] The cross-subgraph entity relationships recorded in the associated dataset are unified and integrated with the entity nodes and relationships in each local subgraph to construct a global fusion graph. Among them, the node set It consists of entity nodes from different local subgraphs, each entity node containing a corresponding name, attribute category, and summary information; edge set Used to represent the connections between entity nodes, each relation edge is labeled with the corresponding relation type, and can be further associated with relation summary information to describe the semantic background of the relation in the investigative evidence.

[0046] Through the above merging process, the entity information and relationship information that were originally scattered in different investigation evidence documents are centrally expressed in a unified fusion diagram, thereby forming a global fusion diagram that can comprehensively reflect the entity relationship between multiple investigation evidence documents.

[0047] In one embodiment, Figure 4 A flowchart for creating a related dataset may include the following steps: Step S201: For each entity node in a local subgraph, extract the standardized name, entity attributes, and corresponding summary text of the entity node, and construct a triplet for a structured description of the entity node.

[0048] For each entity node The standardized name, attribute category, and summary of each entity node are concatenated to form a unified descriptive unit, constituting a triple, which is used to explicitly represent the semantic information and context of the entity node. This triple can be represented as:

[0049] in, For entity nodes Standardized name; For entity nodes Belonging attribute dimension; For entity nodes Summary text.

[0050] Step S202: Convert the triples of entity nodes into semantic feature vectors of entity nodes.

[0051] The generated triples are processed using semantic feature representation to convert textual information into dense semantic vectors. This semantic representation can capture key information features in entity names, attribute categories, and summaries, providing a basis for semantic comparison of entity nodes in different local subgraphs.

[0052] Specifically, the triplet Combine them into a structured prompt text:

[0053] Using semantic vector models Perform vectorization to generate dense vectors:

[0054] in, Represents entity nodes The semantic feature vector is used to characterize the semantic features of entity nodes, including name, attributes, and summary information.

[0055] Step S203: For any two candidate entity node pairs consisting of entity nodes from different local subgraphs, calculate the semantic consistency and graph structure context consistency of the candidate entity node pairs based on semantic feature vectors, and determine the associated dataset based on the calculation results.

[0056] Based on the semantic feature vectors corresponding to the obtained entity nodes, entity nodes from different local subgraphs are paired to form candidate entity node pairs. For each candidate node pair, association determination is performed based on its semantic feature vector. Combined with the structural context information of the nodes in their respective local subgraphs, potential cross-subgraph correspondences are identified, and the identified associated entity nodes are compiled into an association dataset. This association dataset not only records the correspondences between entity nodes but also retains the attributes and summary information of each node to maintain the integrity and traceability of multi-source evidence in subsequent analysis.

[0057] In one embodiment, Figure 5 The diagram illustrates the semantic consistency determination process for entity nodes, which may include the following steps: Step S301: Calculate the cosine similarity of candidate entity node pairs based on the semantic feature vectors corresponding to the candidate entity node pairs.

[0058] Figure 6 This diagram illustrates semantic-based entity consistency recognition, where the induction cooker entity node is shown. The solid node of the wok is By judging the semantic consistency of these two entity nodes, it is possible to confirm whether there is a relationship between them. For any candidate entity node pair from different local subgraphs... and Obtain the corresponding semantic feature vectors respectively. and The cosine similarity between vectors is used to calculate the degree of closeness between them in the semantic space. The cosine similarity measures the similarity between two entity nodes in terms of comprehensive semantic information such as name, attribute category and summary text. Its value ranges from [0, 1], and the larger the value, the higher the degree of semantic similarity.

[0059] Step S302: Use the attribute adjacency coefficient to perform weighted correction on the cosine similarity to obtain the final semantic similarity.

[0060] To avoid distortion in similarity assessment between entities of different attribute categories due to relying solely on textual semantics, an attribute adjacency coefficient is introduced into the semantic similarity calculation process. This is used to characterize the semantic association strength between the attribute dimensions of two entities in a candidate entity node pair, and is specifically defined as follows:

[0061] in, and These represent the attribute dimensions of the two objects respectively; when When , it indicates that the two objects belong to the same attribute dimension, and the attribute adjacency coefficient is 1.0; when Belongs to a predefined set of dimensional adjacency relationships When the adjacency coefficient is 0.6, it indicates that two different attributes have a factual coupling or descriptive intertwining relationship in an electrical fire investigation scenario. This is used to describe combinations of attribute dimensions that are frequently mentioned in electrical fire accidents, such as management behavior with attribute dimension E06 and event process with attribute dimension E07. In other cases, the attribute adjacency coefficient is 0.2.

[0062] After introducing the attribute adjacency coefficient, the final semantic similarity of candidate entity node pairs It can be represented as:

[0063] In the above formula, Indicates cosine similarity; Indicating candidate entity node pairs and The adjacency coefficient between attributes.

[0064] By using the weighting method described above, while preserving semantic vector similarity information, the semantic similarity results can reflect the actual semantic association between different attribute dimensions, thereby improving the rationality of cross-attribute entity semantic consistency determination.

[0065] Step S303: If the semantic similarity is greater than or equal to the first preset threshold, the corresponding candidate entity node pair is stored in the associated dataset.

[0066] Set the first preset threshold Its value range is For example, it can be set When candidate entity nodes pair final semantic similarity The candidate entity node pair is determined to be semantically consistent, and the candidate entity node pair is then set as a separate entity node pair. Add it to the associated dataset.

[0067] In one embodiment, Figure 7 This diagram illustrates the process for determining the context consistency of a graph structure, which may include the following steps: Step S401: For each entity node in the candidate entity node pair, extract the set of neighborhood relations of the entity node in its respective local subgraph.

[0068] For any entity node in the candidate entity node pair From its local subgraph, all relationships centered on the entity node are extracted to form a neighborhood relationship set, which represents the direct structural context of the entity node in the local subgraph. The neighborhood relationship set includes relationships with the entity node... There are directly related adjacent entity nodes. and entity nodes With adjacent entity nodes The corresponding relationship types between entities are used to characterize different semantic relationships such as factual association, spatial association, logical association, temporal association, or co-occurrence association.

[0069] In practical applications, neighborhood relation sets can be represented in the form of relation triples to uniformly describe the local topological structure characteristics of entity nodes, specifically as follows:

[0070] in, Represents entity nodes The set of neighborhood relations in its local subgraph;

[0071] Indicates the presence of entity nodes Local subgraphs; Represents entity nodes With adjacent entity nodes Relationship types; Represents entity nodes Directly connected adjacent entity nodes.

[0072] For example, such as Figure 8 The diagram shown is based on the context consistency recognition of the graph structure. In the left subgraph, the adjacent nodes directly associated with the entity node "AL1 power distribution circuit start point" are extracted, such as "north wall of the kitchen", "3 socket circuits" and "cable". In the right subgraph, the adjacent nodes associated with the entity node "smoke point on the north wall" are extracted, such as "burnt smell", "burnt wire" and "blackened wall".

[0073] Step S402: Semantically vectorize the neighboring entity nodes involved in the neighborhood relationship set to obtain the neighboring entity semantic embedding vector used to characterize the semantic features of the neighborhood relationship.

[0074] For entity nodes Each adjacent entity node Perform semantic vectorization representation. Specifically, represent adjacent entity nodes... The attribute dimension information, standardized names, and summary text are combined to form a structured input for semantic representation, and then vectorized using a semantic vector model to obtain semantic embedding vectors of adjacent entities. It is used to characterize the semantic attribute category features, entity semantic features and their descriptive information in the local scene of adjacent entity nodes, thereby reflecting the semantic role and semantic contribution of the adjacent entity in the neighborhood of the current entity node.

[0075] Among them, the semantic embedding vectors of adjacent entities Specifically, it can be expressed as:

[0076] In the above formula, Representing adjacent entity nodes Belonging attribute dimension; Representing adjacent entity nodes Standardized name; Representing adjacent entity nodes Abstract text; This represents a semantic encoding function used to map structured semantic descriptions to dense vector representations.

[0077] By performing semantic vectorization on all adjacent entity nodes in the neighborhood, entity nodes can be constructed. The neighborhood embedding representation is used to transform its local graph structure context into a computable set of semantic features for subsequent context consistency analysis.

[0078] Step S403: Calculate the context similarity of the candidate entity node pair based on the semantic embedding vector of the adjacent entity corresponding to each entity node in the candidate entity node pair.

[0079] For candidate entity node pairs from different local subgraphs Obtain entity nodes respectively and The neighborhood relationship set is defined, and each neighborhood set is regarded as a semantic bag consisting of semantic embedding vectors of adjacent entities.

[0080] Based on this, entity nodes are calculated by pairwise comparison of the embedding vectors of adjacent entities in two neighboring semantic bags. and Similarity at the graph structure context level. For example... Figure 8 As shown, this process specifically involves quantifying and comparing the semantic matching degree between nodes such as "north wall of the kitchen" and "burnt smell", "cable" and "burnt wire", aiming to measure the overall semantic similarity of the graph structure environment in which the two central nodes are located.

[0081] In practical applications, context similarity can be represented as:

[0082] in, Used to measure entity nodes and The overall similarity between the local graph structure and the semantic environment of the neighborhood. Represents entity nodes The set of neighborhood relations; Represents entity nodes The set of neighborhood relations; Cosine similarity, based on the angle between vectors, is used to measure the similarity between the semantic embedding vectors of two adjacent entities in the semantic space.

[0083] In practical applications, to reduce the impact of noisy neighbors on the similarity calculation results and improve computational efficiency and robustness, the maximum matching mean method can be used to optimize the above calculation process. Its context similarity can be expressed as:

[0084] This approach focuses on the semantically most matching entity pairs in the neighborhood, avoiding a small number of irrelevant neighbors from lowering the overall similarity, thus more accurately reflecting the similarity of entity nodes at the graph structure context level.

[0085] Step S404: If the context similarity is greater than or equal to the second preset threshold, the corresponding candidate entity node pair is stored in the associated dataset.

[0086] Set a second preset threshold Its value range is For example, it can be set When candidate entity nodes pair final semantic similarity The candidate entity node pair is determined to be consistent at the graph structure context level, and the candidate entity node pair is then added to the list. Store it in the associated dataset.

[0087] For example, refer to Figure 8 Although the entity nodes "AL1 power distribution circuit start point" and "north wall smoke point" differ in their literal names, they both contain highly related spatial locations (north wall of the kitchen) and physical carriers (cables and wires) within their neighborhoods. Based on the strong correlation of neighborhood features, the candidate entity node pair is determined to have consistency at the graph structure level, thus achieving accurate alignment of entities across subgraphs.

[0088] In one embodiment, Figure 9The diagram illustrates a process for generating a summary of cross-subgraph associations, which may include the following steps: Step S501: For entity node pairs with related relationships in the associated dataset, merge the summary text of the entity node pairs in the local subgraph, and call the large language model to generate summary text of cross-subgraph association.

[0089] Related datasets This is used to reflect the entity relationships in different local subgraphs that point to the same real-world object or event. In practical applications, it relates to datasets. For semantically consistent entity sets With structural consistency set The union of, i.e. .

[0090] Among them, semantically consistent entity set This represents the set of entity node pairs that are related based on semantic consistency; the structural consistency set. This represents the set of entity node pairs that are related based on the context consistency determination of the graph structure.

[0091] For related datasets The system identifies pairs of entity nodes with related relationships and integrates the summary text of these pairs within their respective local subgraphs as input for cross-subgraph semantic integration. By jointly analyzing the summary texts from different local subgraphs, different descriptions of the same object or event from multi-source evidence can be aligned and synthesized within a unified context.

[0092] Based on the fused summary text, a large language model is invoked to perform joint semantic analysis on entity node pairs, generating summary texts of cross-subgraph associations. The generated cross-subgraph association summary texts are used to objectively describe the cooperative relationships, corroborative relationships, or potential conflicts of the entity node pairs in multi-source evidence, thereby revealing the implicit causal chains, factual correspondences, or semantic differences between different local subgraphs.

[0093] The generated cross-subgraph association summary text is used to objectively describe the collaborative relationship, corroborative relationship or potential conflict of the entity node pair in multi-source evidence, thereby revealing the implicit causal chain, factual correspondence or semantic differences between different local subgraphs.

[0094] Figure 10This diagram illustrates a cross-subgraph association summary in a specific example. The left subgraph records the starting point of the AL1 power distribution circuit and its topological connection to the north wall of the kitchen. The right subgraph details the point of smoke emanating from the north wall, as well as the objective facts found on site, such as burned wires and blackened walls. After identifying the inherent spatiotemporal connection between the two subgraphs, a large language model is invoked for deep joint semantic analysis to generate a cross-subgraph association summary. This summary objectively describes the corroborative relationships or logical conflicts between multiple sources of evidence. For example, the summary text may indicate that the associated location is a potential fault point, thereby revealing the causal chains and factual correspondences hidden between different subgraphs.

[0095] Step S502: Combine the generated summary text with the corresponding entity nodes and their relationships to form a structured relational unit for characterizing cross-subgraph relationships.

[0096] The generated cross-subgraph association summary text is combined with the corresponding entity nodes and the association relationships between entity nodes to form a structured relationship unit for characterizing cross-subgraph association relationships. This structured relationship unit is then incorporated into the electrical fire scene information graph to support subsequent analysis, reasoning, and interpretable display.

[0097] Step S103: Based on the graph structure features, entity semantic features and relation connection patterns in the global fusion graph, identify the entity relation structure that is related across multiple local subgraphs and determine the hotspot relation clusters.

[0098] After representing entities and their relationships in multiple investigative evidence documents using a global fusion graph, the graph structure features of entity nodes in the fusion graph are analyzed to characterize the degree of connection and structural position characteristics of entity nodes in the overall network. At the same time, based on the attribute categories, semantic descriptions, and contexts of entity nodes in different evidence documents, the semantic features of entities are comprehensively analyzed to reflect the semantic information features contained in entities and their neighborhood relationships.

[0099] Furthermore, by analyzing the combinations of relationship types between entity nodes and the connection path patterns they form in the fusion graph, the connection patterns between entities are identified to reveal the organizational characteristics of entity relationships at the global level. Through joint analysis of graph structural features, entity semantic features, and relationship connection patterns, it becomes possible to identify entity relationship structures that exhibit significant correlation characteristics in multi-source evidence from the global fusion graph.

[0100] Based on this, entity relationship structures that exhibit significant characteristics at the structural, semantic, or relational organization levels are identified as hotspot relationship clusters. These hotspot relationship clusters are used to centrally represent sets of entity relationships that are frequently associated, semantically dense, or have significant indicative meaning for event analysis across multiple investigative evidence documents. This enables rapid location of key factual areas and provides support for subsequent information understanding, evidence analysis, and intelligent reasoning.

[0101] In one embodiment, after completing the construction of local subgraphs and cross-subgraph entity associations, and generating a global fusion graph, the entity nodes and their relationship structures in the global fusion graph are analyzed to identify hotspot relationship clusters with outstanding analytical value. These hotspot relationship clusters characterize entity relationship regions that exhibit significant association features in multi-source investigation evidence documents. They can be categorized into three types: topologically interactive, semantically element-intensive, and reasoning path-significant, used to highlight key fact areas from different analytical perspectives.

[0102] For each entity node in the global fusion graph, its neighborhood connectivity is statistically analyzed. By calculating the number of directly connected neighboring entity nodes, the neighborhood interaction strength, which characterizes the local connectivity density of the entity node, is obtained. Specifically, it can be expressed as:

[0103] In the above formula, Represents any entity node in the global fusion graph. Represents entity nodes Adjacent entity nodes directly connected by relation edges This represents the set of relation edges in the global fusion graph. Used to characterize the neighborhood interaction strength of an entity node in the local region of the fusion graph.

[0104] When the neighborhood interaction strength Reaching the third preset threshold (The default value is 5), indicating that the entity node is in a position where it is repeatedly associated or frequently cited in multiple investigation evidence documents. Centered on this entity node, neighboring entity nodes and their corresponding relationships are selected within its neighborhood to form a topological density hotspot cluster. This is used to reflect areas of entity relationships that are frequently mentioned or highly related in multi-source investigation evidence.

[0105] Building upon topology density analysis, a semantic density analysis mechanism is introduced for entity nodes in the global fusion graph to further identify entity relationship regions with significant semantic analytical value. Specifically, based on a preset attribute weighting function, different attribute dimensions of entity nodes are weighted to reflect the relative importance differences of different types of entities in electrical fire investigation and analysis. The attribute weighting function, used to distinguish the importance of different attribute types in the analysis, can be specifically defined as follows:

[0106] For each entity node Based on the attribute distribution of adjacent entities within the neighborhood of an entity node, the neighborhood semantic density of the entity node is calculated, which can be specifically expressed as:

[0107] In the above formula, Represents entity nodes Neighborhood set in the global fusion graph Representing adjacent entity nodes The attribute dimension to which it belongs This represents the weight function set for different attribute dimensions. Used to characterize the concentration of semantic elements within the neighborhood of an entity node.

[0108] When neighborhood semantic density Reaching the fourth preset threshold When the default value is 3, it indicates that multiple high-value survey elements or high-risk semantic information are clustered around the entity node, and the corresponding entity node and its neighborhood set are identified as a semantic density hotspot cluster. It is used to characterize the entity relationship region where high-risk elements or high-value semantic information are relatively concentrated in the fusion graph.

[0109] From the global fusion graph, filter out relationship edges that meet the preset association types. The preset association types can be factual associations, logical associations, temporal associations, and spatial associations, which can be specifically represented as follows:

[0110] in, Represents the relation edges in the global fusion graph. Indicates the association type corresponding to the relation edge. This represents the set of valid relation edges used for reasoning analysis.

[0111] A reasoning subgraph is constructed based on relation edges. Specifically, this reasoning subgraph is... This reasoning subgraph retains only relation edges with clear semantic orientation while keeping the set of entity nodes unchanged, in order to highlight the logical relationship structure of events.

[0112] In the inference subgraph, entity relationship paths are traversed and searched to identify those that meet preset path conditions. Each path that meets the conditions constitutes a hotspot for the step size of an inference chain. The specific preset path conditions are as follows: Path length (Default: 6 nodes, i.e. 5 steps of reasoning); The path must contain at least three entity nodes with different attribute types.

[0113] For example, see reference Figure 11 The diagram illustrates the construction of a hotspot relationship cluster using the inference chain step size. It identifies a long logical chain composed of strong semantic relationships from a complex global network. (Entity nodes in the diagram...) ~ This forms a chain of reasoning.

[0114] The identified hotspot relationship clusters of reasoning chains are used to characterize the entity relationship structure formed by the connection of multiple relationships, in order to reveal the complete logical chain or potential causal path from the initial cause to the result, and to provide support for evidence connection and causal analysis in electrical fire investigations.

[0115] Step S104: Construct an information graph for electrical fire investigation scenarios based on hotspot relationship clusters and the global fusion graph.

[0116] After identifying hotspot relationship clusters, a global fusion graph is used as the basic structural framework to uniformly organize the entity nodes and relationships in the fusion graph. Key content of the information graph is then annotated and enhanced in conjunction with these hotspot relationship clusters. Specifically, the entity nodes in the global fusion graph are treated as the node set of the information graph, with each entity node associated with at least its name, attribute category, and corresponding summary information. The relationship edges in the global fusion graph are treated as the relationship set of the information graph, with each relationship edge labeled with its corresponding association type and relationship summary, used to characterize the factual, logical, or temporal connections between entities.

[0117] Based on this, the identified hotspot relationship clusters are mapped onto the information graph, and entity nodes and relationship edges belonging to the same hotspot relationship cluster are aggregated, identified, or grouped for management to highlight their importance in the investigation and analysis.

[0118] Among them, the topological density hotspot clusters in the hotspot relationship clusters are used to identify the core areas where high-frequency interactions or multi-source evidence converge, the semantic density hotspot clusters are used to identify areas where high-risk elements or key semantic information are concentrated, and the reasoning chain hotspot relationship clusters are used to identify the relationship path structure that can reflect the logic of event development or causal relationship.

[0119] By integrating hotspot relationship clusters with a global fusion graph, an information graph for electrical fire investigation scenarios is formed. This allows entities, relationships, and their associated structures in multiple investigation evidence documents to be presented in a unified and traceable manner, thus providing a foundation for subsequent evidence analysis, causal reasoning, and interpretable presentation.

[0120] In one embodiment, Figure 12 The diagram illustrates a process for constructing an information graph for electrical fire investigation scenarios, which may include the following steps: Step S701: Obtain the hotspot type corresponding to the hotspot relationship cluster, as well as the entity nodes, relationship edges, and entity relationship path information contained therein.

[0121] After identifying hotspot relationship clusters, an information graph for electrical fire investigation scenarios is constructed based on these clusters and the global fusion graph. This construction process first involves information perception processing for different hotspot relationship clusters, specifically acquiring the hotspot type corresponding to each cluster, as well as the entity nodes, relationship edges, and entity relationship paths contained within that cluster. The hotspot type characterizes the basis for the formation of the hotspot relationship cluster and the focus of analysis. Different hotspot types reflect different entity association patterns and investigative values, thus providing constraints and semantic guidance for the subsequent semantic generation process.

[0122] Step S702: Call the large language model to generate hotspot information keywords and summary text to describe hotspot relationship clusters based on the hotspot type and the structural features and semantic content of hotspot relationship clusters.

[0123] After obtaining the structural and semantic information of the hot topic relationship clusters, a large language model is invoked to perform semantic generation processing on the hot topic relationship clusters. During the generation process, the large language model can combine the attribute composition of entities in the hot topic relationship clusters, the organization of associations, and the logical structure reflected by the entity relationship paths to perform semantic abstraction of the survey concerns reflected by the hot topic relationship clusters, thereby generating hot topic information keywords to summarize the core semantics of the hot topic relationship clusters, and simultaneously generating corresponding natural language summary text.

[0124] The summary texts corresponding to different hot topic types differ in their focus to adapt to the analysis objectives and semantic features of different hot topic relationship clusters, thereby enhancing the ability of the summary content to express investigation clues, risk characteristics, or event evolution processes.

[0125] In practical applications, large language models can generate differentiated summary texts for different hot topic types based on pre-defined generation constraints or semantically guided templates. For example: For topological density hotspot clusters, the generated summary text is used to highlight the concentration of multi-source evidence and the high-frequency association of key entities. For example, a summary template can be used: This area is the intersection of multiple pieces of evidence: the inspection record, the interrogation record and the monitoring log all mention [core entity], involving [number] related elements.

[0126] For semantic density hotspot clusters, the generated summary text is used to emphasize the concentrated distribution characteristics of high-risk elements or key semantic information. For example, a summary template can be used: High-risk elements are concentrated here: including [equipment], [combustible materials] and [management deficiencies], which is consistent with the typical electrical fire causation pattern.

[0127] For hotspot relationship clusters in the reasoning chain, the generated summary text is used to depict the complete logical path formed by multiple interconnected relationships, reflecting the event development process or potential causal chains. For example, a summary template can be used to form a complete reasoning chain: This reveals the entire process from management deficiencies to the spread of the fire.

[0128] Step S703: Organize the hot information keywords and summary text generated by the hot relationship clusters in a structured manner to construct a hot summary index structure.

[0129] After generating hot topic keywords and abstract text, the generated results are organized in a structured manner to construct a hot topic abstract index structure. This serves as an auxiliary index for the information graph. The hotspot summary index structure uses hotspot relationship clusters as the basic index unit, and the generated summary text is used to depict the complete logical path formed by multiple interconnected relationships, reflecting the development process of events or potential causal chains.

[0130] In the hotspot summary index structure Each index record includes at least: hot information keywords identifying the core semantics of hot relationship clusters, a list of entity nodes involved in the hot relationship cluster, and a summary text corresponding to that hot relationship cluster. By uniformly formatting and structurally storing the hot information, centralized management, rapid location, and semantic retrieval of different hot relationship clusters are achieved.

[0131] The following is a specific example of an index item: Keywords: Causes of fire; Entities involved: [Distribution box, cotton curtains, no preventative maintenance conducted]; Abstract: High-risk factors are concentrated here... leading to a fire.

[0132] Step S704: The hotspot summary index structure and the global fusion graph are output as an information graph for electrical fire investigation scenarios.

[0133] By linking and integrating the hotspot summary index structure with the global fusion graph, an information graph output is generated for electrical fire investigation scenarios. The information graph also retains a structured representation of entity relationships, i.e., the global fusion graph. and hotspot summary index structure This results in a graph representation that combines structured relational expression with semantic overview capabilities.

[0134] Among them, the global fusion graph As a structured relational carrier of information graphs, it supports entity relation queries, cross-evidence path tracing, and visual analysis of entity relation networks. Hotspot summary index structure. As a semantic auxiliary index for information graphs, it is used to quickly locate high-value relationship areas, focus on key relationships, and efficiently extract survey conclusions or analysis report materials.

[0135] Through the above methods, when conducting relationship queries, path analysis, or event reconstruction, investigators can not only conduct in-depth analysis of entity relationships based on the global fusion graph, but also quickly identify key relationship areas and reasoning clues with significant investigative value in complex relationship networks by leveraging the hotspot summary index structure. This effectively improves the efficiency and relevance of electrical fire investigation and analysis while ensuring the integrity of the analysis.

[0136] In the above embodiments, multiple electrical fire investigation evidence documents are semantically understood and structurally extracted under domain ontology constraints, and uniformly modeled in the form of local subgraphs. Then, based on entity semantic consistency, graph structure context consistency, and summary semantic information, cross-subgraph entity association and fusion are achieved. This allows for the automatic construction of a global fusion graph reflecting the inherent relationships between multi-source investigation evidence while maintaining the original semantic integrity of the evidence. Furthermore, through comprehensive analysis of the entity relationship structure, semantic distribution characteristics, and relationship connection patterns in the global fusion graph, hotspot relationship clusters with outstanding investigation value are automatically identified. Combined with the hotspot summary index structure, key relationship regions are semantically summarized and indexed, enabling the complex entity relationship network to be presented in a structured, interpretable, and easily searchable form.

[0137] Therefore, this method not only significantly reduces the reliance on human experience and manual sorting in the investigation of electrical fires, effectively improving the processing efficiency and fusion depth of multi-source heterogeneous evidence, but also supports the intuitive analysis and visualization of complex causal relationships and event evolution chains. This enables investigators to accurately grasp key facts, evidence conflicts, and potential liability connections in a short period of time, thereby improving the overall efficiency, reliability, and persuasiveness of conclusions in electrical fire investigation and analysis.

[0138] In one embodiment, an electrical fire scene information map construction device is provided, referring to... Figure 13 As shown, the electrical fire scene information map construction device 800 may include: a local subgraph construction module 801, a cross-subgraph association identification module 802, a hotspot relationship cluster determination module 803, and an information map construction module 804.

[0139] Among them, the local subgraph construction module 801 is used to construct a local subgraph corresponding to each investigation evidence file based on the predefined electrical fire domain ontology constraints and by using a large language model to perform semantic understanding and structured extraction on multiple electrical fire investigation evidence files. The local subgraph includes entity nodes, entity attributes describing the characteristics of entity nodes, relation edges representing the interaction logic between entities, and summary text generated by combining factual information of entity nodes or relation edges. The cross-subgraph association recognition module 802 is used to identify entity nodes that are associated across subgraphs based on the semantic consistency of entity nodes in local subgraphs and the contextual consistency of graph structure, combined with the summary text, to form an association dataset. The association dataset is then merged with different local subgraphs to generate a global fusion graph representing the association relationship of multiple investigation evidence documents. The hotspot relationship cluster determination module 803 is used to identify the entity relationship structure that is related across multiple local subgraphs and determine the hotspot relationship clusters based on the graph structure features, entity semantic features and relationship connection patterns in the global fusion graph. The information graph construction module 804 is used to construct an information graph for electrical fire investigation scenarios based on hotspot relationship clusters and the global fusion graph.

[0140] In one embodiment, the cross-subgraph association recognition module 802 is specifically used to: extract the standardized name, entity attributes and corresponding summary text of the entity node in each local subgraph, and construct a triplet for a structured description of the entity node; Transform the triples of entity nodes into semantic feature vectors of entity nodes; For any two candidate entity node pairs consisting of entity nodes from different local subgraphs, the semantic consistency and graph structure context consistency of the candidate entity node pairs are calculated based on semantic feature vectors, and the associated dataset is determined based on the calculation results.

[0141] In one embodiment, the cross-subgraph association recognition module 802 is specifically used to: calculate the cosine similarity of candidate entity node pairs based on the semantic feature vectors corresponding to the candidate entity node pairs; The cosine similarity is weighted and corrected using the attribute adjacency coefficient to obtain the final semantic similarity; where the attribute adjacency coefficient represents the semantic affinity between candidate entity nodes and their respective attributes. If the semantic similarity is greater than or equal to the first preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

[0142] In one embodiment, the cross-subgraph association identification module 802 is specifically used to: extract the neighborhood relationship set of each entity node in the candidate entity node pair in its respective local subgraph; the neighborhood relationship set includes the adjacent entity nodes of the entity node, and the relationship type between the entity node and its adjacent entity nodes; Semantic vectorization representation of adjacent entity nodes involved in the neighborhood relationship set is performed to obtain adjacent entity semantic embedding vectors used to characterize the semantic features of neighborhood relationships; The context similarity of candidate entity node pairs is calculated based on the semantic embedding vectors of adjacent entities corresponding to each entity node in the candidate entity node pair. If the context similarity is greater than or equal to the second preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

[0143] In one embodiment, the cross-subgraph association recognition module 802 is further configured to: for entity node pairs that have an association relationship in the association dataset, fuse the summary text of the entity node pairs in the local subgraph, and call the large language model to generate summary text of cross-subgraph association; The generated summary text is combined with the corresponding entity nodes and their relationships to form a structured relational unit for characterizing cross-subgraph relationships.

[0144] In one embodiment, the hotspot relationship cluster determination module 803 is specifically used to: based on the neighborhood connection of entity nodes in the global fusion graph, to count the neighborhood interaction strength of each entity node, and to determine the entity nodes whose neighborhood interaction strength reaches a third preset threshold and their corresponding neighborhood sets as topological density hotspot clusters. The attribute dimensions of each entity node are weighted based on a preset attribute weight function, the neighborhood semantic density of each node is calculated, and the entity nodes whose neighborhood semantic density reaches the fourth preset threshold and their corresponding neighborhood sets are identified as semantic density hotspot clusters. The system filters out relation edges that meet the preset association type from the global fusion graph, and searches for entity relation paths that meet the preset path conditions in the inference subgraph constructed based on the relation edges. The entity relation paths are then identified as hotspot relation clusters in the inference chain.

[0145] In one embodiment, the information graph construction module 804 is specifically used to: obtain the hotspot type corresponding to the hotspot relationship cluster and the entity nodes, relationship edges and entity relationship path information contained therein; The large language model is invoked to generate hot topic information keywords and summary texts to describe hot topic relationship clusters based on the hot topic type and the structural features and semantic content of hot topic relationship clusters. The hot information keywords and summary text generated by hot relationship clusters are structured and organized to construct a hot summary index structure; The hotspot summary index structure and the global fusion graph are used as the output information graph for electrical fire investigation scenarios.

[0146] Specific limitations regarding the electrical fire scene information map construction device 800 can be found in the limitations of the electrical fire scene information map construction method described above, and will not be repeated here. Each module in the aforementioned electrical fire scene information map construction device 800 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0147] In one embodiment, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement a method for constructing an electrical fire scene information map.

[0148] In one embodiment, a computer storage medium is provided on which a computer program is stored, and when the computer program is executed by a processor, it implements a method for constructing an electrical fire scene information map.

[0149] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0150] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0151] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for constructing an information map of electrical fire scenes, characterized in that, include: Based on predefined ontology constraints in the field of electrical fires, a large language model is used to perform semantic understanding and structured extraction on multiple electrical fire investigation evidence documents, and a local subgraph corresponding to each of the investigation evidence documents is constructed. The local subgraph includes entity nodes, entity attributes describing the characteristics of the entity nodes, relation edges representing the interaction logic between entities, and summary text generated by combining the factual information of the entity nodes or relation edges. Based on the semantic consistency of entity nodes in the local subgraphs and the contextual consistency of the graph structure, and combined with the summary text, entity nodes that are related across subgraphs are identified to form a related dataset. The related dataset is then merged with different local subgraphs to generate a global fusion graph representing the relationship between multiple investigation evidence documents. Based on the graph structure features, entity semantic features, and relation connection patterns in the global fusion graph, the entity relation structure that is related across multiple local subgraphs is identified, and hotspot relation clusters are determined. Based on the hotspot relationship clusters and the global fusion graph, an information graph for electrical fire investigation scenarios is constructed.

2. The method for constructing an electrical fire scene information map according to claim 1, characterized in that, Based on the semantic consistency of entity nodes in the local subgraph and the contextual consistency of the graph structure, and combined with the summary text, entity nodes that are related across subgraphs are identified, forming an association dataset, including: For each entity node in the local subgraph, the standardized name, entity attributes, and corresponding summary text of the entity node are extracted to construct a triple for a structured description of the entity node. The triples of the entity node are transformed into the semantic feature vectors of the entity node; For any two candidate entity node pairs consisting of entity nodes from different local subgraphs, the semantic consistency and graph structure context consistency of the candidate entity node pairs are calculated based on the semantic feature vectors, and the associated dataset is determined based on the calculation results.

3. The method for constructing an electrical fire scene information map according to claim 2, characterized in that, The process of determining the semantic consistency of entity nodes includes: Based on the semantic feature vectors corresponding to the candidate entity node pairs, calculate the cosine similarity of the candidate entity node pairs; The cosine similarity is weighted and corrected using the attribute adjacency coefficient to obtain the final semantic similarity; wherein, the attribute adjacency coefficient represents the semantic affinity between the candidate entity nodes and their respective attributes; If the semantic similarity is greater than or equal to the first preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

4. The method for constructing an electrical fire scene information map according to claim 2, characterized in that, The process for determining the contextual consistency of a graph structure includes: For each entity node in the candidate entity node pair, extract the neighborhood relationship set of the entity node in its respective local subgraph; the neighborhood relationship set includes the neighboring entity nodes of the entity node, and the relationship type between the entity node and its neighboring entity nodes; Semantic vectorization representation is performed on the adjacent entity nodes involved in the neighborhood relationship set to obtain the adjacent entity semantic embedding vector used to characterize the semantic features of the neighborhood relationship; The context similarity of the candidate entity node pair is calculated based on the semantic embedding vector of the adjacent entity corresponding to each entity node in the candidate entity node pair. If the context similarity is greater than or equal to the second preset threshold, the corresponding candidate entity node pair will be stored in the associated dataset.

5. The method for constructing an electrical fire scene information map according to claim 1 or 2, characterized in that, The method further includes: For entity node pairs that are related in the associated dataset, the summary text of the entity node pairs in the local subgraph is fused, and a large language model is called to generate summary text of cross-subgraph association; The generated summary text is combined with the corresponding entity nodes and their relationships to form a structured relational unit for characterizing cross-subgraph relationships.

6. The method for constructing an electrical fire scene information map according to claim 1, characterized in that, The identified hotspot relationship clusters include at least one of the following: Based on the neighborhood connectivity of entity nodes in the global fusion graph, the neighborhood interaction strength of each entity node is statistically analyzed, and the entity nodes whose neighborhood interaction strength reaches the third preset threshold and their corresponding neighborhood sets are determined as topological density hotspot clusters. The attribute dimensions of each entity node are weighted based on a preset attribute weight function, the neighborhood semantic density of each node is calculated, and the entity nodes whose neighborhood semantic density reaches the fourth preset threshold and their corresponding neighborhood sets are identified as semantic density hotspot clusters. Relationship edges that meet the preset association type are selected from the global fusion graph, and entity relationship paths that meet the preset path conditions are searched in the inference subgraph constructed based on the relationship edges. The entity relationship paths are determined as hotspot relationship clusters of the inference chain.

7. The method for constructing an electrical fire scene information map according to claim 1, characterized in that, The step of constructing an information graph for electrical fire investigation scenarios based on the hotspot relationship clusters and the global fusion graph includes: Obtain the hotspot type corresponding to the hotspot relationship cluster, as well as the entity nodes, relationship edges, and entity relationship path information contained therein; The large language model is invoked to generate hotspot information keywords and summary text describing the hotspot relationship clusters based on the hotspot type and the structural features and semantic content of the hotspot relationship clusters. The hot topic information keywords and summary text generated by the hot topic relationship clusters are structured and organized to construct a hot topic summary index structure; The hotspot summary index structure and the global fusion graph are used as the output information graph for electrical fire investigation scenarios.

8. A device for constructing an information map of an electrical fire scene, characterized in that, include: The local subgraph construction module is used to construct a local subgraph corresponding to each of the obtained electrical fire investigation evidence documents by performing semantic understanding and structured extraction on multiple obtained electrical fire investigation evidence documents based on predefined electrical fire domain ontology constraints and using a large language model. The local subgraph includes entity nodes, entity attributes describing the characteristics of the entity nodes, relation edges representing the interaction logic between entities, and summary text generated by combining the factual information of the entity nodes or relation edges. The cross-subgraph association identification module is used to identify entity nodes that are associated across subgraphs based on the semantic consistency of entity nodes in the local subgraphs and the contextual consistency of the graph structure, and in combination with the summary text, to form an association dataset, and to merge the association dataset with different local subgraphs to generate a global fusion graph representing the association relationship of multiple investigation evidence documents. The hotspot relationship cluster determination module is used to identify the entity relationship structure that is related across multiple local subgraphs and determine the hotspot relationship clusters based on the graph structure features, entity semantic features and relationship connection patterns in the global fusion graph. The information graph construction module is used to construct an information graph for electrical fire investigation scenarios based on the hotspot relationship clusters and the global fusion graph.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the electrical fire scene information map construction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the electrical fire scene information map construction method according to any one of claims 1 to 7.

Images