A knowledge graph-based security control method

Abstract

The application discloses a kind of safety control methods based on knowledge graph, belong to safety control field, this method includes the following steps: by crawler program according to keyword to crawl multiple types of data and construct fireproof database;Using natural language processing technology, the data in fireproof database is implemented semantic analysis, and the fireproof database after semantic analysis is carried out data processing by word vector modeling and entity relationship extraction;Using the processed fireproof database constructs internal large language model, adopts Leiden technology to the internal large language model and executes hierarchical clustering;Based on hierarchical clustering result in ancient building space feature, cultural relic feature and disaster-causing factor, establish multidimensional mapping model, based on the multidimensional mapping model, use knowledge graph technology to construct multimodal dynamic correlation fire knowledge graph;Based on user query keyword, dynamically select search mode, based on the search mode, according to the query keyword generates safety control result.

Description

Technical Field

[0001] This invention relates to the field of security control technology, and in particular to a security control method based on knowledge graphs. Background Technology

[0002] my country boasts a rich cultural heritage, including vast collections of cultural relics and precious historical sites. Fires in these historical buildings present high risks, significant hidden dangers, devastating impacts, and significant firefighting challenges. The enormous cost of monitoring and maintenance is a major pain point in the management of these cultural heritage sites, as a fire inevitably leads to irreparable losses. Traditional fire prevention methods have many limitations, making it difficult to achieve precise control and rapid response when dealing with complex and ever-changing fire hazards. With the further advancement of cultural heritage protection efforts, the research, development, and application demonstration of intelligent active fire prevention systems for historical buildings are becoming increasingly important.

[0003] A knowledge graph is a series of graphs that display the development process and structural relationships of knowledge. It mimics the way human structured knowledge is composed through a network structure of nodes (entities) and edges (relationships). Knowledge graphs can capture not only raw information but also high-order relationships across multiple documents and possess powerful reasoning capabilities. They can be widely applied in various industries, including fire prevention for cultural relics and ancient buildings. Based on the above, the problem this invention aims to solve is as follows:

[0004] 1. Traditional fire prevention methods have limitations. Ancient buildings have multiple wooden components, and fire hazards are complex and varied. Current technology is insufficient to meet the fire prevention needs of cultural heritage buildings.

[0005] 2. Ancient buildings and cultural relics present a challenge in accurately locating fire sources, making it difficult to control fires in a timely manner;

[0006] 3. The protection of cultural relics requires adherence to the principle of minimal intervention in the prevention and control of fires in cultural relics buildings, and it is necessary to minimize human intervention in cultural relics and ancient buildings.

[0007] 4. Cultural relics and ancient buildings are easily affected by various disaster-causing factors such as weather, building materials, building structure, and historical restoration, making it difficult to comprehensively assess the impact level and fire-fighting plan. Summary of the Invention

[0008] To address the problems mentioned above, this invention provides a security control method based on knowledge graphs.

[0009] A knowledge graph-based security control method includes the following steps:

[0010] A fire prevention database was constructed by using a web crawler to collect various types of data on ancient building fire prevention, cultural relic fire prevention, fire accidents, and material coefficients based on keywords.

[0011] Natural language processing technology is used to perform semantic parsing on the data in the fire protection database. The fire protection database after semantic parsing is then processed through word vector modeling and entity relation extraction.

[0012] The processed fire protection database was used to build an internal large language model, and the Leiden technique was used to perform hierarchical clustering on the internal large language model.

[0013] A multi-dimensional mapping model is established based on the spatial characteristics of ancient buildings, cultural relics characteristics and disaster-causing factors in the hierarchical clustering results. Based on the multi-dimensional mapping model, a multi-modal dynamic association fire prevention knowledge graph is constructed using knowledge graph technology.

[0014] Based on the user's query keywords, a search mode is dynamically selected, and based on the search mode, a security control result is generated according to the query keywords.

[0015] Preferably, the step of constructing a fire prevention database by crawling various types of data, including ancient building fire prevention, cultural relic fire prevention, fire accidents, and material coefficients, based on keywords using a web crawler program includes:

[0016] It was determined that multiple data sources could be obtained, including fire prevention standards for ancient buildings, fire prevention standards for cultural relics, fire accident data, and fire resistance coefficients of materials.

[0017] Based on the special characteristics of fire prevention for ancient buildings and cultural relics, a multi-level keyword database was constructed.

[0018] The web crawler program retrieves various types of data containing keywords from multiple data sources based on the multi-level keyword database.

[0019] The data of the various types is removed by comparing hash values ​​and then stored in the database, retaining the data source and the timestamp of the crawling.

[0020] The specific parameters of the various types of data in the database are reviewed by experts, and the database is identified as a fire prevention database based on the reliability level of the data source.

[0021] Preferably, the step of using natural language processing technology to perform semantic parsing on the data in the fire protection database, and processing the semantically parsed fire protection database through word vector modeling and entity relation extraction, includes:

[0022] AI technology is used to standardize various types of data in the fire prevention database, converting the data in the database into natural language data.

[0023] The natural language corpus data is segmented into text units, and the text units are sorted according to their relevance and distributed from near to far in the concept vector space.

[0024] The thesaurus is automatically expanded using word vector similarity retrieval, the relevance weights of the synonyms are adjusted, and the distribution of the text units is adjusted based on the weight adjustment results.

[0025] Entities are extracted from the text units, the entities are classified, and relationships are established between each entity category to generate a structured data system.

[0026] Preferably, the step of constructing an internal large-scale language model using the processed fire protection database and performing hierarchical clustering on the internal large-scale language model using Leiden technology includes:

[0027] Obtain the structured data system in the preprocessed fire protection database, and determine the model architecture of the large language model based on the characteristics of the text units in the structured data system;

[0028] The preprocessed structured data is mixed with the original text and input into a large language model to mask the entities in the fire protection database, enabling the model to learn and determine the domain knowledge associations.

[0029] The parameters of the large language model are abstracted into graph nodes, and the weights between nodes are determined by parameter similarity, thus generating a high-dimensional parameter relationship structure.

[0030] The Leiden algorithm is applied to detect the high-dimensional parameter relationship structure, and highly correlated internal groups are merged to form basic functional modules.

[0031] Preferably, the application of the Leiden algorithm to detect the high-dimensional parameter relationship structure, merging highly correlated internal groups to form basic functional modules, includes:

[0032] The internal parameters of the language model are abstracted into graph nodes, which contain the term vector table weight distribution and entity relations of the text unit;

[0033] The correlation of the internal parameters of the language model is calculated by cosine similarity, and the edge weights are defined based on the correlation of the internal parameters of the language model.

[0034] The Leiden algorithm is used to assign each node to an independent community. The nodes are traversed, and the modularity gain when each independent community is moved to an adjacent independent community is calculated. The direction with the maximum gain is selected to iteratively adjust the affiliation of the independent community, and the community set of the independent community is generated.

[0035] If the edge weight between two community sets exceeds a preset threshold, they are merged into the same functional module, and the same functional module is determined as the basic functional module.

[0036] Extract the high-frequency text units of the community set within the basic functional modules, match them with standard functional descriptions, and generate module tags.

[0037] Preferably, the algorithm used to assign each node to an independent community using the Leiden algorithm, traverse the nodes, calculate the modularity gain when moving each independent community to an adjacent independent community, select the direction with the maximum gain to iteratively adjust the affiliation of independent communities, and generate the community set of the independent communities is as follows:

[0038]

[0039] in, For edge weights, For node degree, This represents the total edge weight. The algorithm is used to determine whether nodes belong to the same independent community. It measures the difference between the connection density within the independent community and the random distribution, and determines the specific modularity gain based on the difference.

[0040] Preferably, the step of establishing a multi-dimensional mapping model based on the spatial features of ancient buildings, cultural relic features, and disaster-causing factors in the hierarchical clustering results, and constructing a multi-modal dynamic association fire prevention knowledge graph based on the multi-dimensional mapping model using knowledge graph technology, includes:

[0041] The hidden representations of the spatial features of the ancient buildings and the features of the cultural relics are extracted from the basic functional modules, and standardized feature vectors are generated through graph embedding technology.

[0042] The clustering results of disaster-causing factors are extracted from the basic functional modules, and the clustering results are output as a probability distribution or risk score vector, while retaining the correlation weight between the clustering results and spatial features.

[0043] A multi-branch neural network is constructed to process the feature vector and the catastrophic factor respectively, realizing cross-modal feature interaction. The contribution of different catastrophic factors to the feature vector is dynamically adjusted through an attention mechanism to generate a multi-dimensional model.

[0044] The output of the multi-dimensional model is mapped to graph node attributes. A knowledge graph is built based on the graph nodes. The risk evolution path is modeled and analyzed through the knowledge graph. Association rules are applied to mine potential risk coupling relationships.

[0045] Based on the risk evolution path and the potential risk coupling relationship, a multimodal dynamic association fire prevention knowledge graph is constructed according to the knowledge graph.

[0046] Preferably, the step of dynamically selecting a search mode based on a user query and generating a security control result based on the search mode includes:

[0047] Receive user query dynamics, perform search pattern matching on the user query dynamics, and determine whether the question in the user query dynamics is a holistic question through an intent classification model. If the determination result is no, the first query result is directly generated based on the keywords in the question through the multimodal dynamic association fire protection knowledge graph.

[0048] The user query dynamics are determined to be a holistic question. Multiple batch node reports are generated based on the holistic question by searching the knowledge graph of the multimodal dynamic association fire prevention knowledge graph using the internal large-scale language model.

[0049] The node report is divided into text blocks of a predefined size. Each text block generates an intermediate response, which contains key information fragments about fire prevention of cultural relics. Each key information fragment is assigned a weight coefficient based on its relevance to the overall issue.

[0050] The key information fragments are ranked according to their weight coefficients, and the key information fragments below a preset threshold are filtered to form an aggregated intermediate response. The aggregated intermediate response is then used to generate a second query result.

[0051] The first query result and the second query result are determined as security control results.

[0052] Preferably, the method further includes combining the multimodal dynamic association fire prevention knowledge graph with the cultural relics and ancient buildings information model, dynamically labeling risk information in the three-dimensional model, and automatically generating a cultural relics protection priority list and protection plan based on the risk information, including:

[0053] High-precision geometric models of ancient buildings are reconstructed using lidar point cloud data, preserving detailed features such as brackets, beams, and painted decorations. A digital twin containing material properties, structural parameters, and historical repair records is established, and the model is decomposed into component-level units and bound with data tags.

[0054] Align the ancient building entities and ancient building models in the multimodal dynamic fire prevention knowledge graph with spatial coordinates, establish a two-way index relationship, and synchronize the fire risk data and cultural relic status data in the multimodal dynamic fire prevention knowledge graph to the component attribute panel of each component-level unit in real time.

[0055] Link historical inspection reports and maintenance records in the component properties panel to generate a time-series variation curve;

[0056] A risk quantification index system is constructed based on building risk, environmental disaster risk, and cultural relic vulnerability indicators. The risk weight of each indicator in the risk quantification index system and the correlation coefficient between indicators are determined using the multimodal dynamic association fire prevention knowledge graph.

[0057] Based on the aforementioned risk quantification index system, a risk quantification assessment is performed on each of the aforementioned component-level units to generate a risk label and risk level for each component-level unit;

[0058] Based on the corresponding components of the risk label in the 3D model, risk information is dynamically labeled, and the components are color-rendered based on the risk level.

[0059] Based on the risk label and risk level of each component-level unit, an urgency assessment is performed in the multimodal dynamic association fire prevention knowledge graph, and a comprehensive score is calculated based on the urgency assessment results.

[0060] The priority ranking of each component-level unit is determined based on the comprehensive score of each component-level unit, and a priority list for cultural relic protection is generated based on the priority ranking.

[0061] Based on the risk label of each component-level unit, protective measures are retrieved in the multimodal dynamic association fire prevention knowledge graph, and protective schemes for the corresponding components are generated based on the retrieval results.

[0062] Preferably, the method further includes updating the status of cultural relics and ancient buildings in real time through IoT sensing devices, and dynamically generating the best safety response plan based on visitor flow data, including:

[0063] The multimodal dynamic association fire prevention knowledge graph is used to determine environmental monitoring nodes, structural inspection nodes, and visitor monitoring nodes for cultural relics and ancient buildings;

[0064] Temperature and humidity sensors, smoke detectors, and ultraviolet sensors are deployed at the environmental monitoring nodes of the cultural relics and ancient buildings to collect environmental data of the cultural relics and ancient buildings.

[0065] Strain gauges, inclinometers, and vibration sensors are deployed at the structural detection nodes to collect structural data of cultural relics and ancient buildings;

[0066] Infrared beam detectors and smart cameras equipped with behavior recognition algorithms are deployed at the tourist monitoring nodes to collect tourist flow data and tourist behavior data in the scenic area.

[0067] The ticketing system is used to obtain the current number of visitors and today's reservation data, and the future visitor flow trend is predicted based on the visitor flow data.

[0068] The environmental data, the structural data, the tourist flow data, and the tourist behavior data are unified to a standard timestamp;

[0069] The current environmental data and structural data are input into the multimodal dynamic association fire prevention knowledge graph to determine the real-time risk status of cultural relics and ancient buildings. Based on the tourist flow data and future tourist flow trends, the risk weight of tourist-intensive areas is adjusted to generate a current risk report for cultural relics and ancient buildings.

[0070] Risk identification is performed on tourist behavior data, and when high-risk behaviors are detected in the tourist behavior data, a human-caused disaster warning is generated in real time.

[0071] Based on the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters, the current risk level is determined. When the risk level exceeds the preset threshold, a risk warning is issued on the monitoring display port.

[0072] Based on a multi-constraint decision-making model, the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters are used to generate the best safety disposal plan.

[0073] Through the above-mentioned technical means, the present invention achieves the following beneficial effects:

[0074] 1. Based on knowledge graph technology, integrate and associate the possible causes of fires in cultural relics and ancient buildings to form a knowledge graph in the field of fire safety of cultural relics and establish a knowledge base of emergency response plans for safety scenarios of cultural relics and ancient buildings. Investigate and solve fire hazards in cultural relics and ancient buildings, and meet the fire prevention needs of cultural relics and ancient buildings based on new technologies.

[0075] 2. By updating the status of cultural relics and ancient buildings in real time through IoT sensing devices, the problem of accurately locating fire points in ancient buildings and cultural relics can be solved. Combined with tourist flow data, the best safety response plan can be dynamically generated to control the fire at the first moment.

[0076] 3. By leveraging an emergency response knowledge base and rapid response technology, a swift response can be made when a fire occurs. Multi-type sensor network monitoring and operation and maintenance technology ensure real-time monitoring, laying the groundwork for post-fire handling procedures for cultural heritage buildings, meeting the principle of minimal intervention in the prevention and handling of fires in cultural heritage buildings, and minimizing human intervention in cultural heritage buildings.

[0077] 4. In the field of fire safety of cultural heritage buildings, construct a comprehensive and accurate knowledge graph, and realize the comprehensive assessment of the impact level and fire extinguishing plan based on various disaster-causing factors such as weather, building materials, building structure, and historical restoration to meet the fire prevention needs of cultural heritage buildings, and complete the knowledge retrieval and fire prevention plan generation in related fields.

[0078] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0079] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0080] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0081] Figure 1 A schematic diagram illustrating a knowledge graph-based security control method provided by the present invention;

[0082] Figure 2 Another schematic diagram of a knowledge graph-based security control method provided by the present invention;

[0083] Figure 3 This is another schematic diagram of a knowledge graph-based security control method provided by the present invention;

[0084] Figure 4 A schematic diagram of the knowledge graph construction process for a knowledge graph-based security control method provided by the present invention;

[0085] Figure 5 A schematic diagram of the auxiliary decision-making software interface for a knowledge graph-based security control method provided by the present invention;

[0086] Figure 6 This is a schematic diagram of the visitor behavior recognition software interface for a knowledge graph-based security control method provided by the present invention. Detailed Implementation

[0087] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0088] my country possesses a rich cultural heritage, including vast cultural relics and precious historical sites. Fires in these historical buildings present high risks, significant hidden dangers, devastating damage, and significant firefighting challenges. The enormous cost of monitoring and maintenance is a major pain point in the management of these cultural relic protection units, as a fire inevitably leads to irreparable losses. Traditional fire prevention methods have many limitations, making it difficult to achieve precise control and rapid response when dealing with complex and ever-changing fire hazards. With the further advancement of cultural relic protection work, the importance of researching and demonstrating intelligent active fire prevention systems for historical buildings is becoming increasingly prominent. Knowledge graphs are a series of graphs that display the development process and structural relationships of knowledge. They mimic the composition of human structural knowledge through a network structure of nodes (entities) and edges (relationships). Knowledge graphs can capture not only raw information but also high-order relationships across multiple documents and possess powerful reasoning capabilities, making them widely applicable in various industries, including the fire prevention of historical buildings. Based on the above, the problems that this invention aims to solve are as follows:

[0089] 1. Traditional fire prevention methods have limitations. Ancient buildings have multiple wooden components, and fire hazards are complex and varied. Current technology is insufficient to meet the fire prevention needs of cultural heritage buildings.

[0090] 2. Ancient buildings and cultural relics present a challenge in accurately locating fire sources, making it difficult to control fires in a timely manner;

[0091] 3. The protection of cultural relics requires adherence to the principle of minimal intervention in the prevention and control of fires in cultural relics buildings, and it is necessary to minimize human intervention in cultural relics and ancient buildings.

[0092] 4. Cultural relics and ancient buildings are easily affected by various disaster-causing factors such as weather, building materials, building structure, and historical restoration, making it difficult to comprehensively assess the impact level and fire-fighting plan.

[0093] A knowledge graph-based security control method, such as Figure 1 As shown, it includes the following steps:

[0094] Step S101: Use a web crawler to crawl various types of data on ancient building fire prevention, cultural relic fire prevention, fire accidents, and material coefficients based on keywords to build a fire prevention database;

[0095] Step S102: Use natural language processing technology to perform semantic parsing on the data in the fire protection database, and process the fire protection database after semantic parsing by word vector modeling and entity relation extraction;

[0096] In some embodiments, this embodiment uses natural language processing technology to perform deep semantic analysis on unstructured text data in the fire protection database. It employs a pre-trained language model combined with a domain-adaptive entity recognition algorithm to extract key information such as building parameters, material properties, and fire protection codes. It constructs a semantic representation space for fire protection professional terms based on word vector modeling, uses an attention mechanism to capture the contextual features of concepts such as "fire resistance limit" and "fire separation distance," and establishes triplet relationships between building components, fire protection facilities, and disaster-causing factors through entity relationship extraction technology.

[0097] Step S103: Construct an internal large-scale language model using the processed fire protection database, and perform hierarchical clustering on the internal large-scale language model using Leiden technology;

[0098] In some embodiments, this embodiment uses a structured fire protection database as training corpus, adopts the Transformer architecture to build a pre-trained language model, learns deep semantic associations of professional knowledge such as building parameters and fire protection codes through tasks such as masked language modeling, uses the Leiden algorithm to perform hierarchical clustering on the hidden parameters of the model, and optimizes the model structure and inference efficiency by pruning redundant units and strengthening the dynamic paths between modules, and finally forms an interpretable and easily extensible domain-specific model to support fire correlation analysis and safety decision generation.

[0099] Step S104: Based on the spatial features of ancient buildings, cultural relics features and disaster-causing factors in the hierarchical clustering results, establish a multi-dimensional mapping model, and use knowledge graph technology to construct a multi-modal dynamic association fire prevention knowledge graph based on the multi-dimensional mapping model;

[0100] In some embodiments, this embodiment constructs a cross-modal feature mapping network based on the clustering results of ancient building spatial characteristics, cultural relic attributes and disaster-causing factors. Heterogeneous data is encoded into a unified vector space through graph embedding algorithm. An attention mechanism is designed to dynamically fuse the correlation weights of structural risk, cultural relic vulnerability and disaster probability. Graph neural networks are used to realize functions such as fire evolution prediction and risk path tracing. The coupling relationship between protective measures and risk status is dynamically updated through time series graphs to form an interpretable and reasonable intelligent fire prevention decision center.

[0101] Step S105: Dynamically select a search mode based on the user's query keywords, and generate a security control result based on the query keywords according to the search mode.

[0102] In some embodiments, this embodiment uses a domain-adaptive semantic parsing model to identify the intent and extract entities from user query keywords, matching user questions to global search and specified search patterns respectively. Global search aims to handle holistic questions involving the entire cultural relic fire prevention data corpus, such as "Which ancient buildings are most prone to fire?" or "What are the common fire prevention measures for cultural relics?". Specified search focuses on handling queries with specific keywords (such as a certain ancient building or a certain fire prevention measure), such as "What are the fire prevention measures for Jinci?" or "How to protect wooden ancient buildings from fire?".

[0103] The knowledge graph construction process is as follows: Figure 4 As shown.

[0104] The working principle of the above technical solution is as follows: First, a fire prevention database is constructed by crawling various types of data through a web crawler; second, semantic parsing and data processing are performed on the data in the fire prevention database; third, the processed fire prevention database is used to construct an internal large-scale language model, perform hierarchical clustering, and establish a multi-dimensional mapping model based on the spatial characteristics of ancient buildings, cultural relics characteristics, and disaster-causing factors in the hierarchical clustering results, and construct a multimodal dynamic association fire prevention knowledge graph using knowledge graph technology; finally, based on user query keywords, the search mode is dynamically selected, and safety control results are generated according to the query keywords.

[0105] The beneficial effects of the above technical solution are as follows: First, by crawling various types of data through a web crawler to construct a fire prevention database, multiple types of data can be obtained from multiple data sources, improving data quality. Furthermore, semantic parsing and data processing are performed on the data in the fire prevention database, which unifies the database data format and facilitates the construction of a large language model. Further, the processed fire prevention database is used to construct an internal large-scale language model, performing hierarchical clustering. Based on the spatial characteristics of ancient buildings, cultural relics, and disaster-causing factors in the hierarchical clustering results, a multi-dimensional mapping model is established. Knowledge graph technology is used to construct a multimodal, dynamically related fire prevention knowledge graph. A large language model can be constructed based on the database, and the knowledge graph can be generated using the large language model, making the knowledge graph practical and relevant. Furthermore, based on user query keywords, the search mode can be dynamically selected, and safety control results can be generated based on the query keywords, solving the problem of users obtaining effective and feasible safety control solutions based on knowledge graphs.

[0106] In one embodiment, such as Figure 2 As shown, the fire prevention database is constructed by using a web crawler to collect various types of data, including fire prevention of ancient buildings, fire prevention of cultural relics, fire accidents, and material coefficients, based on keywords.

[0107] Step S201: Identify multiple data sources from which fire prevention standards for ancient buildings, fire prevention standards for cultural relics, fire accident data, and fire resistance coefficients of materials can be obtained;

[0108] In some embodiments, multiple data sources include: standards and specifications, such as the State Administration of Cultural Heritage's "Guidelines for Fire Protection Design of Cultural Relics Buildings" and the International Council on Monuments and Sites (ICOMOS) Conservation Guidelines; accident case studies, including historical fire archives of the Ministry of Emergency Management, accident investigation reports of key cultural relics protection units, and technical analysis reports on major international cultural relics fires; academic research, such as fire prevention technology papers and experimental data from platforms like CNKI and Elsevier; and material parameters, integrating GB8624 building material combustion performance classification data and the NIST materials thermodynamics database.

[0109] Step S202: Based on the special characteristics of fire prevention for ancient buildings and cultural relics, construct a multi-level keyword database;

[0110] In some embodiments, this embodiment targets special scenarios such as ancient wooden frames and brick and stone cultural relics, and constructs a multi-dimensional keyword library covering the building's inherent attributes, disaster causes, and protection technologies: the first-level classification includes building type, material characteristics (such as wood carbonization rate, brick and stone thermal conductivity), and fire prevention elements; the second level refines disaster-causing factors (such as weather causes, lightning strikes, incense burning, electrical aging), environmental parameters, and protection standards; the third level integrates historical repair records and new protection technologies to form a dynamically updated terminology network, supporting the construction of knowledge graphs and the semantic analysis of risk models;

[0111] Step S203: Using a web crawler program, crawl various types of data containing keywords from multiple data sources based on the multi-level keyword database;

[0112] In some embodiments, this embodiment develops an intelligent crawler program based on a predefined multi-level keyword system to achieve cross-platform data collection. For heterogeneous data sources such as standard document libraries, academic databases, and accident reporting platforms, it configures differentiated crawling strategies, combines a rule engine to automatically label the credibility level of data sources, and finally stores diverse data such as text, tables, and drawings into the architecture database to form a dynamically expandable fire prevention-themed corpus.

[0113] Step S204: Remove duplicate data of the various types by comparing hash values, store the various types of data in the database, and retain the data source and crawling timestamp;

[0114] Step S205: The specific parameters of the various types of data in the database are reviewed by experts, and the database is identified as a fire prevention database based on the reliability level of the data source.

[0115] In some embodiments, this embodiment establishes a multidisciplinary expert group including fire protection engineering, cultural relic protection, and materials science to conduct multi-dimensional verification of core data in the database, such as material combustion parameters, fire protection code clauses, and accident causation chains. For national standard data, it verifies the validity of the version; for accident cases, it verifies the completeness of spatiotemporal elements; and it performs consistency verification based on the authority of the source.

[0116] The beneficial effects of the above technical solution are as follows: it can obtain corresponding ancient building fire prevention standards, cultural relic fire prevention standards, fire accident data and material fire resistance coefficient data from multiple data sources, and automatically remove duplicate data of the various types, conduct expert review, make the data authentic and reliable, and make the content more detailed, which facilitates subsequent unified processing and enhances the reliability and intelligence of the system.

[0117] In one embodiment, such as Figure 3 As shown, the step of using natural language processing technology to perform semantic parsing on the data in the fire protection database, and processing the semantically parsed fire protection database through word vector modeling and entity relation extraction, includes:

[0118] Step S301: Use AI technology to standardize the various types of data in the fire protection database, and convert the data in the database into natural language data;

[0119] In some embodiments, this embodiment generates structured natural descriptions by using serialization templates for table-type parameters; extracts spatial topological features and converts them into directional semantic descriptions for fire compartment markings in drawing documents; and uses a domain-fine-tuned model to perform entity-attribute-value parsing and coherent paragraph reconstruction for unstructured text, eliminates data source representation differences based on adversarial generative networks, and corrects unit formats and professional terms using a rule engine, ultimately outputting a standardized corpus that conforms to natural language logic and retains original semantic details.

[0120] Step S302: Perform text unit segmentation on the natural language corpus data, and distribute the text units from near to far in the concept vector space according to their relevance.

[0121] In some embodiments, this embodiment uses a pre-trained language model to perform semantic vectorization encoding on the segmented text units, combines similarity calculation constrained by the domain knowledge graph, constructs a multi-dimensional concept space, identifies semantically similar unit clusters through a hierarchical clustering algorithm, projects high-dimensional vectors onto a two-dimensional plane to form a topology-preserving distribution mapping, dynamically adjusts the unit spacing based on the association strength gradient, and separates low-association content.

[0122] Step S303: Utilize word vector similarity retrieval to automatically expand the thesaurus, adjust the relevance weights of synonyms, and adjust the distribution of the text units based on the weight adjustment results;

[0123] In some embodiments, this embodiment uses a pre-trained word vector model to calculate the cosine similarity of domain terms, sets a threshold to filter highly relevant candidate words, constructs a dynamic synonym expansion library, and designs a weighting strategy based on contextual dependency: increases the association weight for strongly context-bound word pairs and decreases the weight for weakly related word pairs. The weight matrix is ​​integrated into the text unit vector space through a graph embedding algorithm, and the force-directed layout model is used to adjust the repulsive and attractive forces between units, so that high-weight association clusters are clustered and low-weight units are marginalized.

[0124] Step S304: Extract entities from the text units, classify the entities, build relationships between each entity category, and generate a structured data system.

[0125] In some embodiments, this embodiment combines a domain dictionary to extract architectural entities from text units, such as the Hall of Supreme Harmony in the Forbidden City, the East Hall of Foguang Temple, material entities, and specification entities. These are then divided into four categories—architectural entities, fire prevention measures, disaster-causing factors, and environmental parameters—through a predefined ontology library. Dependency parsing and semantic matching are performed to identify cross-category entity associations. A rule engine is used to construct explicit relationships, and a graph attention network is combined to mine implicit associations, enabling multi-level data linkage queries and dynamic expansion.

[0126] The beneficial effects of the above technical solution are as follows: by using AI technology to standardize the acquired data of various types, the acquired data is further optimized, which facilitates the subsequent use of the data and improves the data processing efficiency. Using AI technology to standardize the raw data improves the standardization level and processing speed of the map and increases the loading efficiency of the system.

[0127] In one embodiment, the step of constructing an internal large-scale language model using the processed fire protection database and performing hierarchical clustering on the internal large-scale language model using Leiden technology includes:

[0128] Obtain the structured data system in the preprocessed fire protection database, and determine the model architecture of the large language model based on the characteristics of the text units in the structured data system;

[0129] In some embodiments, the structured data system in this embodiment includes entities, relations, attributes, etc., while the characteristics of text units involve data type, structure, correlation, etc. The local attention mechanism is mainly designed for text attribute fields to strengthen numerical semantic correlation. Based on the characteristics of structured data, the correlation between entities, the multimodality of attributes, etc., the domain-specific model architecture of large language models is determined.

[0130] The preprocessed structured data is mixed with the original text and input into a large language model to mask the entities in the fire protection database, enabling the model to learn and determine the domain knowledge associations.

[0131] In some embodiments, this embodiment converts structured data into pseudo-natural language descriptions, concatenates them with the original text, inputs them into the model, and designs a dynamic masking strategy to implement a high proportion of random masking for core entities such as building components and material parameters. The masking ratio is set to 30%-50%, which forces the model to infer the masked entities through context. At the same time, logical tags such as specification numbers are retained to maintain syntactic constraints. Positive and negative sample pairs are constructed through contrastive learning to strengthen the co-occurrence probability learning of strongly associated entities such as "wood structure-flame retardant treatment", suppress erroneous associations such as "metal components-combustibility", and improve the model's logical reasoning ability in the knowledge system of fire protection.

[0132] The parameters of the large language model are abstracted into graph nodes, and the weights between nodes are determined by parameter similarity, thus generating a high-dimensional parameter relationship structure.

[0133] In some embodiments, this embodiment abstracts the neuron weights, attention heads and other parameters of the language model into graph nodes, constructs a high-dimensional relational network based on functional similarity to build a dense graph for the fully connected layer parameters, filters low-relevance noise edges according to the attention mechanism parameters to form a sparse parameter relational graph, identifies functionally coupled neuron clusters, such as material combustion prediction module, fire spread simulation module, etc., performs group division, and realizes interpretable representation and targeted optimization of the knowledge structure inside the model;

[0134] The Leiden algorithm is applied to detect the high-dimensional parameter relationship structure, and highly correlated internal groups are merged to form basic functional modules.

[0135] The beneficial effects of the above technical solution are as follows: by performing hierarchical clustering on the preprocessed fire protection database, the fire protection database is further optimized, which facilitates the subsequent use of the data and improves the data processing efficiency. The Leiden algorithm is applied to detect the high-dimensional parameter relationship structure, merge highly related internal groups to form basic functional modules, and enable the data to form clusters according to similarity and correlation, thereby improving the effectiveness of this method.

[0136] In one embodiment, the application of the Leiden algorithm to detect the high-dimensional parameter relationship structure, merging highly correlated internal groups to form basic functional modules, including:

[0137] The internal parameters of the language model are abstracted into graph nodes, which contain the term vector table weight distribution and entity relations of the text unit;

[0138] In some embodiments, the hidden layer parameters of the language model are abstracted into heterogeneous graph nodes. Each node encodes multidimensional features. The basic parameter features include a weight distribution histogram and a gradient update trajectory. The semantic association features are fused with the embedded representations of corresponding words in the term vector table. The structural association features inherit the role weights in the entity relationship triples. The multimodal features of adjacent nodes are aggregated through a graph attention network, and the community detection algorithm is used to identify functional modules that have both numerical similarity and semantic consistency.

[0139] The correlation of the internal parameters of the language model is calculated by cosine similarity, and the edge weights are defined based on the correlation of the internal parameters of the language model.

[0140] In some embodiments, this embodiment uses cosine similarity to calculate the response correlation of parameters in the activation mode of training data, and defines the edge weights between nodes based on the response correlation.

[0141] The Leiden algorithm is used to assign each node to an independent community. The nodes are traversed, and the modularity gain when each independent community is moved to an adjacent independent community is calculated. The direction with the maximum gain is selected to iteratively adjust the affiliation of the independent community, and the community set of the independent community is generated.

[0142] In some embodiments, this embodiment uses the Leiden algorithm to initially partition each node in the network, so that each node constitutes an independent community. It traverses all nodes, and for each node, calculates the modularity gain brought about by merging its independent community into other independent communities adjacent to it. By comparing the modularity changes caused by merging different adjacent communities, the merging direction with the largest modularity gain is selected to adjust the community affiliation of the node, and the community partitioning is gradually optimized, so that the modularity of the network is gradually improved, and finally a stable and optimized set of communities is obtained, realizing the effective partitioning and community discovery of the original network structure.

[0143] If the edge weight between two community sets exceeds a preset threshold, they are merged into the same functional module, and the same functional module is determined as the basic functional module.

[0144] Extract the high-frequency text units of the community set within the basic functional modules, match them with standard functional descriptions, and generate module tags.

[0145] In some embodiments, extracting high-frequency text units refers to finding words or phrases that appear frequently from the data of the community set; a basic functional module refers to a core part of the system, and the community set refers to user-generated content or datasets; matching standard functional descriptions and generating module tags mainly involves matching high-frequency text units with a predefined library of fire prevention functional descriptions for ancient buildings and cultural relics to find the most suitable standard description.

[0146] The beneficial effects of the above technical solution are as follows: by abstracting the internal parameters of the language model into graph nodes and calculating the correlation of the internal parameters of the language model through cosine similarity, the graph nodes can be sorted from near to far according to their correlation. Furthermore, based on the edge weights between community sets, the same functional modules are merged and the same functional modules are determined as basic functional modules, which improves the degree of data clustering and enhances the practicality and effectiveness of this method.

[0147] In one embodiment, the algorithm used to assign each node to an independent community using the Leiden algorithm, traverse the nodes, calculate the modularity gain when moving each independent community to an adjacent independent community, select the direction with the maximum gain to iteratively adjust the independent community affiliation, and generate the community set of the independent communities is as follows:

[0148]

[0149] in, For edge weights, For node degree, This represents the total edge weight. The algorithm is used to determine whether nodes belong to the same independent community. It measures the difference between the connection density within the independent community and the random distribution, and determines the specific modularity gain based on the difference.

[0150] The beneficial effects of the above technical solution are as follows: by measuring the difference between the internal connection density and random distribution of the independent community, the specific modularity gain is determined, and by iteratively adjusting the affiliation of the independent community by selecting the direction of maximum gain, a community set is generated, which further improves the practicality of this method and the relevance of the user's search results to the question.

[0151] In one embodiment, the step of establishing a multi-dimensional mapping model based on the spatial features of ancient buildings, cultural relics features, and disaster-causing factors in the hierarchical clustering results, and constructing a multi-modal dynamic association fire prevention knowledge graph based on the multi-dimensional mapping model using knowledge graph technology, includes:

[0152] The hidden representations of the spatial features of the ancient buildings and the features of the cultural relics are extracted from the basic functional modules, and standardized feature vectors are generated through graph embedding technology.

[0153] In some embodiments, this embodiment captures the semantics of the hidden layer space through the neighborhood information aggregation mechanism of the graph neural network, and uses graph embedding technology to map high-dimensional discrete features to a low-dimensional continuous vector space to achieve standardized representation of cross-modal features, and further generates feature vectors, wherein the multi-feature extraction of cultural relic features includes shape, texture, color, etc.

[0154] The clustering results of disaster-causing factors are extracted from the basic functional modules, and the clustering results are output as a probability distribution or risk score vector, while retaining the correlation weight between the clustering results and spatial features.

[0155] In some embodiments, the clustering results of disaster-causing factors extracted based on basic functional modules need to achieve risk quantification and spatial correlation preservation through multidimensional modeling. First, hierarchical clustering method is used to group multi-source disaster-causing factors. The clustering quality is evaluated and core clusters are selected by silhouette coefficient. The discrete clustering results are mapped to continuous probability space to generate a probability distribution function that represents the possibility of disaster occurrence. The weights of multidimensional disaster-causing factors are determined, and the risk scoring vector is optimized to construct a hybrid weight matrix.

[0156] A multi-branch neural network is constructed to process the feature vector and the catastrophic factor respectively, realizing cross-modal feature interaction. The contribution of different catastrophic factors to the feature vector is dynamically adjusted through an attention mechanism to generate a multi-dimensional model.

[0157] In some embodiments, the disaster-causing factor branch extracts dynamic evolution patterns through time-series modeling. To achieve cross-modal interaction, a bidirectional cross-attention mechanism is introduced to construct a dynamic correlation matrix between modalities. The contribution weights of different disaster-causing factors to the knowledge graph are calculated through multi-head self-attention. The collaborative representation between modalities is optimized through a loss function. The generated fusion representation contains the multi-dimensional risk attributes of disaster-causing factors and the mapping relationship between the knowledge graph and the knowledge graph.

[0158] The output of the multi-dimensional model is mapped to graph node attributes. A knowledge graph is built based on the graph nodes. The risk evolution path is modeled and analyzed through the knowledge graph. Association rules are applied to mine potential risk coupling relationships.

[0159] In some embodiments, the dynamic parameters such as architectural spatial features, cultural relic vulnerability index and disaster probability output by the multi-dimensional model are mapped to knowledge graph node attributes to establish a network containing building entities, protective facilities and environmental elements, construct a spatiotemporally integrated knowledge graph, use temporal graph convolutional networks to model risk evolution paths, track causal chains, mine cross-modal association rules, identify implicit risk coupling patterns, quantify association strength through confidence and enhancement, match sensing technology to locate high-risk topologies, and generate risk governance solutions containing "hazard identification - path deduction - measure recommendation";

[0160] Based on the risk evolution path and the potential risk coupling relationship, a multimodal dynamic association fire prevention knowledge graph is constructed according to the knowledge graph.

[0161] In some embodiments, the construction of a multimodal dynamic association fire prevention knowledge graph based on risk evolution paths and potential risk coupling relationships mainly integrates multi-source heterogeneous data and dynamic reasoning mechanisms, integrates multimodal data features, performs cross-modal semantic alignment between real-time sensor monitoring data and text reports, realizes the association mapping between sensor data and text descriptions through a bidirectional attention mechanism, and reveals the risk transmission chain under the coupling conditions of multiple disasters in cultural relics and ancient buildings through interpretable reasoning of the knowledge graph.

[0162] The beneficial effects of the above technical solution are as follows: by establishing a multi-dimensional mapping model based on the spatial characteristics of ancient buildings, cultural relics characteristics and disaster-causing factors in the hierarchical clustering results, the intelligence of this method is improved; by constructing a multi-modal dynamic association fire prevention knowledge graph based on the knowledge graph based on the risk evolution path and potential risk coupling relationship, the practicality and professionalism of the method are enhanced, and the user experience is improved.

[0163] In one embodiment, the step of dynamically selecting a search mode based on a user query and generating a security control result based on the search mode includes:

[0164] Receive user query dynamics, perform search pattern matching on the user query dynamics, and determine whether the question in the user query dynamics is a holistic question through an intent classification model. If the determination result is no, the first query result is directly generated based on the keywords in the question through the multimodal dynamic association fire protection knowledge graph.

[0165] In some embodiments, the system first segments and extracts entities from the query text input by the user through a pattern recognition engine, calculates the semantic similarity between word vectors and domain ontology using a pre-trained language model, identifies core keywords such as "direction of fire spread" or "key points for fire prevention in Jinci", analyzes the contextual logical structure of the query statement using a hierarchical intent classification model, determines whether the problem involves overall analysis requirements through decision boundary thresholds, and activates the knowledge graph direct connection retrieval channel to map keywords to entity nodes in the fire prevention knowledge graph. The system then uses a graph traversal algorithm to perform multi-hop reasoning along predefined relationship paths such as "building material - combustion threshold" and "meteorological conditions - fire risk level".

[0166] The user query dynamics are determined to be a holistic question. Multiple batch node reports are generated based on the holistic question by searching the knowledge graph of the multimodal dynamic association fire prevention knowledge graph using the internal large-scale language model.

[0167] In some embodiments, when it is determined that the user query involves a holistic issue such as regional fire evolution or comprehensive fire early warning analysis of cultural relics and ancient buildings, the deep knowledge reasoning engine needs to be activated. First, the semantics of the question are deconstructed through a pre-trained language model, and global analysis elements are extracted by combining the context attention mechanism. This enables adaptive retrieval of large-scale graph data, and the core nodes related to the global question are selected through vector similarity calculation.

[0168] The node report is divided into text blocks of a predefined size. Each text block generates an intermediate response, which contains key information fragments about fire prevention of cultural relics. Each key information fragment is assigned a weight coefficient based on its relevance to the overall issue.

[0169] In some embodiments, the report is divided into text blocks of equal length according to preset semantic paragraph units, and keyword boundaries (such as core concepts such as fire separation distance of historical buildings) are identified to ensure that each information block contains a complete fire prevention logic unit. Each text block is extracted by a semantic parsing engine to extract key information fragments, such as core parameters such as the critical combustion temperature of wooden buildings and the flame retardant threshold of mural restoration materials, and is standardized and labeled. The semantic similarity between the information fragments and the user's overall question is calculated based on a pre-trained language model, and the structural association strength between the information fragments and the source question is measured by the length of the knowledge graph relational path. The weighted information fragments are then screened and reorganized according to the confidence threshold.

[0170] The key information fragments are ranked according to their weight coefficients, and the key information fragments below a preset threshold are filtered to form an aggregated intermediate response. The aggregated intermediate response is then used to generate a second query result.

[0171] In some embodiments, this embodiment uses the matching degree between semantic similarity calculation and user intent as the main ranking factor, and introduces the node centrality index in the knowledge graph as an auxiliary weight coefficient to semantically align heterogeneous data such as text, spatial coordinates and fire protection code clauses, and transform the structured intermediate response into query results that conform to industry standards.

[0172] The first query result and the second query result are determined as security control results.

[0173] The beneficial effects of the above technical solution are as follows: by performing search pattern matching based on user query dynamics, it is possible to accurately match search patterns corresponding to user needs, thereby improving the intelligence of the method and the comprehensiveness of the retrieval results. Furthermore, when user needs and user query dynamics are holistic issues, a large language model can be used to perform a global search in the knowledge graph, which improves the practicality of the method and enhances the user experience.

[0174] In one embodiment, the method further includes combining the multimodal dynamic association fire prevention knowledge graph with the cultural relics and ancient buildings information model, dynamically labeling risk information in the three-dimensional model, and automatically generating a cultural relics protection priority list and protection plan based on the risk information, including:

[0175] High-precision geometric models of ancient buildings are reconstructed using lidar point cloud data, preserving detailed features such as brackets, beams, and painted decorations. A digital twin containing material properties, structural parameters, and historical repair records is established, and the model is decomposed into component-level units and bound with data tags.

[0176] In some embodiments, this embodiment uses three-dimensional laser scanning technology to obtain point cloud data with millimeter-level precision, fully preserving the detailed features such as the mortise and tenon structure of the brackets, the bending curves of the beam frame, and the painted patterns. The ancient building is decomposed into component-level units such as platform, column base, and brackets. Each unit is bound with a digital tag containing material parameters, such as wood texture and paint thickness, and structural mechanical properties, such as load-bearing coefficient and internal stress. The resulting digital twin supports component-level reverse disassembly analysis and can simulate structural deformation under different environmental loads.

[0177] Align the ancient building entities and ancient building models in the multimodal dynamic fire prevention knowledge graph with spatial coordinates, establish a two-way index relationship, and synchronize the fire risk data and cultural relic status data in the multimodal dynamic fire prevention knowledge graph to the component attribute panel of each component-level unit in real time.

[0178] In some embodiments, this embodiment uses a spatial coordinate system transformation algorithm to match the latitude and longitude positioning data of ancient building entities in the knowledge graph with the spatial grid of the model, and uses feature point registration technology to eliminate spatial drift errors between multi-source data. On the knowledge graph side, a three-dimensional model component ID hash value is configured for each bracket and beam frame node; on the model side, a query interface is embedded in the component attribute panel to support reverse retrieval of related entities in the knowledge graph based on the component ID.

[0179] Link historical inspection reports and maintenance records in the component properties panel to generate a time-series variation curve;

[0180] In some embodiments, the time-series variation curve includes key physical quantities (such as crack width and load-bearing stress) and major maintenance event nodes. The spatial mapping between the event marker line and the curve fluctuation range reveals the effectiveness of intervention measures and supports multi-scale time window adjustment.

[0181] A risk quantification index system is constructed based on building risk, environmental disaster risk, and cultural relic vulnerability indicators. The risk weight of each indicator in the risk quantification index system and the correlation coefficient between indicators are determined using the multimodal dynamic association fire prevention knowledge graph.

[0182] In some embodiments, this embodiment is based on a weight calculation engine of a multimodal dynamic association fire prevention knowledge graph, which aggregates node attributes and relationship path features, and determines the correlation strength between building ventilation structure and fire spread speed through the cosine similarity between node embedding vectors; it extracts the potential transmission paths of environmental temperature and humidity nodes and wood moisture content nodes, quantifies the nonlinear coupling coefficient of cross-modal indicators, introduces a dynamic attention mechanism in the weight allocation process, and adjusts the sensitivity weight of indicators according to the real-time fire risk warning level, such as automatically increasing the failure risk coefficient of lightning protection system during thunderstorm weather;

[0183] Based on the aforementioned risk quantification index system, a risk quantification assessment is performed on each of the aforementioned component-level units to generate a risk label and risk level for each component-level unit;

[0184] In some embodiments, this embodiment aligns the risk quantification index system with the component attribute data in the three-dimensional model through ontology mapping technology, and uses a dynamic weighted summation algorithm to calculate the comprehensive risk value: the index weights derived from the knowledge graph are combined with real-time monitoring data such as component material aging degree and environmental temperature and humidity, and the risk level classification generates a corresponding risk label system based on the component type and the preset differentiated grading standard. When the sensor detects abnormal fading of the paint or the addition of fire prevention facilities in the knowledge graph, the system automatically updates the risk matrix parameters to ensure the spatiotemporal synchronization of the assessment results with the physical state.

[0185] Based on the corresponding components of the risk label in the 3D model, risk information is dynamically labeled, and the components are color-rendered based on the risk level.

[0186] In some embodiments, the color rendering rule for the component based on the risk level is as follows: red corresponds to high risk, such as wooden components + surrounding flammable materials; yellow corresponds to medium risk, such as brick and stone structures + aging circuits; and green corresponds to protected areas.

[0187] Based on the risk label and risk level of each component-level unit, an urgency assessment is performed in the multimodal dynamic association fire prevention knowledge graph, and a comprehensive score is calculated based on the urgency assessment results.

[0188] In some embodiments, this embodiment extracts key parameters such as the risk transmission intensity, historical disaster frequency, and secondary disaster triggering probability of each component node by performing multi-hop traversal along the spatiotemporal association path in the knowledge graph, and uses a time decay function to process the timeliness characteristics of risk labels. For components whose risk level has continued to rise in the past three months, the weight coefficient is automatically increased to assess the urgency of processing and to give a comprehensive score.

[0189] The priority ranking of each component-level unit is determined based on the comprehensive score of each component-level unit, and a priority list for cultural relic protection is generated based on the priority ranking.

[0190] Based on the risk label of each component-level unit, protective measures are retrieved in the multimodal dynamic association fire prevention knowledge graph, and protective schemes for the corresponding components are generated based on the retrieval results.

[0191] In some embodiments, risk tags are parsed into semantic feature vectors in a knowledge graph, and graph traversal retrieval is performed along the three-layer relationship path of "risk type-disaster path-protective measures". For conventional risk tags, the component ontology is matched with the preset fireproof coating construction specifications. For composite risk tags, a graph neural network is used to calculate cross-modal correlation and select combined measures with synergistic protection effectiveness. The retrieval results are processed by a dynamic priority ranking module, and a protection plan is generated by comprehensively considering the timeliness of measure implementation, protection value weight and resource constraints.

[0192] The beneficial effects of the above technical solution are as follows: by combining the multimodal dynamic association fire prevention knowledge graph with the cultural relics and ancient building information model, a three-dimensional model of the ancient building can be generated in virtual space. The three-dimensional model is annotated, which further improves the flexibility of the method, reduces the intervention of the ancient building entity, and dynamically annotates risk information in the three-dimensional model. Based on the risk information, a list of cultural relics protection priorities and protection plans are automatically generated, which improves the intelligence level of the method and is conducive to the development of cultural relics and ancient building protection work.

[0193] In one embodiment, the method further includes, as Figure 5 As shown, the status of cultural relics and ancient buildings is updated in real time through IoT sensing devices, and the best safety response plan is dynamically generated based on visitor flow data, including:

[0194] The multimodal dynamic association fire prevention knowledge graph is used to determine environmental monitoring nodes, structural inspection nodes, and visitor monitoring nodes for cultural relics and ancient buildings;

[0195] In some embodiments, this embodiment uses a spatial coordinate system transformation algorithm to match the latitude and longitude positioning data of ancient building entities in the knowledge graph with the spatial grid of the model, and determines the specific location and type of sensor placement for the ancient building entity based on the retrieval function of the knowledge graph.

[0196] Temperature and humidity sensors, smoke detectors, and ultraviolet sensors are deployed at the environmental monitoring nodes of the cultural relics and ancient buildings to collect environmental data of the cultural relics and ancient buildings.

[0197] In some embodiments, environmental data includes temperature and humidity data, fire data, and ultraviolet intensity data, used to determine whether the building is in the optimal environmental range and whether a fire has occurred.

[0198] Strain gauges, inclinometers, and vibration sensors are deployed at the structural detection nodes to collect structural data of cultural relics and ancient buildings;

[0199] In some embodiments, strain gauges are used to test the internal stress of ancient buildings, inclinometers are used to determine the tilt angle of components, and vibration sensors are used to determine the vibration of components.

[0200] Infrared beam detectors and smart cameras equipped with behavior recognition algorithms are deployed at the tourist monitoring nodes to collect tourist flow data and tourist behavior data in the scenic area.

[0201] In some embodiments, infrared beam detectors are used to determine whether tourists are passing through key locations to prevent tourists from entering prohibited areas, and smart cameras equipped with behavior recognition algorithms are used to determine the number of tourists and can identify dangerous behaviors of tourists.

[0202] The ticketing system is used to obtain the current number of visitors and today's reservation data, and the future visitor flow trend is predicted based on the visitor flow data.

[0203] In some embodiments, by combining the current number of visitors and today's reservation data obtained in real time through the ticketing system with the historical time series characteristics of visitor flow, a multi-dimensional visitor flow prediction and analysis framework can be constructed. First, the structured data stream is extracted from the ticketing database, including real-time ticket gate passage records, dynamic remaining ticket quotas in the online reservation system, frequency of refund and change operations, and other core indicators. Simultaneously, environmental variables such as weather warning information, holiday tags, and traffic control notices from external data sources are integrated to generate visitor flow prediction curves.

[0204] The environmental data, the structural data, the tourist flow data, and the tourist behavior data are unified to a standard timestamp;

[0205] In some embodiments, this embodiment is used to perform time-series alignment of all data;

[0206] The current environmental data and structural data are input into the multimodal dynamic association fire prevention knowledge graph to determine the real-time risk status of cultural relics and ancient buildings. Based on the tourist flow data and future tourist flow trends, the risk weight of tourist-intensive areas is adjusted to generate a current risk report for cultural relics and ancient buildings.

[0207] In some embodiments, this embodiment integrates component status data from 3D laser scanning with real-time monitoring results from infrared thermal imaging; it mines cross-modal risk transmission relationships through a knowledge graph reasoning engine to generate a risk heat map; and the decision-making level combines emergency response plan knowledge base with immediate response measures and long-term planning guidance to generate a risk trend report through comparative analysis.

[0208] Risk identification is performed on tourist behavior data, and when high-risk behaviors are detected in the tourist behavior data, a human-caused disaster warning is generated in real time.

[0209] In some embodiments, such as Figure 6 As shown, this embodiment uses the Internet of Things (IoT) sensing layer to collect real-time data on tourist trajectory density, abnormal clustering areas captured by thermal imagers, and climbing and boundary-crossing actions identified by smart cameras. It performs multi-dimensional cross-verification of behaviors such as abnormal tourist stay time, deviation from the preset route, and intrusion into dangerous areas, and issues real-time warning information.

[0210] Based on the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters, the current risk level is determined. When the risk level exceeds the preset threshold, a risk warning is issued on the monitoring display port.

[0211] In some embodiments, this embodiment uses a knowledge graph reasoning engine to align structured risk reports with warnings of human-caused disasters, such as abnormal gatherings of tourists in popular areas or illegal carrying of fire sources, in a spatiotemporal manner. A dynamic weighted algorithm is used to construct a risk level algorithm, and an adaptive threshold mechanism is introduced into the risk level calculation. The deviation between historical disaster damage data and monitoring values ​​is analyzed in real time. When the comprehensive risk value exceeds the dynamic safety baseline, a risk warning is triggered.

[0212] Based on a multi-constraint decision-making model, the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters are used to generate the best safety disposal plan.

[0213] In some embodiments, this embodiment aligns environmental monitoring parameters, structural health indicators, and man-made disaster warnings in real-time risk reports in a spatiotemporal manner to construct a multi-dimensional decision space that includes constraints such as risk level, emergency resource distribution, and response timeliness requirements. It also mines the characteristics of the optimal disposal path from historical emergency cases contained in the knowledge graph and outputs the best current safety disposal plan in real time.

[0214] The beneficial effects of the above technical solution are as follows: By updating the status of cultural relics and ancient buildings nodes in real time through Internet of Things (IoT) sensing devices, real-time monitoring of cultural relics and ancient buildings can be carried out based on environmental data, structural data, visitor flow data, and visitor behavior data, and the real-time risk status of cultural relics and ancient buildings can be determined. Furthermore, by combining visitor flow data, the best current safety handling plan can be dynamically generated, reducing the possibility of human-caused disasters. The best current safety handling plan can be generated using the current cultural relics and ancient building risk report and human-caused disaster warning, thus improving the safety and intelligence of the method.

[0215] Those skilled in the art should understand that the "first" and "second" in this invention simply refer to different application stages.

[0216] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0217] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A security control method based on knowledge graphs, characterized in that, Includes the following steps: The system acquires multiple data sources, including fire prevention standards for ancient buildings, fire prevention standards for cultural relics, fire accident data, and fire resistance coefficients of materials; constructs a multi-level keyword database; uses a web crawler to crawl various types of data containing keywords from multiple data sources based on the multi-level keyword database; removes duplicate data of various types, stores the data in the database, and retains the data source and crawling timestamp. Multiple types of data were reviewed by experts, and the database was identified as a fire prevention database based on the reliability level of the data source. Semantic parsing is performed on the data in the fire protection database, and the data in the semantically parsed fire protection database is then processed. Based on the structured data system in the preprocessed fire protection database, the model architecture of the large-scale language model is determined; Preprocessed structured data and raw text are mixed and input into a large language model to mask entities in the fire protection database, enabling the model to learn and determine domain knowledge associations. Leiden technology is used to perform hierarchical clustering on the internal large language model. A multi-dimensional mapping model is established based on the hierarchical clustering results, and a multimodal dynamic association fire prevention knowledge graph is constructed using knowledge graph technology. Based on the user's query keywords, select the search mode, and generate security control results; A high-precision geometric model of the ancient building was reconstructed using LiDAR point cloud data, preserving detailed features, and a digital twin was established. The model was then decomposed into component-level units and bound with data tags. Align the ancient building entities and ancient building models in the multimodal dynamic fire prevention knowledge graph with spatial coordinates, establish a two-way index relationship, and synchronize the fire risk data and cultural relic status data in the multimodal dynamic fire prevention knowledge graph to the component attribute panel of each component-level unit. The risk weight of each indicator and the correlation coefficient between indicators in the risk quantification indicator system are determined using a multimodal dynamic association fire prevention knowledge graph.

2. The knowledge graph-based security control method according to claim 1, characterized in that, The aforementioned method uses a web crawler to collect various types of data based on keywords, including fire prevention data for ancient buildings, cultural relics, fire accidents, and material coefficients, to construct a fire prevention database. It was determined that multiple data sources could be obtained, including fire prevention standards for ancient buildings, fire prevention standards for cultural relics, fire accident data, and fire resistance coefficients of materials. Based on the special characteristics of fire prevention for ancient buildings and cultural relics, a multi-level keyword database was constructed. The web crawler program retrieves various types of data containing keywords from multiple data sources based on the multi-level keyword database. The data of the various types is removed by comparing hash values ​​and then stored in the database, retaining the data source and the timestamp of the crawling. The specific parameters of the various types of data in the database are reviewed by experts, and the database is identified as a fire prevention database based on the reliability level of the data source.

3. The knowledge graph-based security control method according to claim 1, characterized in that, Natural language processing (NLP) techniques are used to perform semantic parsing on the data in the fire protection database. This involves word vector modeling and entity relation extraction to process the semantically parsed fire protection database, including: AI technology is used to standardize various types of data in the fire prevention database, converting the data in the database into natural language data. The natural language corpus data is segmented into text units, and the text units are sorted according to their relevance and distributed from near to far in the concept vector space. By using word vector similarity retrieval, the thesaurus is automatically expanded, the relevance weights of the synonyms are adjusted, and the distribution of the text units is adjusted based on the weight adjustment results; Entities are extracted from the text units, the entities are classified, and relationships are established between each entity category to generate a structured data system.

4. The knowledge graph-based security control method according to claim 1, characterized in that, An internal large-scale language model was constructed using the processed fire protection database. Hierarchical clustering was then performed on this internal large-scale language model using Leiden techniques, including: Obtain the structured data system in the preprocessed fire protection database, and determine the model architecture of the large language model based on the characteristics of the text units of the structured data system; The preprocessed structured data is mixed with the original text and input into a large language model to mask the entities in the fire protection database, enabling the model to learn and determine the domain knowledge associations. The parameters of the large language model are abstracted into graph nodes, and the weights between nodes are determined by parameter similarity, thus generating a high-dimensional parameter relationship structure. The Leiden algorithm is applied to detect the high-dimensional parameter relationship structure, and highly correlated internal groups are merged to form basic functional modules.

5. The knowledge graph-based security control method according to claim 4, characterized in that, The application of the Leiden algorithm detects the high-dimensional parameter relationship structure, merges highly correlated internal groups, and forms basic functional modules, including: The internal parameters of the language model are abstracted into graph nodes, which contain the term vector table weight distribution and entity relations of the text unit; The correlation of the internal parameters of the language model is calculated by cosine similarity, and the edge weights are defined based on the correlation of the internal parameters of the language model. The Leiden algorithm is used to assign each node to an independent community. The nodes are traversed, and the modularity gain when each independent community is moved to an adjacent independent community is calculated. The direction with the maximum gain is selected to iteratively adjust the affiliation of the independent community, and the community set of the independent community is generated. If the edge weight between two community sets exceeds a preset threshold, they are merged into the same functional module, and the same functional module is determined as the basic functional module. Extract the high-frequency text units of the community set within the basic functional modules, match them with standard functional descriptions, and generate module tags.

6. The knowledge graph-based security control method according to claim 5, characterized in that, The algorithm used to assign each node to an independent community using the Leiden algorithm, traverse the nodes, calculate the modularity gain when moving each independent community to an adjacent independent community, select the direction with the maximum gain to iteratively adjust the affiliation of independent communities, and generate the community set of the independent communities is as follows: ,in, For edge weights, For node degree, This represents the total edge weight. The algorithm is used to determine whether nodes belong to the same independent community. It measures the difference between the connection density within the independent community and the random distribution, and determines the specific modularity gain based on the difference.

7. The knowledge graph-based security control method according to claim 4, characterized in that, A multi-dimensional mapping model is established based on the spatial characteristics of ancient buildings, cultural relic characteristics, and disaster-causing factors in the hierarchical clustering results. Based on the multi-dimensional mapping model, a multi-modal dynamic association fire prevention knowledge graph is constructed using knowledge graph technology, including: The hidden representations of the spatial features of the ancient buildings and the features of the cultural relics are extracted from the basic functional modules, and standardized feature vectors are generated through graph embedding technology. The clustering results of disaster-causing factors are extracted from the basic functional modules, and the clustering results are output as a probability distribution or risk score vector, while retaining the correlation weight between the clustering results and spatial features. A multi-branch neural network is constructed to process the feature vector and the catastrophic factor respectively, realizing cross-modal feature interaction. The contribution of different catastrophic factors to the feature vector is dynamically adjusted through an attention mechanism to generate a multi-dimensional model. The output of the multi-dimensional model is mapped to graph node attributes. A knowledge graph is built based on the graph nodes. The risk evolution path is modeled and analyzed through the knowledge graph. Association rules are applied to mine potential risk coupling relationships. Based on the risk evolution path and the potential risk coupling relationship, a multimodal dynamic association fire prevention knowledge graph is constructed according to the knowledge graph.

8. The knowledge graph-based security control method according to claim 1, characterized in that, Based on the user query, a search mode is dynamically selected, and based on the search mode, a security control result is generated, including: Receive user query dynamics, perform search pattern matching on the user query dynamics, and determine whether the question in the user query dynamics is a holistic question through an intent classification model. If the determination result is no, the first query result is directly generated based on the keywords in the question through the multimodal dynamic association fire protection knowledge graph. The user query dynamics are determined to be a holistic question. Multiple batch node reports are generated based on the holistic question by searching the knowledge graph of the multimodal dynamic association fire prevention knowledge graph using the internal large-scale language model. The node report is divided into text blocks of a predefined size. Each text block generates an intermediate response, which contains key information fragments about fire prevention of cultural relics. Each key information fragment is assigned a weight coefficient based on its relevance to the overall issue. The key information fragments are ranked according to their weight coefficients, and the key information fragments below a preset threshold are filtered to form an aggregated intermediate response. The aggregated intermediate response is then used to generate a second query result. The first query result and the second query result are determined as security control results.

9. The knowledge graph-based security control method according to claim 1, characterized in that, The method further includes combining the multimodal dynamic association fire prevention knowledge graph with the cultural relics and ancient buildings information model, dynamically labeling risk information in the three-dimensional model, and automatically generating a cultural relics protection priority list and protection plan based on the risk information, including: High-precision geometric models of ancient buildings are reconstructed using lidar point cloud data, preserving detailed features such as brackets, beams, and painted decorations. A digital twin containing material properties, structural parameters, and historical repair records is established, and the model is decomposed into component-level units and bound with data tags. Align the ancient building entities and ancient building models in the multimodal dynamic fire prevention knowledge graph with spatial coordinates, establish a two-way index relationship, and synchronize the fire risk data and cultural relic status data in the multimodal dynamic fire prevention knowledge graph to the component attribute panel of each component-level unit in real time. Link historical inspection reports and maintenance records in the component properties panel to generate a time-series variation curve; A risk quantification index system is constructed based on building risk, environmental disaster risk, and cultural relic vulnerability indicators. The risk weight of each indicator in the risk quantification index system and the correlation coefficient between indicators are determined using the multimodal dynamic association fire prevention knowledge graph. Based on the aforementioned risk quantification index system, a risk quantification assessment is performed on each of the aforementioned component-level units to generate a risk label and risk level for each component-level unit; Based on the corresponding components of the risk label in the 3D model, risk information is dynamically labeled, and the components are color-rendered based on the risk level. Based on the risk label and risk level of each component-level unit, an urgency assessment is performed in the multimodal dynamic association fire prevention knowledge graph, and a comprehensive score is calculated based on the urgency assessment results. The priority ranking of each component-level unit is determined based on the comprehensive score of each component-level unit, and a priority list for cultural relic protection is generated based on the priority ranking. Based on the risk label of each component-level unit, protective measures are retrieved in the multimodal dynamic association fire prevention knowledge graph, and protective schemes for the corresponding components are generated based on the retrieval results.

10. The knowledge graph-based security control method according to claim 1, characterized in that, The method also includes updating the status of cultural relics and ancient buildings in real time through IoT sensing devices, and dynamically generating the best safety response plan based on visitor flow data, including: The multimodal dynamic association fire prevention knowledge graph is used to determine environmental monitoring nodes, structural inspection nodes, and visitor monitoring nodes for cultural relics and ancient buildings; Temperature and humidity sensors, smoke detectors, and ultraviolet sensors are deployed at the environmental monitoring nodes of the cultural relics and ancient buildings to collect environmental data of the cultural relics and ancient buildings. Strain gauges, inclinometers, and vibration sensors are deployed at the structural detection nodes to collect structural data of cultural relics and ancient buildings; Infrared beam detectors and smart cameras equipped with behavior recognition algorithms are deployed at the tourist monitoring nodes to collect tourist flow data and tourist behavior data in the scenic area. The ticketing system is used to obtain the current number of visitors and today's reservation data, and the future visitor flow trend is predicted based on the visitor flow data. The environmental data, the structural data, the tourist flow data, and the tourist behavior data are unified to a standard timestamp; The current environmental data and structural data are input into the multimodal dynamic association fire prevention knowledge graph to determine the real-time risk status of cultural relics and ancient buildings. Based on the tourist flow data and future tourist flow trends, the risk weight of tourist-intensive areas is adjusted to generate a current risk report for cultural relics and ancient buildings. Risk identification is performed on tourist behavior data, and when high-risk behaviors are detected in the tourist behavior data, a human-caused disaster warning is generated in real time. Based on the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters, the current risk level is determined. When the risk level exceeds the preset threshold, a risk warning is issued on the monitoring display port. Based on a multi-constraint decision-making model, the current risk report of cultural relics and ancient buildings and the warning of human-caused disasters are used to generate the best safety disposal plan.

Images