Fire-fighting supervision and communication global data-oriented knowledge graph construction and correlation analysis method and system

Abstract

The invention relates to the technical field of intelligent fire protection and big data analysis, and discloses a knowledge graph construction and association analysis method and system for fire protection supervision and communication global data. In order to solve the technical problems of heterogeneous fire-fighting data and single association, the invention provides a domain ontology-driven knowledge graph automatic construction and deep analysis method. The core innovation of the method is that a multi-source heterogeneous data fusion model is adopted to unify semantic representation of multi-modal data such as fire-fighting documents and Internet of Things perception; performing entity relationship joint extraction by utilizing the pre-training model, and efficiently constructing a fire-fighting knowledge graph of massive triples; and a graph neural network is introduced to carry out deep association mining and reasoning, a risk conduction path is quantitatively evaluated, and potential threats are predicted. According to the invention, discrete fire-fighting data is improved into a computable and reasonable knowledge system, and deep perception and intelligent decision-making of a fire-fighting safety situation are realized.

Description

Technical Field

[0001] This invention relates to the fields of smart fire protection and big data analysis technology, and in particular to a method and system for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain. Background Technology

[0002] In the development of fire protection information systems, massive amounts of multimodal and cross-domain business data are generated across various business processes. Currently, the application of this data mainly relies on their respective independent business systems, achieving preliminary statistical analysis and simple rule-based alarms, which constitutes the basic technical status of current smart fire protection data applications.

[0003] However, the aforementioned existing technologies reveal significant technical bottlenecks when conducting comprehensive fire safety risk analysis. First, severe "information silos" have formed between various business systems, with inconsistent data standards and semantic gaps, making effective cross-domain data integration extremely difficult. Second, existing data analysis methods generally remain at a superficial statistical level of direct correlations, failing to address indirect or hidden connections that require multiple intermediate steps to establish links. For example, the potential risk transmission chain between a unit's historical violations, abnormal communication by its responsible person, and the malfunctioning state of surrounding facilities is undetectable by traditional methods. Most importantly, existing systems are essentially static knowledge bases, lacking the ability to learn from analysis results and evolve independently. This hinders the intelligent transformation from "post-event response" to "pre-event prediction," failing to meet the urgent needs of modern fire protection work for proactive risk warning and precise decision-making.

[0004] Therefore, this invention aims to solve the key problems existing in the prior art, namely, on the basis of breaking through the semantic barriers of multi-source heterogeneous data in the fire protection field and realizing the deep integration of data across the entire domain, further realizing the in-depth mining of hidden and diverse risk transmission patterns and abnormal correlations, and finally constructing a dynamic cognitive closed loop that can learn from the analysis results and drive the self-evolution of knowledge structure, so as to comprehensively improve the level of intelligent and forward-looking early warning decision-making for fire safety risks. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method and system for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain, solving the technical problems of heterogeneous and singular associations in existing fire data.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: The first aspect is a method for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain, which includes the following steps: Step 1: Format the multi-source heterogeneous data from the fire safety field into a unified data object stream, where each data object contains metadata describing its source, time, and credibility; Step 2: Initialize the basic ontology containing core fire protection concepts, and based on the data object stream, use the entity relationship joint extraction model to synchronously identify entities, multi-relationships and event hyperedges from the data object stream, so as to iteratively and incrementally construct a fire protection knowledge graph with a hypergraph structure; Step 3: Deploy a graph neural network that integrates a temporal causal attention mechanism to perform deep association analysis on the fire protection knowledge graph. This attention mechanism dynamically weights the paths in the fire protection knowledge graph based on the timeliness of events and the causal patterns of historical fires. Step 4: Establish a cognitive feedback and graph self-optimization closed loop. The deep analysis results obtained by the graph neural network are used in reverse to the fire protection knowledge graph. By automatically adjusting the connection weights between entities in the fire protection knowledge graph and proposing new relation types to the basic ontology, the dynamic self-optimization and continuous evolution of the fire protection knowledge graph are realized.

[0008] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, step 1, which involves formatting multi-source heterogeneous data from the fire safety field into a unified data object stream, specifically includes: pre-defining a standardized semantic data structure, wherein the semantic data structure consists of original data fields and associated multi-dimensional metadata fields, and parsing and mapping the multi-source heterogeneous data to the semantic data structure to generate data objects constituting the data object stream; The multidimensional metadata fields include at least: source metadata to characterize the data source and collection time, quality assessment metadata to characterize the data reliability and accuracy, and preprocessed semantic metadata extracted from the multi-source heterogeneous data through a preliminary analysis model.

[0009] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, in step 2, the working mode of the entity relationship joint extraction model includes: using a deep semantic encoder to encode the data in the data object stream to generate context-related word vector representations that can characterize context semantics; Based on the context-related word vector representation, the potential entities identified in the data are used as nodes and dynamically constructed into a local graph structure. The features of the nodes are initialized by aggregating the vector representations of the words they constitute. A graph neural network decoder is deployed to update node features through multi-round information transmission between nodes on the local graph structure, and to perform joint decoding based on the updated node features to directly output event hyperedges that represent the complex multi-dimensional relationships between the potential entities.

[0010] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, in step 3, the way in which the attention mechanism dynamically weights the paths in the fire knowledge graph includes: calculating a dynamic attention coefficient for adjusting the information transmission intensity for any side between nodes in the fire knowledge graph, wherein the dynamic attention coefficient is formed by weighted fusion of temporal attention component and causal attention component; The value of the temporal attention component decreases as the time interval between the events associated with the edge increases, while the value of the causal attention component increases as the historical risk level of the association type represented by the edge increases.

[0011] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, step 4, the step of automatically adjusting the connection weights between entities in the fire knowledge graph, includes: identifying the transmission path connecting risk source entity nodes and risk target entity nodes in the fire knowledge graph, and calculating the comprehensive risk score of the transmission path. When the comprehensive risk score is higher than a preset threshold, the transmission path is identified as a high-risk transmission path. Starting from the risk target entity node, trace back along the high-risk transmission path, and perform non-linear enhancement adjustments on the connection weights represented by each edge constituting the high-risk transmission path. The magnitude of the enhancement adjustment for any side is positively correlated not only with the comprehensive risk score of the high-risk transmission path, but also with the dynamic attention coefficient value calculated for that specific side.

[0012] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, step 4, the step of proposing new relation types to the basic ontology includes: obtaining the semantic vector of each entity node output by the graph neural network after deep association analysis, wherein the semantic vector integrates the structural information of the node in its neighborhood; In the vector space composed of the semantic vectors, cluster analysis is used to identify entity node clusters whose distance between vectors is less than a preset threshold, but which have no direct edge connection in the fire protection knowledge graph. The common context information of entity pairs within the entity node cluster in the multi-source heterogeneous data is automatically extracted, and based on the common context information, a semantic description of potential new relationship types is summarized and generated.

[0013] As a preferred embodiment of the knowledge graph construction and association analysis method for fire supervision and communication full-domain data described in this invention, the deep association analysis in step 3 further includes an abnormal association detection step, which includes: extracting subgraphs that conform to a preset conventional pattern from the fire knowledge graph as conventional pattern subgraph samples, and training a graph autoencoder model composed of a graph encoder and a graph decoder based on the conventional pattern subgraph samples; The subgraph structure to be detected in the fire protection knowledge graph is input into the graph encoder to generate a low-dimensional latent vector representing the subgraph structure. The low-dimensional latent vector is then input into the graph decoder to generate a reconstructed subgraph. The reconstruction error is determined by calculating the comprehensive difference between the subgraph structure to be detected and the reconstructed subgraph in terms of topology and node attributes. When the reconstruction error is greater than the normal error baseline obtained from the statistical analysis of the conventional pattern subgraph samples, the subgraph to be detected is determined to contain abnormal correlation.

[0014] Furthermore, this invention also provides a knowledge graph construction and association analysis system for fire supervision and communication full-domain data, wherein a unified semantic representation engine is configured to: receive multi-source heterogeneous data from the fire safety field, and convert the multi-source heterogeneous data into a unified data object stream in real time and without patterning according to a predefined standardized semantic data structure containing original data fields and multi-dimensional metadata fields; A cognitive graph reasoning engine, connected to the unified semantic representation engine, is configured to: continuously acquire data objects from the data object stream and perform knowledge construction and association analysis operations; wherein, the knowledge construction operation iteratively maintains a fire protection knowledge graph with a hypergraph structure through an entity relation joint extraction model; the association analysis operation performs deep analysis on the fire protection knowledge graph by deploying a graph neural network integrating a temporal causal attention mechanism to output analysis results containing high-risk transmission chains and abnormal association patterns; The graph evolution and ontology management engine is connected to the cognitive graph reasoning engine and is configured to receive the analysis results and perform a dual-channel self-optimization operation based on the results. The first channel adjusts the connection weights in the fire protection knowledge graph in reverse according to the high-risk transmission chain, while the second channel automatically generates potential new relation type definitions to expand the basic ontology based on the entity node semantic vectors output by the graph neural network. The three engines work together to form an adaptive cognitive loop from data processing and knowledge reasoning to knowledge evolution.

[0015] As a preferred embodiment of the knowledge graph construction and association analysis system for fire supervision and communication full-domain data described in this invention, the knowledge construction operation of the cognitive graph reasoning engine is performed on each data object received from the data object stream by executing the following steps: First, the data object is subjected to deep semantic encoding to generate a context-related word vector representation that can characterize the contextual meaning of each word element; Secondly, based on the vector representation, the potential entities identified in the data object are used as nodes to dynamically construct a local graph structure, and the features of the nodes are initialized by aggregating the vector representations of their constituent words. Then, a graph neural network decoder is deployed on the local graph structure to update node features by performing multiple rounds of information transmission between nodes, and joint decoding is performed based on the updated node features to directly output event hyperedges that characterize the complex multi-dimensional relationships between the potential entities. Finally, the event hyperedge, along with the confirmed entities and relationships contained therein, is integrated into the fire protection knowledge graph as an indivisible transaction unit, thereby ensuring data consistency of the knowledge graph during incremental updates.

[0016] As a preferred embodiment of the knowledge graph construction and association analysis system for fire supervision and communication full-domain data described in this invention, the graph evolution and ontology management engine is specifically configured to achieve its self-optimization function by executing operations including the following first and second channels in parallel: The first channel is a connection weight enhancement channel. In this channel, when a high-risk transmission path and its analysis metadata are received from the cognitive graph inference engine, a reverse tracing is performed starting from the risk target entity node of the path, and a weight enhancement factor is calculated for any side of the path. The weight enhancement factor is jointly determined by the comprehensive risk score of the path and the dynamic attention coefficient value of the side in forward inference. Then, the weight enhancement factor is applied to update the basic connection weight of the side in the fire protection knowledge graph. The second channel is a new relationship discovery channel. In this channel, the latest semantic vectors of entity nodes are periodically obtained from the cognitive graph reasoning engine. Cluster analysis is performed in the vector space composed of the semantic vectors to generate entity node clusters. Clusters that meet the following conditions are selected from the entity node clusters: the nodes inside are semantically similar, but are not directly connected in the topology of the fire protection knowledge graph. Then, the common context information of entity pairs in the selected clusters is automatically extracted to summarize and generate semantic descriptions of potential new relationship types, and these are submitted to the review module as ontology evolution proposals.

[0017] The beneficial effects of this invention are as follows: By unifying semantic representation, this invention completely breaks down data barriers between various fire protection business systems, achieving the first deep fusion of comprehensive, multimodal data, providing a solid foundation for global security situation awareness. Utilizing event hyperedges and deeply customized graph neural networks, this invention greatly enhances the depth of risk analysis, automatically discovering hidden, complex risk transmission chains and abnormal correlation patterns that traditional methods cannot reach, elevating risk identification from judging isolated hazard points to accurately understanding dynamic risk networks. The invention's unique cognitive feedback and self-optimizing closed loop endow the system with the ability to learn from analysis results and achieve self-evolution of its knowledge structure. This "growthable" cognitive characteristic powerfully propels fire safety management from a passive "post-event response" mode to a proactive "pre-event warning and precise decision-making" mode, significantly improving the accuracy of fire risk prediction and the scientific level of emergency command. Attached Figure Description

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

[0019] Figure 1 A flowchart outlining the overall process for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain.

[0020] Figure 2 This is an architecture diagram of a knowledge graph construction and association analysis system for fire supervision and communication data across the entire domain.

[0021] Figure 3 This is an internal workflow diagram of the graph evolution and ontology management engine in a knowledge graph construction and association analysis system for fire supervision and communication full-domain data. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0025] Example 1: Refer to Figure 1 This is one embodiment of the present invention. The present invention proposes a method for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain. The implementation of this method begins with the formatting of multi-source heterogeneous data from the fire safety field to form a unified data object stream, including the following steps: Step 1: Format the multi-source heterogeneous data from the fire safety field into a unified data object stream, where each data object contains metadata describing its source, time, and credibility; Step 2: Initialize the basic ontology containing core fire protection concepts, and based on the data object flow, use the entity relationship joint extraction model to synchronously identify entities, multi-relationships and event hyperedges from the data object flow, so as to iteratively and incrementally build a fire protection knowledge graph with a hypergraph structure. Step 3: Deploy a graph neural network that integrates a temporal causal attention mechanism to perform deep correlation analysis on the fire protection knowledge graph. This attention mechanism dynamically weights the paths in the fire protection knowledge graph based on the timeliness of events and the causal patterns of historical fires. Step 4: Establish a cognitive feedback and graph self-optimization closed loop. The deep analysis results from the graph neural network are used in reverse to affect the fire protection knowledge graph. By automatically adjusting the connection weights between entities in the fire protection knowledge graph and proposing new relation types to the basic ontology, the dynamic self-optimization and continuous evolution of the fire protection knowledge graph are achieved.

[0026] Step 1, which involves formatting multi-source heterogeneous data from the fire safety field into a unified data object stream, specifically includes: pre-defining a standardized semantic data structure, which consists of original data fields and associated multi-dimensional metadata fields; and parsing and mapping multi-source heterogeneous data to the semantic data structure to generate data objects that constitute the data object stream. The multidimensional metadata fields should include at least: source metadata to characterize the data source and collection time, quality assessment metadata to characterize the data reliability and accuracy, and preprocessed semantic metadata extracted from multi-source heterogeneous data through a preliminary analysis model.

[0027] In step 2, the entity relation joint extraction model works by: using a deep semantic encoder to encode the data in the data object stream to generate context-related word vector representations that can characterize the context semantics; Based on context-related word vector representations, potential entities identified in the data are used as nodes and dynamically constructed into a local graph structure. The features of the nodes are initialized by aggregating the vector representations of the constituent words. The deployed graph neural network decoder updates node features through multi-round information transmission between nodes in the local graph structure, and performs joint decoding based on the updated node features to directly output event hyperedges that represent complex multi-dimensional relationships between potential entities.

[0028] In step 3, the attention mechanism dynamically weights the paths in the fire protection knowledge graph by: calculating a dynamic attention coefficient for any side between nodes in the fire protection knowledge graph to adjust the intensity of information transmission. The dynamic attention coefficient is formed by weighted fusion of temporal attention components and causal attention components. The value of the temporal attention component decreases as the time interval between the events associated with the edge increases, while the value of the causal attention component increases as the historical risk level of the association type represented by the edge increases.

[0029] In step 4, the steps of automatically adjusting the connection weights between entities in the fire protection knowledge graph include: identifying the transmission paths connecting risk source entity nodes and risk target entity nodes in the fire protection knowledge graph, calculating the comprehensive risk score of the transmission path, and determining the transmission path as a high-risk transmission path when the comprehensive risk score is higher than a preset threshold. Starting from the risk target entity node, trace back along the high-risk transmission path, and perform non-linear enhancement adjustments on the connection weights represented by each edge constituting the high-risk transmission path. The magnitude of the enhancement adjustment for any side is positively correlated not only with the comprehensive risk score of the high-risk transmission path, but also with the dynamic attention coefficient value calculated for that specific side.

[0030] In step 4, the step of proposing new relation types to the basic ontology includes: obtaining the semantic vector of each entity node output by the graph neural network after deep association analysis, wherein the semantic vector integrates the structural information of the node in its neighborhood; In a vector space composed of semantic vectors, cluster analysis is used to identify clusters of entity nodes whose distance between vectors is less than a preset threshold but have no direct edge connection in the fire protection knowledge graph. It automatically extracts common context information of entity pairs in multi-source heterogeneous data within the entity node cluster, and summarizes and generates semantic descriptions of potential new relationship types based on the common context information.

[0031] The deep association analysis in step 3 further includes an abnormal association detection step, which includes: extracting subgraphs that conform to a preset conventional pattern from the fire protection knowledge graph as conventional pattern subgraph samples, and training a graph autoencoder model composed of a graph encoder and a graph decoder based on the conventional pattern subgraph samples. The structure of the subgraph to be detected in the fire protection knowledge graph is input into the graph encoder to generate a low-dimensional latent vector representing the structure of the subgraph. The low-dimensional latent vector is then input into the graph decoder to generate a reconstructed subgraph. The reconstruction error is determined by calculating the comprehensive difference between the subgraph to be detected and the reconstructed subgraph in terms of topology and node attributes. When the reconstruction error is greater than the normal error baseline obtained from the statistical analysis of the subgraph samples of the regular pattern, the subgraph to be detected is determined to contain abnormal associations.

[0032] The data sources may include, but are not limited to: structured law enforcement record databases exported from the fire supervision and management information system, semi-structured fire inspection documents in Word or PDF format filled out by law enforcement officers, voice communication transcripts from the emergency command center, and real-time unstructured alarm data reported by various fire IoT sensors (such as smoke detectors, temperature detectors, and water pressure sensors).

[0033] To process this data, this embodiment first designs and implements a unified fire protection data representation format based on JSON-based linked data. To illustrate this more clearly, the aforementioned JSON-based linked data formatting technology is explained in detail. Traditional JSON data can only be read by machines, but their true meaning cannot be understood. For example, a machine cannot distinguish whether the field "responsible person" refers to a person or a position. This semantic uncertainty is the root cause of the ineffective integration of data and the formation of information silos. This invention, by using JSON-based linked data technology, adds a "semantic dictionary" (technically called @context) to each piece of data entering the system. The function of this "dictionary" is to assign a globally unified, unambiguous, and clear definition to each field in the data (such as "responsible person" and "unit address"). For example, it explicitly tells the machine that the field "responsible person" corresponds to the concept of "person," and "unit address" corresponds to the concept of "geographical location."

[0034] In this way, all raw data from diverse sources and with varying structures are transformed into unified data objects with clear, standardized semantic tags, and encapsulated into a standard fire protection unified data representation format. Each fire protection unified data representation object contains five core metadata fields: identifying the data source, recording the collection time, assessing data credibility, a list of potential entities pre-identified through a lightweight model, and storing the original data. Through this complete processing, all input data is integrated and standardized into a unified data object stream, fundamentally solving the problem of data heterogeneity and laying a solid foundation for the accuracy and efficiency of subsequent knowledge graph construction.

[0035] After obtaining a standardized data object stream, the system begins constructing a fire safety knowledge graph. This process first requires domain experts to predefine a basic ontology containing core concepts such as "social units," "fire protection facilities," "law enforcement personnel," "fire hazards," and "laws and regulations," along with their hierarchical relationships. Subsequently, based on the data object stream, the system employs an end-to-end entity relationship joint extraction model using a pre-trained language model (BERT) as the encoder and a graph attention network as the decoder.

[0036] In this embodiment, BERT is chosen as the encoder because its bidirectional encoding mechanism based on a self-attention network architecture (Transformer) can deeply capture the contextual dependencies of words. Compared to traditional models, BERT can generate dynamic word vector representations rich in deep semantics based on specific contexts, thereby accurately understanding the polysemy commonly found in fire safety documents and providing a high-precision semantic foundation for subsequent analysis. In conjunction with this, a graph attention network is chosen as the decoder because its self-attention mechanism allows the model to dynamically assign different attention weights to different neighboring nodes when aggregating neighboring node information. When extracting complex fire safety events, this means the model can intelligently determine which potential entity relationships are more critical, thus differing from the indiscriminate processing of traditional graph networks and significantly improving the ability to accurately decode complex event structures. Combining BERT's deep semantic understanding capabilities with the dynamic structural reasoning capabilities of graph attention networks constitutes the core of the joint extraction model of this invention. It transforms the problem of understanding linear text into a reasoning problem on a non-linear graph structure, thereby efficiently and accurately extracting complex event hyperedges containing multiple relationships directly from unstructured text.

[0037] To illustrate this more clearly, the "event hyperedge" used in this embodiment will be explained in detail. Traditional knowledge graphs use "edges" that can only connect two entity nodes, forming a binary relationship, which is insufficient to fully describe complex fire events involving multiple participants. For example, a fire inspection involves multiple elements such as inspectors, units, time, equipment, and issues. If represented using traditional binary relationships, this would have to be broken down into multiple fragmented associations, thus losing the overall integrity of the event. To solve this problem, this invention introduces the "event hyperedge" as the core unit of knowledge representation. A "hyperedge" is a data structure that can connect multiple nodes at once. In this invention, it is specifically used to capture and represent a complete, multi-faceted fire event, along with all its participating entities, holistically and atomically. This is analogous to a traditional "edge" being a thin line, while the "event hyperedge" is a "net" that can "encircle" all relevant parties of the event at once, greatly ensuring the fidelity and completeness of the knowledge representation. For example, an inspection event can be modeled as a hyperedge containing multiple roles. This hyperedge holistically describes several key roles in the event, such as the inspector: "Captain Wang," the inspected unit: "XX Building," the inspected equipment: "Fire Hydrant No. 3," the problem found: "Insufficient Pressure," and the inspection time: "Year XX Month XX Day," etc. These newly generated event hyperedges are incrementally incorporated into the global fire protection knowledge graph, causing its size and knowledge density to continuously grow.

[0038] As the knowledge graph is constructed and updated, the system performs deep correlation analysis on it. This embodiment deploys a graph neural network model specifically designed for this knowledge graph. To illustrate this more clearly, the graph neural network model deployed in this embodiment is described in detail. A graph neural network is a deep learning model specifically designed to process graph-structured data. Its core is to learn feature representations containing network topology through message passing between nodes. However, a standard graph neural network model often treats neighboring node information indiscriminately, failing to meet the needs of fire risk identification, which is highly sensitive to timeliness and causality. To solve this technical problem, this invention innovates the architecture of the graph neural network by deeply integrating the temporal causal attention mechanism proposed in this invention into its information aggregation layer. This means that when the graph neural network of this invention transmits information, the importance of information from neighboring nodes is no longer equal, but is dynamically modulated by an attention coefficient jointly determined by "event timeliness" and "historical causal patterns." In this way, the graph neural network of this invention is endowed with an "intuition" similar to that of a fire expert, enabling it to adaptively prioritize information flows that are more urgent in time and more dangerous in causal chains, thereby achieving accurate and forward-looking identification of high-risk transmission chains.

[0039] Specifically, the model's attention mechanism incorporates two dimensions of information when calculating the attention weights for connecting any two edges: the first is a temporal attention component, the value of which is calculated by a negative exponential function based on the timestamp difference of the events associated with the edge, with higher weights for more recent events; the second is a causal attention component, the value of which is determined by querying a "fire causal knowledge base" built through pre-analysis of tens of thousands of historical fire accident reports. For example, the historical risk level of the association between "aging electrical wiring" and "short circuit occurrence" is assigned a very high weight value.

[0040] In this embodiment, the dynamic attention coefficient That is, from the adjacent node To the central node The information transmission strength is calculated as follows: First, calculate the temporal attention components separately. and causal attention components Time attention component The calculation aims to quantify the timeliness of an event, and its value is negatively correlated with the time interval between the event's occurrence. A specific implementation uses a negative exponential decay function:

[0041] in, and Representing nodes respectively and The timestamp of the associated event Indicates the time interval between the two. It is an adjustable time decay hyperparameter used to control the rate at which the weights decay over time.

[0042] Causal attention components The calculation aims to incorporate domain knowledge, with values ​​positively correlated with historical risk levels for specific association types. A concrete implementation involves directly querying a pre-built "fire causal knowledge base":

[0043] in, Represents the statistical results obtained from historical data, derived from nodes. The fire safety concepts (reasons) and key points represented The strength of the prior causal relationship between the fire safety concepts (outcomes) they represent is a pre-calculated quantitative indicator that reflects the importance of a specific risk transmission pattern in historical experience.

[0044] Finally, the final dynamic attention coefficients are obtained through weighted fusion. :

[0045] in, and There are two weighting coefficients that satisfy... This is used to adjust the importance of information in both the temporal and causal dimensions. This was subsequently used in the information aggregation step of graph neural networks to modulate information from neighboring nodes. Information.

[0046] Through this deeply customized attention mechanism, graph neural networks can prioritize paths with high timeliness and high-risk causal orientation when performing path search and risk scoring on the graph, thereby accurately identifying high-risk transmission chains.

[0047] Finally, the system completes the entire process by establishing a cognitive feedback and graph self-optimization closed loop. This closed loop contains two parallel self-optimization channels. The first channel is weight adjustment: when a high-risk transmission chain with an extremely high comprehensive risk score is identified, the system will trace back along the chain from the target node to the source node, and based on the risk score of the chain and the attention coefficient value of any side of the path, perform a significant nonlinear enhancement on the basic connection weights of these sides.

[0048] In this embodiment, for either side of the high-risk transmission path (from node) Pointing to node Basic connection weights The update rule for nonlinear enhancement adjustments can be expressed by the following formula:

[0049] in, These are the original connection weights. This is the calculated weight enhancement amount. Weight Enhancement Amount The calculation aims to take into account both the overall risk of the path and the individual contribution of each edge:

[0050] In this formula, It is a global learning rate hyperparameter used to control the overall update magnitude. This is a comprehensive risk score for this high-risk transmission path, reflecting the principle that "the higher the risk of the path, the more significant the enhancement." In the deep association analysis step of forward reasoning, for this edge The calculated, normalized dynamic attention coefficient precisely quantifies the contribution of that edge to the formation of this high-risk path. Through this update mechanism, the knowledge graph can reinforce the memory of key risk patterns in a quantitative and interpretable manner.

[0051] This mechanism allows verified high-risk patterns to be reinforced in the graph. The second channel is for new relationship discovery: the system periodically extracts the semantic vectors of all entity nodes and uses a density-based, noisy applied spatial clustering algorithm to find entity clusters that are close in vector space but not yet directly connected in the graph.

[0052] To illustrate this more clearly, the density-based noisy spatial clustering algorithm used in this embodiment will be described in detail. The density-based noisy spatial clustering algorithm is an unsupervised clustering algorithm based on sample distribution density. This invention chose the density-based noisy spatial clustering algorithm because it possesses the following three key advantages: First, the number of potential new relation patterns is unknown. The density-based noisy spatial clustering algorithm does not require pre-specifying the number of clusters and can automatically discover any number of entity clusters based on the data's own distribution, perfectly matching the exploratory discovery requirements of this invention. Second, the distribution of entity vectors representing the same potential relation in space may be of arbitrary shape. The density-based noisy spatial clustering algorithm can identify and divide clusters of arbitrary shapes, ensuring the ability to capture complex patterns. Finally, not all entities participate in potential new relations. The density-based noisy spatial clustering algorithm can effectively identify isolated entities as noise points and exclude them, thereby ensuring the high purity of the discovered entity clusters and improving the accuracy of defining new relation types.

[0053] The purpose of this mechanism is to discover hidden relationships based on deep fire safety logic, rather than superficial textual descriptions. For example, in the field of fire safety, a major fire hazard known as "three-in-one" refers to the illegal combination of production, warehousing, and accommodation within the same building. However, routine inspection documents do not directly label data as "three-in-one," but rather describe it as scattered descriptions like "small production workshops" and "employee dormitories." By learning from a large number of historical cases, the system of this invention can discover in the semantic vector space that the entity vectors of numerous "small production workshops" and "employee dormitories" are always closely clustered together, even though there is no explicit relationship between them in the records. At this point, the system will automatically extract the original textual context of these entity pairs, which may reveal descriptions such as "accommodation on the second floor, production on the first floor" or "fire escape routes are blocked by goods." Based on this context, the system will summarize and generate a new relationship type definition of "suspected mixed production and accommodation use" and push it to experts for review. Once experts confirm that this is a typical "three-in-one" hidden danger pattern, the new relationship is formally added to the basic ontology, thereby greatly enhancing the knowledge graph's ability to proactively identify and warn of such major hidden risks, and truly expanding its cognitive boundaries.

[0054] Example 2: Refer to Figure 2 This is one embodiment of the present invention. In a preferred embodiment of the present invention, a knowledge graph construction and association analysis system for fire supervision and communication full-domain data is provided. The system is logically designed as a cognitive closed loop consisting of three highly collaborative intelligent modules, specifically including: A unified semantic representation engine is configured to receive multi-source heterogeneous data from the field of fire safety and, based on a predefined standardized semantic data structure containing original data fields and multi-dimensional metadata fields, convert the multi-source heterogeneous data into a unified data object stream in real time and without schema. The cognitive graph reasoning engine, connected to the unified semantic representation engine, is configured to continuously extract data objects from the data object stream and perform knowledge construction and association analysis operations. The knowledge construction operation iteratively maintains a fire safety knowledge graph with a hypergraph structure through an entity relation joint extraction model. The association analysis operation performs deep analysis of the fire safety knowledge graph by deploying a graph neural network integrating a temporal causal attention mechanism to output analysis results containing high-risk transmission chains and abnormal association patterns. The graph evolution and ontology management engine is connected to the cognitive graph reasoning engine and is configured to receive analysis results and perform a dual-channel self-optimization operation based on the results. The first channel adjusts the connection weights in the fire protection knowledge graph in reverse according to the high-risk transmission chain, while the second channel automatically generates potential new relation type definitions to expand the basic ontology based on the semantic vectors of entity nodes output by the graph neural network. The three engines work together to form an adaptive cognitive loop from data processing and knowledge reasoning to knowledge evolution.

[0055] The knowledge construction operation of the cognitive graph reasoning engine is accomplished for each data object received from the data object stream by performing the following steps: First, the data object is deeply semantically encoded to generate a context-related word vector representation that can characterize the contextual meaning of each word. Secondly, based on vector representation, the potential entities identified in the data object are used as nodes to dynamically construct a local graph structure, and the features of the nodes are initialized by aggregating the vector representations of the constituent words. Then, a graph neural network decoder is deployed on the local graph structure to update node features by passing information between nodes in multiple rounds, and joint decoding is performed based on the updated node features to directly output event hyperedges that represent complex multi-dimensional relationships between potential entities. Finally, the event hyperedge, along with the confirmed entities and relationships it contains, is integrated into the fire protection knowledge graph as an indivisible transaction unit, thereby ensuring data consistency during the incremental update process of the knowledge graph.

[0056] The graph evolution and ontology management engine is specifically configured to achieve its self-optimization function by executing operations in parallel, including the following first and second channels: The first channel is the connection weight enhancement channel. In this channel, when a high-risk transmission path and its analysis metadata are received from the cognitive graph reasoning engine, the risk target entity node of the path is traced backward, and a weight enhancement factor is calculated for any side of the path. The weight enhancement factor is jointly determined by the comprehensive risk score of the path and the dynamic attention coefficient value of the side in forward reasoning. Then, the weight enhancement factor is applied to update the basic connection weight of the side in the fire protection knowledge graph. The second channel is the new relationship discovery channel. In this channel, the latest semantic vectors of entity nodes are periodically obtained from the cognitive graph reasoning engine. Cluster analysis is performed in the vector space composed of semantic vectors to generate entity node clusters. Clusters that meet the following conditions are selected from the entity node clusters: the nodes inside are semantically similar, but are not directly connected in the topology of the fire protection knowledge graph. Then, the common context information of entity pairs in the selected clusters is automatically extracted to summarize and generate semantic descriptions of potential new relationship types, and these are submitted to the review module as ontology evolution proposals.

[0057] These modules can be physically deployed centrally or distributed, and are decoupled and communicated through asynchronous message queues, ensuring high availability and scalability of the system. To illustrate this more clearly, the asynchronous message queue technology used in this system is explained. This is an advanced mechanism for achieving asynchronous communication by establishing a shared data buffer between engine modules. In this invention, this mechanism, by achieving loose coupling between modules, not only caches data to ensure system stability and fault tolerance when module processing speeds are mismatched or there are brief failures, but also allows for parallel data processing by adding module instances to achieve elastic scaling.

[0058] The system's data access and semantic unification module serves as the data entry point and preprocessing hub for the entire cognitive loop. This module incorporates pluggable data adapters for different data sources (such as relational databases, document servers, and IoT gateways). When external data flows in, the adapter first loads it, followed by the activation of a core semantic mapping processor. This processor, based on a predefined fire protection domain ontology conceptual model using JSON-based linked data, forcibly transforms diverse input data into a standardized, metadata-rich intermediate object representation. This process not only unifies the data format but, more importantly, injects a clear, machine-readable semantic context into each piece of raw data, eliminating ambiguity caused by data heterogeneity at the source and providing high-quality "digital nourishment" for subsequent intelligent processing.

[0059] Next, the knowledge reasoning and graph construction module, as the cognitive core of the system, continuously processes the standardized data objects delivered by the data access and semantic unification module. The core of the knowledge reasoning and graph construction module is an innovative knowledge extraction workflow based on graph reasoning. For each data object, the workflow first uses a deep language model based on a self-attention network architecture (Transformer) to transform its text content into high-dimensional context-aware word vectors. Then, a dynamic graph builder materializes the initially identified entity references in the text into nodes, instantly generating a temporary micro-graph network containing only current context information, and initializing the node states with word vectors. Subsequently, a graph attention network performs reasoning on this micro-graph, simulating information interactions between nodes to understand their deep connections, ultimately decoding a complex multi-faceted event holistically into one or more event hyperedge structures. Finally, a transaction manager atomically incorporates these newly discovered knowledge into a global fire safety knowledge graph hosted by a high-performance graph database (such as Neo4j or Janus Graph) in a manner that ensures data consistency. Simultaneously, this module also runs a global analysis task in parallel, using a deeply customized graph neural network to perform risk transmission analysis and anomaly pattern scanning on the entire knowledge graph, and sends the analytical insights to the next module.

[0060] Finally, the Knowledge Evolution and Ontology Self-Optimization module, acting as the system's "evolutionary engine," is responsible for receiving and digesting knowledge insights from the reasoning module to drive the self-improvement of the entire knowledge system. This module contains two parallel feedback loops. The first is the "Memory Reinforcement" loop: upon receiving an analysis report of a high-risk transmission chain, this loop immediately initiates a weighted backpropagation task, tracing the path along the chain and significantly increasing the weight values ​​of all connections on the path based on the chain's danger level and its attention contribution during the reasoning process. This is equivalent to deepening the "memory" of the dangerous pattern in the neural network structure of the knowledge graph. The second is the "Cognitive Boundary Expansion" loop: this loop periodically analyzes the global semantic vectors of all entity nodes, using a density-based clustering algorithm to identify a special type of entity cluster: whose internal entity nodes have a distance of less than a preset threshold in the semantic vector space, but are not directly connected by edges in the topology of the knowledge graph. Once such a cluster is discovered, the system traces back to its original context, automatically summarizes and generates a hypothesis about a potential new relationship, and submits it to domain experts for review. This mechanism enables the system's knowledge ontology to continuously generate new concepts from the data, achieving genuine cognitive growth.

[0061] Example 3: Reference Figure 3This is one embodiment of the present invention. This embodiment provides a specific application scenario for a knowledge graph construction and association analysis method for fire supervision and communication full-domain data. This scenario is specifically designed for in-depth, dynamic risk inspections and hidden risk discovery of key fire safety units. In this scenario, the system's deep association analysis function has been further specialized.

[0062] In this embodiment, the system first focuses on entity nodes in the fire safety knowledge graph marked as "high-risk locations" or "key fire safety units," along with all surrounding entities and relationships directly associated with or indirectly associated with this core entity through an intermediate entity, forming a high-risk subgraph. Subsequently, the system calls a module specifically designed for anomaly association detection to analyze this subgraph. The core of this module is an anomaly detection model based on graph contrastive learning, which can serve as a superior alternative to the graph autoencoder model in Embodiment 1.

[0063] Specifically, this graph-contrastive learning model learns from a massive amount of high-risk subgraphs under normal conditions from different units, acquiring a compact and unique vector representation, or "structural prototype vector," for each type of normal subgraph structure (e.g., a knowledge subgraph describing the state of a compliant hazardous materials warehouse). During online detection, the system inputs the real-time knowledge subgraph of the key unit to be detected (e.g., warehouse A) into the model, similarly generating a real-time vector representation. Subsequently, the system calculates the distance between this real-time vector and all known "structural prototype vectors" in the vector space.

[0064] If the distance between the real-time vector and the nearest "structural prototype vector" (e.g., the prototype vector of a compliant hazardous chemicals warehouse) significantly exceeds the normal fluctuation threshold determined by historical statistical data, the system determines that the current state of unit A has a structural anomaly. In this case, the system not only outputs a general "anomaly" alert but also further identifies the specific node or edge causing the largest vector representation shift through attribution analysis. For example, the system might find that the overall structural shift is caused by the sudden appearance of a new node pointing to "temporary worker dormitory" in the subgraph of warehouse A, and the abnormally shortened path distance between the "hazardous chemicals" node and the "domestic ignition source" node. In this way, the system can accurately locate and report to regulators in natural language: "Warehouse A is suspected of having newly added illegal occupancy, and the safety distance between hazardous chemicals and ignition sources is insufficient, posing a significant safety hazard." This achieves accurate and interpretable automatic detection of highly concealed and dynamically changing complex risks.

[0065] In summary, this invention achieves the following: First, by unifying semantic representation, it completely breaks down data barriers between various fire protection systems, realizing the first-ever deep fusion of multi-modal data across the entire domain, providing a solid foundation for global security situation awareness. Second, by utilizing event hyperedges and deeply customized graph neural networks, it significantly enhances the depth of risk analysis, automatically discovering hidden, complex risk transmission chains and abnormal correlation patterns that traditional methods cannot reach, elevating risk identification from judging isolated hazard points to accurately understanding dynamic risk networks. Third, its unique cognitive feedback and self-optimizing closed loop endow the system with the ability to learn from analysis results and achieve self-evolution of its knowledge structure. This "growthable" cognitive characteristic powerfully propels fire safety management from a passive "post-event response" mode to a proactive "pre-event warning and precise decision-making" mode, significantly improving the accuracy of fire risk prediction and the scientific level of emergency command.

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing and analyzing knowledge graphs for fire supervision and communication data across the entire domain, characterized in that, Includes the following steps: Step 1: Format the multi-source heterogeneous data from the fire safety field into a unified data object stream, where each data object contains metadata describing its source, time, and credibility; Step 2: Initialize the basic ontology containing core fire protection concepts, and based on the data object stream, use the entity relationship joint extraction model to synchronously identify entities, multi-relationships and event hyperedges from the data object stream, so as to iteratively and incrementally construct a fire protection knowledge graph with a hypergraph structure; Step 3: Deploy a graph neural network that integrates a temporal causal attention mechanism to perform deep association analysis on the fire protection knowledge graph. This attention mechanism dynamically weights the paths in the fire protection knowledge graph based on the timeliness of events and the causal patterns of historical fires. Step 4: Establish a cognitive feedback and graph self-optimization closed loop. The deep analysis results obtained by the graph neural network are used in reverse to the fire protection knowledge graph. By automatically adjusting the connection weights between entities in the fire protection knowledge graph and proposing new relation types to the basic ontology, the dynamic self-optimization and continuous evolution of the fire protection knowledge graph are realized.

2. The knowledge graph construction and association analysis method for fire supervision and communication full-domain data as described in claim 1, characterized in that, In step 1, the step of formatting multi-source heterogeneous data from the field of fire safety into a unified data object stream specifically includes: pre-defining a standardized semantic data structure, which consists of original data fields and associated multi-dimensional metadata fields, and parsing and mapping the multi-source heterogeneous data to the semantic data structure to generate data objects that constitute the data object stream; The multidimensional metadata fields include at least: source metadata to characterize the data source and collection time, quality assessment metadata to characterize the data reliability and accuracy, and preprocessed semantic metadata extracted from the multi-source heterogeneous data through a preliminary analysis model.

3. The knowledge graph construction and association analysis method for fire supervision and communication full-domain data as described in claim 1, characterized in that, In step 2, the working method of the entity relationship joint extraction model includes: using a deep semantic encoder to encode the data in the data object stream to generate context-related word vector representations that can characterize context semantics; Based on the context-related word vector representation, the potential entities identified in the data are used as nodes and dynamically constructed into a local graph structure. The features of the nodes are initialized by aggregating the vector representations of the words they constitute. A graph neural network decoder is deployed to update node features through multi-round information transmission between nodes on the local graph structure, and to perform joint decoding based on the updated node features to directly output event hyperedges that represent the complex multi-dimensional relationships between the potential entities.

4. The knowledge graph construction and association analysis method for fire supervision and communication full-domain data as described in claim 1, characterized in that, In step 3, the attention mechanism dynamically weights the paths in the fire protection knowledge graph by: calculating a dynamic attention coefficient for any side between nodes in the fire protection knowledge graph to adjust the intensity of information transmission. The dynamic attention coefficient is formed by weighted fusion of temporal attention components and causal attention components. The value of the temporal attention component decreases as the time interval between the events associated with the edge increases, while the value of the causal attention component increases as the historical risk level of the association type represented by the edge increases.

5. The method for knowledge graph construction and association analysis of fire supervision and communication full-domain data as described in claim 1, characterized in that, In step 4, the step of automatically adjusting the connection weights between entities in the fire protection knowledge graph includes: identifying the transmission path connecting the risk source entity node and the risk target entity node in the fire protection knowledge graph, calculating the comprehensive risk score of the transmission path, and determining the transmission path as a high-risk transmission path when the comprehensive risk score is higher than a preset threshold. Starting from the risk target entity node, trace back along the high-risk transmission path, and perform non-linear enhancement adjustments on the connection weights represented by each edge constituting the high-risk transmission path. The magnitude of the enhancement adjustment for any side is positively correlated not only with the comprehensive risk score of the high-risk transmission path, but also with the dynamic attention coefficient value corresponding to the side.

6. The method for knowledge graph construction and association analysis of fire supervision and communication full-domain data as described in claim 1, characterized in that, In step 4, the step of proposing a new relation type to the basic ontology includes: obtaining the semantic vector of each entity node output by the graph neural network after deep association analysis, wherein the semantic vector integrates the structural information of the node in its neighborhood; In the vector space composed of the semantic vectors, cluster analysis is used to identify entity node clusters whose distance between vectors is less than a preset threshold, but which have no direct edge connection in the fire protection knowledge graph. The common context information of entity pairs within the entity node cluster in the multi-source heterogeneous data is automatically extracted, and based on the common context information, a semantic description of potential new relationship types is summarized and generated.

7. The method for knowledge graph construction and association analysis of fire supervision and communication full-domain data as described in claim 1, characterized in that, The deep association analysis in step 3 further includes an abnormal association detection step, which includes: extracting subgraphs that conform to a preset conventional pattern from the fire knowledge graph as conventional pattern subgraph samples, and training a graph autoencoder model composed of a graph encoder and a graph decoder based on the conventional pattern subgraph samples. The subgraph structure to be detected in the fire protection knowledge graph is input into the graph encoder to generate a low-dimensional latent vector representing the subgraph structure. The low-dimensional latent vector is then input into the graph decoder to generate a reconstructed subgraph. The reconstruction error is determined by calculating the comprehensive difference between the subgraph structure to be detected and the reconstructed subgraph in terms of topology and node attributes. When the reconstruction error is greater than the normal error baseline obtained from the statistical analysis of the conventional pattern subgraph samples, the subgraph to be detected is determined to contain abnormal correlation.

8. A knowledge graph construction and association analysis system for fire supervision and communication full-domain data, based on the knowledge graph construction and association analysis method for fire supervision and communication full-domain data as described in any one of claims 1 to 7, characterized in that, include: A unified semantic representation engine is configured to receive multi-source heterogeneous data from the field of fire safety and, based on a predefined standardized semantic data structure containing original data fields and multi-dimensional metadata fields, convert the multi-source heterogeneous data into a unified data object stream in real time and without schema. A cognitive graph reasoning engine, connected to the unified semantic representation engine, is configured to: continuously acquire data objects from the data object stream and perform knowledge construction and association analysis operations; wherein, the knowledge construction operation iteratively maintains a fire protection knowledge graph with a hypergraph structure through an entity relation joint extraction model; the association analysis operation performs deep analysis on the fire protection knowledge graph by deploying a graph neural network integrating a temporal causal attention mechanism to output analysis results containing high-risk transmission chains and abnormal association patterns; The graph evolution and ontology management engine is connected to the cognitive graph reasoning engine and is configured to receive the analysis results and perform a dual-channel self-optimization operation based on the results. The first channel adjusts the connection weights in the fire protection knowledge graph in reverse according to the high-risk transmission chain, while the second channel automatically generates potential new relation type definitions to expand the basic ontology based on the entity node semantic vectors output by the graph neural network. The three engines work together to form an adaptive cognitive loop from data processing and knowledge reasoning to knowledge evolution.

9. The knowledge graph construction and association analysis system for fire supervision and communication full-domain data as described in claim 8, characterized in that, The knowledge construction operation of the cognitive graph reasoning engine is performed on each data object received from the data object stream by executing the following steps: First, the data object is subjected to deep semantic encoding to generate a context-related word vector representation that can characterize the contextual meaning of each word. Secondly, based on the vector representation, the potential entities identified in the data object are used as nodes to dynamically construct a local graph structure, and the features of the nodes are initialized by aggregating the vector representations of their constituent words. Then, a graph neural network decoder is deployed on the local graph structure to update node features by performing multiple rounds of information transmission between nodes, and joint decoding is performed based on the updated node features to directly output event hyperedges that characterize the complex multi-dimensional relationships between the potential entities. Finally, the event hyperedge, along with the confirmed entities and relationships contained therein, is integrated into the fire protection knowledge graph as an indivisible transaction unit, thereby ensuring data consistency of the knowledge graph during incremental updates.

10. The knowledge graph construction and association analysis system for fire supervision and communication full-domain data as described in claim 8, characterized in that, The graph evolution and ontology management engine is specifically configured to achieve its self-optimization function by executing operations in parallel, including the following first and second channels: The first channel is a connection weight enhancement channel. In this channel, when a high-risk transmission path and its analysis metadata are received from the cognitive graph inference engine, a reverse tracing is performed starting from the risk target entity node of the path, and a weight enhancement factor is calculated for any side of the path. The weight enhancement factor is jointly determined by the comprehensive risk score of the path and the dynamic attention coefficient value of the side in forward inference. Then, the weight enhancement factor is applied to update the basic connection weight of the side in the fire protection knowledge graph. The second channel is a new relationship discovery channel. In this channel, the latest semantic vectors of entity nodes are periodically obtained from the cognitive graph reasoning engine. Cluster analysis is performed in the vector space composed of the semantic vectors to generate entity node clusters. Clusters that meet the following conditions are selected from the entity node clusters: the nodes inside are semantically similar, but are not directly connected in the topology of the fire protection knowledge graph. Then, the common context information of entity pairs in the selected clusters is automatically extracted to summarize and generate semantic descriptions of potential new relationship types, and these are submitted to the review module as ontology evolution proposals.

Images