Flood mapping method and system based on adaptive entity relation extraction

CN123031840APending Publication Date: 2026-11-03NORTHWEST ENGINEERING CORPORATION LIMITED
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611557672.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-09
Publication Date
2026-11-03

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively unify and link flood information from different sources and in different forms, resulting in scattered disaster information and unclear relationships, making it difficult to accurately reflect the impact of different disaster-causing factors on flood risk and affecting the accuracy of flood risk analysis.

Method used

By using adaptive entity relationship extraction and dynamic graph relationship mining, a flood knowledge graph is constructed to obtain standardized flood entities and business relationships, and incremental updates are performed. The dynamic flood knowledge graph is then used to conduct flood factor risk extrapolation and analysis.

Benefits of technology

It improves the consistency of flood knowledge organization and the accuracy of flood risk analysis, accurately reflects the impact of different disaster-causing factors on flood risk, and enhances the reliability of flood risk assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123031840A_ABST
    Figure CN123031840A_ABST
Patent Text Reader

Abstract

The disclosure provides a flood atlas construction method and system based on adaptive entity relationship extraction, relating to the technical field of data processing. The method comprises: acquiring multi-source flood data of a target flood control area, wherein the multi-source flood data comprises flood text data and hydrological time series data; inputting the flood text data into an adaptive entity extraction model to obtain a plurality of standardized flood entities; constructing a dynamic entity graph according to the plurality of standardized flood entities, inputting the dynamic entity graph into a dynamic graph relationship mining model to obtain standardized business relationships between entities; constructing a flood knowledge graph according to the plurality of standardized flood entities, the standardized business relationships and the multi-source flood data; incrementally updating the flood knowledge graph according to new flood data to obtain a dynamic flood knowledge graph; and performing flood element risk deduction analysis based on the dynamic flood knowledge graph to obtain a flood risk analysis result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and more specifically, to a method and system for constructing flood maps based on adaptive entity relationship extraction. Background Technology

[0002] Urban flooding is characterized by its suddenness, wide impact, and complex contributing factors. Flood control operations generate a wealth of disaster records, hydrological monitoring information, engineering operation data, and emergency response materials, providing a crucial data foundation for flood risk assessment.

[0003] However, existing technologies lack a unified and effective organization and correlation of flood information from different sources and in different forms, which easily leads to scattered disaster information and unclear relationships, making it difficult to form a complete flood disaster knowledge system. At the same time, existing flood analysis methods are insufficient in exploring the complex relationships between various disaster-causing factors and their degree of influence, making it difficult to accurately reflect the role of different disaster-causing factors in flood risk, thus affecting the accuracy of flood risk analysis. Summary of the Invention

[0004] The purpose of this disclosure is to provide a flood knowledge graph construction method, a flood knowledge graph construction system, an electronic device, and a computer-readable storage medium based on adaptive entity relationship extraction. This method can obtain standardized flood entities and standardized business relationships through adaptive entity extraction and dynamic graph relationship mining, construct and incrementally update a flood knowledge graph, and perform flood element risk inference and analysis based on the dynamic flood knowledge graph, thereby improving the accuracy of flood knowledge organization and flood risk analysis.

[0005] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0006] According to a first aspect of the present disclosure, a flood map construction method based on adaptive entity relationship extraction is provided, comprising: acquiring multi-source flood data of a target flood control area, wherein the multi-source flood data includes flood text data and hydrological time-series data; inputting the flood text data into an adaptive entity extraction model to obtain multiple standardized flood entities; constructing a dynamic entity graph based on the multiple standardized flood entities, and inputting the dynamic entity graph into a dynamic graph relationship mining model to obtain standardized business relationships between entities; constructing a flood knowledge graph based on the multiple standardized flood entities, the standardized business relationships, and the multi-source flood data; incrementally updating the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph; and performing flood element risk inference and analysis based on the dynamic flood knowledge graph to obtain flood risk analysis results.

[0007] In an example embodiment of this disclosure, based on the aforementioned scheme, the acquisition of multi-source flood data for the target flood control area includes: acquiring flood text data, hydrological time-series data, geographic information spatial data, and engineering ledger data; wherein, the flood text data includes flood plans, disaster reports, dispatching procedures, flood control bulletins, hazard reports, and emergency response records; the hydrological time-series data includes monitoring data on rainfall, river levels, pipeline flow, and water depth; the geographic information spatial data includes spatial data corresponding to rivers, drainage networks, sluice gates, low-lying areas, administrative regions, and flood-prone points; and the engineering ledger data includes water conservancy project archives, flood control material ledgers, hazard point ledgers, and facility operation and maintenance data.

[0008] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of inputting the flood text data into an adaptive entity extraction model to obtain multiple standardized flood entities includes: performing word segmentation and positional encoding on the flood text data, and fusing domain enhancement features corresponding to flood domain proper nouns to obtain text input features; inputting the text input features into the domain semantic enhancement layer of the adaptive entity extraction model for semantic encoding to obtain domain semantic features; inputting the domain semantic features into a temporal feature extraction layer for bidirectional temporal encoding to obtain bidirectional temporal features that integrate contextual information; and inputting the bidirectional temporal features into a label constraint decoding layer to perform global constraint decoding on label transfers according to entity annotation rules to obtain multiple standardized flood entities.

[0009] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of globally constraining and decoding the tag transfer according to entity labeling rules to obtain multiple standardized flood entities includes: acquiring pre-established flood domain entity types and entity labeling rules, wherein the flood domain entity types include disaster-causing factor entities, water conservancy project entities, disaster event entities, spatial location entities, emergency resource entities, and control and dispatch entities, and the entity labeling rules include the entity's starting position, the entity's internal position, and the tags corresponding to non-entity positions; determining the legitimate transfer relationship between tags according to the flood domain entity types and the entity labeling rules; globally constraining the candidate tag sequence corresponding to the bidirectional temporal features according to the legitimate transfer relationship; determining the target entity tag sequence from the candidate tag sequence that satisfies the global constraints, and obtaining multiple standardized flood entities according to the target entity tag sequence.

[0010] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of constructing a dynamic entity graph based on multiple standardized flood entities and inputting the dynamic entity graph into a dynamic graph relationship mining model to obtain standardized business relationships between entities includes: using multiple standardized flood entities as entity nodes, generating corresponding node representations based on the semantic features, category features, and location features of each entity node; calculating the semantic similarity between each entity node, establishing candidate connection edges between the entity nodes based on the semantic similarity, and constructing the dynamic entity graph based on the entity nodes and the candidate connection edges; calculating the association weights of each candidate connection edge under different relationship types through a multi-head relationship attention mechanism based on the node representations of each entity node and the relationship type features corresponding to different relationship types, and obtaining the corresponding relationship features based on the association weights; classifying the relationship features to obtain the standardized business relationships corresponding to each candidate connection edge.

[0011] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of establishing candidate connection edges between entity nodes according to the semantic similarity includes: comparing the semantic similarity between any two entity nodes with a preset similarity threshold; when the semantic similarity reaches the preset similarity threshold, determining the corresponding two entity nodes as dynamic adjacent entities; and establishing the candidate connection edges between the two dynamic adjacent entities to determine the adjacency relationship of the dynamic entity graph.

[0012] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of classifying the relational features to obtain the standardized business relationships corresponding to each candidate connection edge includes: classifying the relational features corresponding to each candidate connection edge to obtain relational classification probabilities corresponding to different relational types; determining the target relational type corresponding to each candidate connection edge from a preset business relational system based on the relational classification probabilities; and determining the target relational type as the standardized business relationship between corresponding entity nodes. The preset business relational system includes disaster causal relationships, spatial affiliation relationships, engineering linkage relationships, emergency response relationships, and attribute association relationships. The disaster causal relationships include induction leading to, exacerbation of, mitigation of, and co-existence. The spatial affiliation relationships include jurisdictional inclusion, locational adjacency, upstream and downstream distribution, and low-lying affiliation. The engineering linkage relationships include coordinated scheduling, pipeline connectivity, gate pump linkage, and water storage replenishment. The emergency response relationships include emergency response, material allocation, personnel duty, and early warning issuance. The attribute association relationships include threshold correspondence, state binding, level matching, and time-period coupling.

[0013] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of constructing a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data includes: determining a head entity and a tail entity based on two standardized flood entities associated with each standardized business relationship; combining the head entity, the standardized business relationship, and the tail entity to form a triple; forming multi-source heterogeneous knowledge based on the multi-source flood data; performing semantic alignment on the triples and the multi-source heterogeneous knowledge to obtain semantically aligned knowledge; performing adaptive fusion on the semantically aligned knowledge to obtain fused knowledge; performing entity disambiguation on the fused knowledge to obtain disambiguated knowledge; performing relationship verification and temporal denoising on the disambiguated knowledge to obtain standardized fused knowledge; and constructing the flood knowledge graph based on the standardized fused knowledge.

[0014] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of semantically aligning the triples and the multi-source heterogeneous knowledge to obtain semantically aligned knowledge includes: obtaining information on different source objects corresponding to the standardized flood entity from the multi-source heterogeneous knowledge; extracting features corresponding to each of the different source object information and determining the semantic matching degree between the different source object information; mapping the different source object information corresponding to the same flood object to a unified standardized flood entity according to the semantic matching degree; and establishing the association between the triples and the mapped different source object information to obtain the semantically aligned knowledge.

[0015] In an example embodiment of this disclosure, based on the aforementioned scheme, the adaptive fusion of the semantic alignment knowledge to obtain fused knowledge includes: acquiring different source data mapped to the same standardized flood entity in the semantic alignment knowledge; determining the information entropy corresponding to each of the different source data; determining the corresponding fusion weight based on the information entropy corresponding to each of the different source data; and fusing the different source data mapped to the same standardized flood entity based on the fusion weight corresponding to each of the different source data to obtain the fused knowledge.

[0016] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of performing entity disambiguation on the fused knowledge to obtain disambiguated knowledge includes: constructing corresponding joint features based on the entity semantic features, entity category features, spatial location features, and business attribute features of the entities to be matched; determining the comprehensive similarity between the entities to be matched based on the joint features; when the comprehensive similarity meets a preset entity fusion condition, determining the corresponding entities to be matched as homologous entities and performing entity fusion; when the names of the entities to be matched are the same and the comprehensive similarity does not meet the preset entity fusion condition, retaining the corresponding entities to be matched as different entities, thereby obtaining the disambiguation knowledge.

[0017] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of performing relation verification and temporal denoising on the disambiguation knowledge to obtain standardized fusion knowledge includes: establishing legality constraints on relationships between different entity types based on the standardized business relationships; using the legality constraints to verify the legality of entity relationships in the disambiguation knowledge and eliminating entity relationships that do not meet the legality constraints; obtaining the time information corresponding to each triple and verifying the temporal order relationship between the triples based on the time information; eliminating or deduplicating triples with relational conflicts, temporal conflicts, or repetitions to obtain the standardized fusion knowledge.

[0018] In an example embodiment of this disclosure, based on the aforementioned scheme, the incremental update of the flood knowledge graph according to newly added flood data to obtain a dynamic flood knowledge graph includes: dividing the knowledge in the flood knowledge graph into static inherent knowledge, dynamic temporal knowledge, and temporary working condition knowledge; determining the knowledge type corresponding to the newly added flood data, and determining the corresponding update method according to the knowledge type, wherein the static inherent knowledge adopts a batch update method, the dynamic temporal knowledge adopts a streaming update method, and the temporary working condition knowledge adopts an event-triggered update method; performing entity extraction and relation mining on the newly added flood data to obtain new triples, and performing validity verification and redundancy removal on the new triples; matching the verified new triples with the existing knowledge in the flood knowledge graph to determine the new knowledge and the graph substructure related to the new knowledge; updating the entity nodes, entity relationships, network parameters, and graph topology in the graph substructure according to the update method corresponding to the new knowledge to obtain the dynamic flood knowledge graph.

[0019] In an example embodiment of this disclosure, based on the foregoing scheme, updating the entity nodes, entity relationships, network parameters, and graph topology in the graph substructure includes: adding new entity nodes to the graph substructure according to the new knowledge, and updating the entity information corresponding to entity nodes that match existing entity nodes; freezing historical network parameters unrelated to the new knowledge, and incrementally updating network parameters related to the new knowledge; obtaining historical triplet samples corresponding to historical flood knowledge, and jointly training the historical triplet samples with the new samples corresponding to the new knowledge; redetermining the adjacency relationships between entity nodes related to the new knowledge based on the jointly trained network parameters; and updating the corresponding entity relationships and graph topology according to the redetermined adjacency relationships to complete the update of the graph substructure.

[0020] In an example embodiment of this disclosure, based on the aforementioned scheme, the step of performing flood risk extrapolation analysis based on the dynamic flood knowledge graph to obtain flood risk analysis results includes: obtaining entity attributes corresponding to multiple standardized flood entities from the dynamic flood knowledge graph, and constructing flood risk analysis features based on the entity attributes; inputting the flood risk analysis features into a temporal attention neural network to mine the nonlinear coupling relationship between each flood risk analysis feature and flood risk; determining the risk contribution weight of each flood-causing factor and the flood risk prediction result based on the nonlinear coupling relationship; using each risk contribution weight as the contribution degree of the corresponding flood-causing factor, and obtaining the flood risk analysis result based on the contribution degree and the flood risk prediction result.

[0021] According to a second aspect of the present disclosure, a flood map construction system based on adaptive entity relationship extraction is provided. The system includes: a data acquisition module for acquiring multi-source flood data of a target flood control area, wherein the multi-source flood data includes flood text data and hydrological time-series data; an entity extraction module for inputting the flood text data into an adaptive entity extraction model to obtain multiple standardized flood entities; a relationship mining module for constructing a dynamic entity graph based on the multiple standardized flood entities, and inputting the dynamic entity graph into a dynamic graph relationship mining model to obtain standardized business relationships between entities; a map construction module for constructing a flood knowledge graph based on the multiple standardized flood entities, the standardized business relationships, and the multi-source flood data; an incremental update module for incrementally updating the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph; and a risk inference module for performing flood element risk inference analysis based on the dynamic flood knowledge graph to obtain flood risk analysis results.

[0022] According to a third aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions that, when executed by the processor, implement the flood map construction method based on adaptive entity relation extraction as described in the first aspect.

[0023] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the flood map construction method based on adaptive entity relation extraction as described in the first aspect.

[0024] The technical solutions provided in this disclosure may have the following beneficial effects: According to the flood knowledge graph construction method based on adaptive entity relationship extraction in this example embodiment, on the one hand, standardized flood entities are obtained by adaptively extracting entities from flood text data. A dynamic entity graph is then constructed based on these standardized entities, and a dynamic graph relationship mining model is used to determine standardized business relationships between entities. This allows entity information and its relationships in the flood text to be represented in a unified form, improving the correlation and structure of flood information. On the other hand, a flood knowledge graph is constructed based on standardized flood entities, standardized business relationships, and multi-source flood data. The knowledge graph is incrementally updated based on new flood data, ensuring that flood information from different sources and in different forms is continuously incorporated into a unified knowledge structure, improving the completeness and consistency of flood knowledge. Furthermore, by performing flood factor risk extrapolation analysis based on the dynamic flood knowledge graph, the contribution level of different flood-causing factors is determined, and flood risk analysis results are generated. This allows for further utilization of various flood elements and their relationships in the graph to reflect the impact of different disaster-causing factors on flood risk, thereby improving the reliability of flood risk assessment. This can improve the accuracy of flood knowledge organization and flood risk analysis.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0027] Figure 1 The illustration shows a flowchart of a flood map construction method based on adaptive entity relationship extraction according to some embodiments of the present disclosure.

[0028] Figure 2 The diagram illustrates a schematic representation of the entity relationship structure of a flood knowledge graph according to some embodiments of the present disclosure.

[0029] Figure 3 The diagram illustrates a flood knowledge graph construction and incremental update structure according to some embodiments of the present disclosure.

[0030] Figure 4 The diagram illustrates a block diagram of a flood map construction system based on adaptive entity relationship extraction according to some embodiments of the present disclosure.

[0031] Figure 5The schematic diagram illustrates the structural schematic of a computer system of an electronic device according to some embodiments of the present disclosure.

[0032] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0033] 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 numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of systems and methods consistent with some aspects of this specification as detailed in the appended claims.

[0034] The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “the,” and “the” as used in this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0035] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be more thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art.

[0036] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, systems, steps, etc., can be employed. In other instances, well-known methods, systems, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0037] Furthermore, the accompanying drawings are for illustrative purposes only and are not necessarily drawn to scale. The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0038] In the exemplary embodiments disclosed herein, a flood map construction method based on adaptive entity relation extraction is first provided. This method can be implemented by a flood map construction system based on adaptive entity relation extraction. Figure 1 The illustration schematically shows a flowchart of a flood map construction method based on adaptive entity relationship extraction according to some embodiments of the present disclosure. Reference Figure 1 As shown, the flood map construction method based on adaptive entity relationship extraction may include the following steps: Step S110: Obtain multi-source flood data for the target flood control area, including flood text data and hydrological time series data; Step S120: Input the flood text data into the adaptive entity extraction model to obtain multiple standardized flood entities; Step S130: Construct a dynamic entity graph based on multiple standardized flood entities, input the dynamic entity graph into the dynamic graph relationship mining model, and obtain the standardized business relationships between entities; Step S140: Construct a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data; Step S150: Incrementally update the flood knowledge graph based on the newly added flood data to obtain a dynamic flood knowledge graph; Step S160: Based on the dynamic flood knowledge graph, perform risk extrapolation and analysis of flood factors to obtain flood risk analysis results.

[0039] According to the flood knowledge graph construction method based on adaptive entity relationship extraction in this example embodiment, on the one hand, standardized flood entities are obtained by adaptively extracting entities from flood text data. Furthermore, a dynamic entity graph is constructed based on these standardized entities, and a dynamic graph relationship mining model is used to determine standardized business relationships between entities. This allows entity information and its relationships in the flood text to be represented in a unified form, improving the correlation and structure of flood information. On the other hand, a flood knowledge graph is constructed based on standardized flood entities, standardized business relationships, and multi-source flood data. The knowledge graph is incrementally updated based on new flood data, ensuring that flood information from different sources and in different forms is continuously incorporated into a unified knowledge structure, improving the completeness and consistency of flood knowledge. Furthermore, by performing flood factor risk extrapolation analysis based on the dynamic flood knowledge graph, the contribution level of different flood-causing factors is determined, and flood risk analysis results are generated. This allows for further utilization of various flood elements and their relationships in the graph to reflect the impact of different disaster-causing factors on flood risk, thereby improving the reliability of flood risk assessment. This can improve the accuracy of flood knowledge organization and flood risk analysis.

[0040] The following will further explain the flood map construction method based on adaptive entity relationship extraction in this example embodiment.

[0041] In step S110, multi-source flood data of the target flood control area is acquired, including flood text data and hydrological time series data.

[0042] The target flood control area can represent the geographical region where flood knowledge organization and flood risk analysis are required, and its scope can be determined based on actual flood control operational needs. Multi-source flood data can represent a collection of data from different flood operational data sources used to describe flood-related information concerning the target flood control area. Flood text data can represent data in text form recording flood operational objects, flood events, operational status, and handling information. Hydrological time-series data can represent data acquired continuously or periodically in chronological order, reflecting changes in the hydrological status of the target flood control area.

[0043] Specifically, the data acquisition scope can be determined based on the target flood control area. Flood text data and hydrological time-series data can be obtained from data sources related to the target flood control area, and the acquired data can be aggregated into the flood data set corresponding to the target flood control area. Different data may have different data generation methods, data update cycles, and data representation formats. During the acquisition process, the source information and time information of each data point should be maintained so that the acquired multi-source flood data can correspond to the same target flood control area.

[0044] By acquiring flood text data and hydrological time-series data of the target flood control area, flood information scattered from different data sources can be centrally collected, enabling subsequent processing to simultaneously utilize the flood business information recorded in the text and the hydrological information that changes over time, thereby improving the completeness of flood information acquisition.

[0045] In step S120, the flood text data is input into the adaptive entity extraction model to obtain multiple standardized flood entities.

[0046] The Adaptive Entity Extraction Model (AEM) can be defined as an entity extraction model that adapts to the semantics of flood-related texts and can identify text units with predefined flood-related business meanings from flood-related text data. For example, the AEM can employ a neural network model capable of performing contextual semantic representation of the input text and entity sequence recognition. Standardized flood entities can represent flood-related business semantic units uniformly represented according to predefined flood entity representation rules, used to represent objects, events, states, or other information with independent business meanings in flood-related business.

[0047] Specifically, the flood text data obtained in step S110 is used as input to the adaptive entity extraction model. The adaptive entity extraction model analyzes the contextual semantics in the flood text data, identifies text content with flood business meaning, and determines the entity scope and entity semantics corresponding to the text content. For a flood text, one or more flood entities can be identified based on the different flood business semantics contained therein, and output according to a unified entity representation rule, thereby obtaining multiple standardized flood entities.

[0048] Compared to directly processing the original text for subsequent knowledge, using an adaptive entity extraction model to convert unstructured or semi-structured flood text into standardized flood entities with a unified representation can reduce the differences in entity representation caused by different text expression methods and improve the consistency of flood entity recognition and organization.

[0049] In step S130, a dynamic entity graph is constructed based on multiple standardized flood entities. The dynamic entity graph is then input into a dynamic graph relationship mining model to obtain standardized business relationships between entities.

[0050] The dynamic entity graph can represent a graph structure with multiple standardized flood entities as graph nodes, forming node connection relationships based on the current associations between these standardized flood entities. The dynamic graph relationship mining model can represent a relationship mining model capable of identifying business relationships between different entities based on node information and graph structure information between nodes in the dynamic entity graph. For example, the dynamic graph relationship mining model can employ a graph neural network model capable of aggregating graph node and node association information and outputting relationship categories. Standardized business relationships can represent relationship information that uniformly describes the business associations between two standardized flood entities according to preset flood business relationship representation rules.

[0051] Specifically, the standardized flood entities obtained in step S120 are used as entity nodes. Connections between these nodes are established based on potential semantic relationships, thus forming a dynamic entity graph. Since different flood texts contain different entities and their relationships, the node composition and node connections of the constructed dynamic entity graph can change depending on the flood text being processed.

[0052] Furthermore, the dynamic entity graph is input into the dynamic graph relationship mining model. This model, combined with information about the entity nodes themselves and their associations within the dynamic entity graph, analyzes the relationships between entities with potential business connections and determines the corresponding business relationships. For different relationship representation formats, the identified relationships are represented according to a unified business relationship representation rule, thus obtaining standardized business relationships between entities.

[0053] By organizing multiple standardized flood entities into a dynamic entity graph and mining entity relationships based on the graph structure, the relationship identification process can incorporate the overall association between different entities, reducing the omissions caused by relying solely on local text information for relationship judgment, and improving the completeness and accuracy of business relationship identification between flood entities.

[0054] In step S140, a flood knowledge graph is constructed based on multiple standardized flood entities, standardized business relationships, and multi-source flood data.

[0055] Among them, the flood knowledge graph can represent a structured knowledge set that takes flood domain knowledge as the object and organizes flood entities, business relationships between entities, and data related to entities or relationships through a graph structure.

[0056] Specifically, the standardized flood entities obtained in step S120 are used as knowledge objects in the flood knowledge graph, and the standardized business relationships obtained in step S130 are used as the association information between different knowledge objects to establish a relationship structure between the standardized flood entities. Simultaneously, based on the flood objects described by the multi-source flood data and their correspondence with the standardized flood entities, relevant data are associated with corresponding entities or business relationships, enabling flood information that originally existed in textual or temporal form to be incorporated into a unified knowledge structure. Thus, a flood knowledge network can be formed based on the business relationships between multiple standardized flood entities, and relevant information in the multi-source flood data can be linked with corresponding knowledge objects, thereby constructing a flood knowledge graph.

[0057] In step S150, the flood knowledge graph is incrementally updated based on the newly added flood data to obtain a dynamic flood knowledge graph.

[0058] Specifically, newly added flood data refers to newly generated or acquired flood data related to the target flood control area after the flood knowledge graph is constructed. Incremental updates refer to the method of incorporating new knowledge or knowledge changes reflected in the newly added flood data into the existing flood knowledge graph while retaining the existing knowledge. Dynamic flood knowledge graphs refer to flood knowledge graphs that reflect changes in flood knowledge in the target flood control area after continuous updates with newly added flood data.

[0059] Specifically, after constructing the flood knowledge graph in step S140, newly generated flood data in the target flood control area is continuously acquired. When new flood data is acquired, the flood information contained in the new flood data is identified, and the knowledge changes relative to the existing flood knowledge graph are determined. Based on the determined knowledge changes, the relevant graph content in the existing flood knowledge graph is updated, so that the new flood information is incorporated into the existing knowledge structure, resulting in an updated dynamic flood knowledge graph.

[0060] As new flood data continues to be generated, the flood knowledge graph can be incrementally updated in the manner described above, enabling the dynamic flood knowledge graph to be updated in accordance with changes in flood information in the target flood control area. Compared to rebuilding the complete knowledge graph after new data is generated, this method can reduce the duplication of processing existing knowledge, maintain the continuous connection between new knowledge and historical knowledge, and improve the timeliness of flood knowledge updates.

[0061] In step S160, flood risk analysis is performed based on the dynamic flood knowledge graph to obtain flood risk analysis results.

[0062] Among them, flood elements can represent entities, entity attributes, and inter-entity relationships related to the occurrence, development, and risk status of flood disasters in a dynamic flood knowledge graph. Flood element risk extrapolation analysis can represent the process of using different flood elements and their associated information in the dynamic flood knowledge graph to perform correlation analysis and extrapolation of flood risk status. Flood risk analysis results can represent the analytical information obtained through flood element risk extrapolation analysis, used to characterize the flood risk status of the target flood control area.

[0063] Specifically, the dynamic flood knowledge graph obtained in step S150 is read, and flood elements related to the current flood risk and the correlation information between these elements are extracted from the graph. Combined with relevant data recorded in the dynamic flood knowledge graph, a comprehensive analysis of the correlation status between multiple flood elements is performed. Based on the obtained element correlation information, the flood risk of the target flood control area under the current or analyzed conditions is extrapolated, and a flood risk analysis result is formed based on the extrapolated risk information.

[0064] Because the dynamic flood knowledge graph organizes scattered flood elements and their business relationships into a unified knowledge structure, when conducting risk simulations based on this dynamic flood knowledge graph, it can comprehensively utilize the correlation information between different flood elements, reduce the information gaps caused by relying solely on a single flood element for risk judgment, and thus improve the accuracy of flood risk analysis.

[0065] The following is a detailed explanation of the steps described above.

[0066] In some embodiments, acquiring multi-source flood data for a target flood control area can be achieved through the following technical steps: The first step is to acquire flood text data, hydrological time-series data, geographic information spatial data, and engineering ledger data.

[0067] Flood text data can represent data recorded in natural language text format, including operational information such as flood prevention, disaster occurrence, disaster development, flood control scheduling, and emergency response in a target flood control area. For example, flood text data can include flood plans, disaster reports, scheduling procedures, flood control bulletins, hazard reports, and emergency response records. For instance, flood plans can record early warning, inspection, and response measures adopted under different rainfall and water conditions; disaster reports can record the time, location, type, and impact of flood events; scheduling procedures can record the operation and scheduling requirements of water conservancy projects such as gates, pumping stations, and storage facilities; flood control bulletins can record real-time or periodic rainfall, water conditions, engineering conditions, and hazards; hazard reports can record disaster situations such as road flooding, pipeline overflows, river flooding, and dike seepage; and emergency response records can record information such as rescue teams, pumping equipment, flood control materials, and on-site response measures.

[0068] Hydrological time-series data can represent data collected continuously or periodically in chronological order, reflecting changes in the hydrological status of a target flood control area. Hydrological time-series data includes monitoring data on rainfall, river levels, pipeline flow, and water depth. For example, rainfall can be recorded at the minute, hour, or per event time granularity; river levels can record real-time water levels at different hydrological stations or river cross-sections; pipeline flow can reflect the drainage and discharge status of the drainage network at different times; and water depth can reflect changes in water accumulation in roads, underpasses, low-lying areas, or other flood-prone locations. For hydrological time-series data, information such as monitoring time, monitoring location, monitored physical quantity, measured value, and data source can be retained to ensure that data from different times can be correlated with specific flood-affected areas.

[0069] Geographic spatial data can represent data with geographical location, spatial extent, or spatial connectivity. Geographic spatial data includes spatial data corresponding to rivers, drainage networks, sluice gates, low-lying areas, administrative regions, and flood-prone points. For example, geographic spatial data can originate from a Geographic Information System (GIS), and its data format can include point spatial objects, line spatial objects, area spatial objects, or raster data. Rivers and drainage networks can be represented using line spatial objects, sluice gates and flood-prone points can be represented using point spatial objects, and low-lying areas and administrative regions can be represented using area spatial objects. By storing the coordinate information or spatial extent information of each spatial object, a correspondence can be established between different flood control operational information and the actual geographic space.

[0070] Engineering ledger data represents data used to record flood control engineering facilities, flood control resources, potential hazards, and facility operation and maintenance information. Engineering ledger data includes water conservancy project archives, flood control material ledgers, potential hazard ledgers, and facility operation and maintenance data. Water conservancy project archives record the names, locations, design parameters, and engineering attributes of engineering facilities such as drainage pumping stations, control gates, box culverts, dikes, and regulating lakes; flood control material ledgers record the types, quantities, and storage locations of resources such as mobile water pumps, flood control sandbags, and emergency vehicles; potential hazard ledgers record potential hazards such as flood-prone areas, weak dike sections, and areas with insufficient drainage capacity; and facility operation and maintenance data records the inspection, malfunctions, start-up, shutdown, and operational status of engineering facilities.

[0071] In other implementations, multi-source flood data can be expanded to include meteorological and hydrological data, pipeline infrastructure data, topographic surface data, water conservancy scheduling data, disaster record data, and operation and maintenance emergency data, depending on the actual flood control business scope. For example, topographic surface data can record topographic elevation, topographic slope, road hardening status, and green space distribution; water conservancy scheduling data can record pump station start-up and shutdown status, gate opening degree, and storage facility operation status, etc.

[0072] During data acquisition, data acquisition boundaries can be determined according to the administrative scope, water system scope, or drainage zone scope corresponding to the target flood control area, and data corresponding to this scope from different business systems can be aggregated. Simultaneously, basic information such as the source, generation time, and spatial location of the original data can be retained, enabling unified management of data from different business systems according to the target flood control area.

[0073] By acquiring the above-mentioned different types of data, we can simultaneously obtain flood business texts, hydrological status changes, spatial distribution, and engineering operation information, reducing the problem that a single data source cannot fully describe the flood business objects and flood status, and improving the integrity of multi-source flood data.

[0074] In some embodiments, inputting flood text data into an adaptive entity extraction model to obtain multiple standardized flood entities can be achieved through the following technical steps: The first step is to segment and positionally encode the flood-related text data, and then integrate the domain-enhanced features corresponding to proper nouns in the flood field to obtain the text input features.

[0075] Among these, word segmentation refers to the serialization of flood-related text data according to characters, words, or sub-words, transforming a flood-related text into a sequence of text units that can be input into the model. Positional encoding represents information used to characterize the relative or absolute position of each text unit in the flood-related text. Domain-enhanced features represent features formed based on the semantic information corresponding to proper nouns in the flood-related domain, used to enhance the representation ability of flood-related professional terms.

[0076] For example, a flood-specific thesaurus can be pre-established, which may include proper nouns related to flood control such as stormwater culverts, river floodplains, pipeline overflows, drainage pumping stations, low-lying areas, storage lakes, and pre-release scheduling. After segmenting the flood-related text data, it can be determined whether each text unit corresponds to a proper noun in the flood control field or is part of a proper noun, and the corresponding domain-enhanced vector can be obtained based on the determination result.

[0077] For input flood-related text, its text input features can be represented as: ; in, Represents text input features; Represents the basic word vectors; Indicates positional encoding; This represents the domain enhancement vector obtained from a flood-specific thesaurus.

[0078] For example, for the text "High water level operation of stormwater culverts," we can use basic word vectors to describe the general semantics of each word, positional encoding to describe the sequential position of each word in the text, and domain enhancement vectors to enhance the domain semantics of specialized terms such as "stormwater culverts." Combining these three types of encoding yields the corresponding text input features. By adding domain enhancement features to the basic text encoding, we can enhance the distinguishability of specialized terms in the model input and reduce the semantic weakening that occurs when flood-related specialized terms are represented using general semantics.

[0079] The second step is to input the text input features into the domain semantic enhancement layer of the adaptive entity extraction model for semantic encoding to obtain domain semantic features.

[0080] The domain semantic enhancement layer can be represented as a semantic feature extraction layer formed by adapting the basic semantic encoding model to flood-related text. For example, the domain semantic enhancement layer can be constructed using a Bidirectional Encoder Representations from Transformers (BERT) model, and the parameters of the basic model can be adjusted using flood-related text samples.

[0081] During training or adaptation, flood-related text with completed entity annotations can be input into the basic semantic encoding model. By updating the model parameters, the model's semantic response to flood control documents, disaster descriptions, and professional water conservancy engineering texts can be improved, while reducing the interference of general, everyday semantics unrelated to current flood-related business on entity extraction results. After model adaptation, corresponding domain semantic features can be output for each text position in the input text features. For example, in "Chengdong Drainage Pumping Station Starts Full-Speed ​​Drainage," the domain semantic enhancement layer can form a domain semantic representation distinct from general everyday text based on the common occurrence patterns of words like "drainage," "pumping station," and "start drainage" in flood-related texts and the contextual semantics. Semantic encoding through the domain semantic enhancement layer allows for a more complete representation of the meaning of different flood-related professional terms in context, helping to reduce the probability of professional entities being missed.

[0082] The third step involves inputting the domain semantic features into the temporal feature extraction layer for bidirectional temporal encoding, thereby obtaining bidirectional temporal features that integrate contextual information.

[0083] The temporal feature extraction layer can be used to perform forward and backward information transmission of domain semantic features according to the arrangement order of characters or words in the flood text. For example, the temporal feature extraction layer can adopt a bidirectional long short-term memory (BiLSTM) network.

[0084] For example, in lengthy flood control documents, the name of a water conservancy project may be separated from the words describing its status, dangers, or dispatching measures by a considerable distance. By extracting forward features, information preceding the current text position can be utilized; by extracting backward features, information following the current text position can be utilized. This allows the bidirectional temporal features corresponding to the current position to simultaneously reflect the preceding and following context.

[0085] By using bidirectional temporal coding, the impact of large semantic distances in long sentences, complex sentences, and lengthy official documents on entity identification can be reduced, thereby improving the completeness of long-boundary entity recognition.

[0086] The fourth step involves inputting the bidirectional temporal features into the label constraint decoding layer, and performing global constraint decoding on the label transfer according to the entity annotation rules to obtain multiple standardized flood entities.

[0087] A label-constrained decoding layer can be represented as a decoding layer that makes an overall judgment on each candidate label based on the combination of labels in the complete text sequence. For example, a label-constrained decoding layer can employ a Conditional Random Field (CRF). The label-constrained decoding layer can receive bidirectional temporal features corresponding to each text position, determine the overall score of different candidate label sequences based on the candidate scores of different labels corresponding to each text position and the transition scores between adjacent labels, and output the entity label sequence whose overall score satisfies the constraints.

[0088] In one training implementation, the adaptive entity extraction model can employ... The model was trained with a learning rate of 16, a batch size of 16, and 80 training epochs. In one set of experiments, the precision, recall, and F1 score of the combined bidirectional long short-term memory network and conditional random field model were 72.36%, 70.12%, and 71.22%, respectively; the precision, recall, and F1 score of the combined basic semantic encoding model and conditional random field model were 81.45%, 79.63%, and 80.53%, respectively; the precision, recall, and F1 score of the combined basic semantic encoding model and the combined bidirectional long short-term memory network and conditional random field model were 83.71%, 82.04%, and 82.87%, respectively; and the adaptive entity extraction model adapted to flood-related text achieved precision, recall, and F1 score of 92.14%, 90.68%, and 91.40%, respectively, with an F1 score improvement of 8.53 percentage points.

[0089] In some embodiments, global constraint decoding of label transfer according to entity annotation rules is performed to obtain multiple standardized flood entities, which can be achieved through the following technical steps: The first step is to obtain the pre-established entity types and entity labeling rules in the flood domain. The entity types in the flood domain include disaster-causing factor entities, water conservancy project entities, disaster event entities, spatial location entities, emergency resource entities, and control and dispatch entities. The entity labeling rules include the starting position of the entity, the internal position of the entity, and the labels corresponding to non-entity positions.

[0090] Among them, disaster-causing entities can represent natural environmental or hydrological factors that can influence the occurrence or development of floods. Disaster-causing entities can include rainfall, strong winds, tides, high temperatures, river levels, and soil moisture content. For example, short-duration heavy rainfall, water levels exceeding warning levels, and moisture from the outer periphery of typhoons can be specific instances of disaster-causing entities.

[0091] A water conservancy project entity can refer to engineering facilities that participate in urban flood control, drainage, water storage, or water volume control. Water conservancy project entities can include drainage pumping stations, control gates, stormwater pipe networks, box culverts, dikes, storage lakes, and water distribution outlets. For example, the drainage pumping station in the east of the city, the stormwater box culvert in the south of the city, and a flood control dike along a certain section of a river can all be considered water conservancy project entities.

[0092] A disaster event entity can represent a specific disaster phenomenon or emergency that occurs during the occurrence and development of flooding. Disaster event entities can include road flooding, pipeline overflows, levee seepage, river flooding, and flooding of underground parking garages. For example, flooding on main roads during the morning rush hour and the risk of flooding blocking underpasses can be considered disaster event entities.

[0093] Spatial point entities can represent the geographical location or spatial area corresponding to flood control business objects. Spatial point entities can include administrative areas, road locations, hydrological stations, low-lying areas, and catchment areas. For example, the old town catchment area, the low-lying area in the south of the city, and river hydrological monitoring stations can be used as spatial point entities.

[0094] Emergency resource entities can represent personnel, equipment, supplies, or emergency locations that can be mobilized during flood emergency response. Emergency resource entities can include rescue teams, pumping equipment, flood control supplies, emergency sites, and on-duty personnel. For example, mobile water pumps, district-level flood control and rescue teams, and emergency shelters can serve as emergency resource entities.

[0095] A control and dispatch entity can represent orders, levels, measures, or timelines generated during flood control or engineering scheduling. A control and dispatch entity can include dispatch instructions, control levels, warning levels, response measures, and control periods. For example, a pumping station operating at full speed, a river channel discharge restriction order, or a specific warning level can all serve as a control and dispatch entity.

[0096] Entity annotation rules can use the BIO (Browser-Initial-Output) annotation method. Here, the B tag indicates the starting position of the entity, the I tag indicates a position inside the entity, and the O tag indicates a non-entity position. For different entity types, the B and I tags can be combined with the corresponding entity category to distinguish between different entities.

[0097] For example, for the sentence "Short-term heavy rainfall caused flooding in the old town's catchment area, and drainage pumping stations started at full speed," the first text unit corresponding to "short-term heavy rainfall" can be labeled as a B-tag for the disaster-causing entity, and subsequent text units belonging to the same entity can be labeled as I-tags for the disaster-causing entity; "old town catchment area" is labeled according to spatial location entities; "road flooding" is labeled according to disaster event entities; "drainage pumping stations" is labeled according to water conservancy project entities; and "starting at full speed" is labeled according to control and dispatch entities.

[0098] The second step is to determine the legitimate transfer relationships between labels based on the entity types and entity labeling rules in the flood field.

[0099] In this context, a valid transition relationship indicates that two entity tags corresponding to adjacent text positions can form a tag combination that conforms to entity boundary rules. For example, an entity's "I" tag should appear after a "B" tag or another "I" tag of the same entity category, and should not appear directly without an entity start tag.

[0100] A label transition constraint table or label transition matrix can be pre-established to record the allowed or prohibited transition relationships between labels. For the six types of flood-related entities, corresponding starting labels and internal labels can be established respectively. If the current text position corresponds to the internal label of a certain water conservancy project entity, the next text position can continue to correspond to the internal label of the same water conservancy project entity, or it can enter other legal label states after the current entity ends, but it is not allowed to directly form the internal label of another entity that lacks a starting boundary.

[0101] The third step is to apply global constraints to the candidate label sequences corresponding to the bidirectional temporal features based on the legal transition relationships.

[0102] Specifically, the tag score for each candidate tag at each text position can be calculated based on bidirectional temporal features, and the overall score of the complete candidate tag sequence can be calculated by combining the transition scores between adjacent tags. For tag combinations that violate legal transition relationships, their transition scores can be set to lower values ​​or they can be directly excluded, so that the corresponding candidate tag sequence cannot be determined as the final result.

[0103] Unlike text-position-independent classification, global constraints consider the combination of labels corresponding to the preceding and following text positions when determining the label of a certain text position. Therefore, it can reduce the problem of illegal entity boundaries formed by the highest probability label combination in the local area.

[0104] The fourth step is to determine the target entity label sequence from the candidate label sequences that satisfy the global constraints, and to obtain multiple standardized flood entities based on the target entity label sequence.

[0105] The target entity label sequence can be represented by the label sequence with the highest overall score among candidate label sequences that satisfy the legal transition relationship. Entity boundaries can be determined based on the text range covered by B labels and consecutive I labels in the target entity label sequence, and entity categories can be determined based on the corresponding label types.

[0106] For example, when the target entity label sequence determines that a certain continuous text range is a water conservancy project entity, the text corresponding to that range can be output as a standardized flood entity; for adjacent but different entities, their boundaries can be determined separately through the starting label, reducing the situation of entities sticking together or being incorrectly split.

[0107] In some embodiments, a dynamic entity graph is constructed based on multiple standardized flood entities. The dynamic entity graph is then input into a dynamic graph relationship mining model to obtain standardized business relationships between entities. This can be achieved through the following technical steps: The first step is to use multiple standardized flood entities as entity nodes, and generate corresponding node representations based on the semantic features, category features, and location features of each entity node.

[0108] In this context, entity nodes can represent nodes in a dynamic entity graph used to characterize standardized flood entities. Node representation can transform the semantic information, entity category information, and the entity's location information in the flood text of an entity into feature vectors that can be used for graph structure computation.

[0109] For example, a dynamic graph relation mining model can employ a Relational Graph Attention Network (RGAT). For the... The initial node representation of an entity node can be expressed as: ; in, Indicates the entity node index; Indicates the first The initial node representation of each entity node; Indicates the first Semantic information corresponding to each entity; This indicates entity semantic embedding; Indicates the first The entity category corresponding to each entity; Indicates entity category embedding; Indicates the first The location of an entity in the flood text. Indicates feature splicing; This indicates entity location embedding. For example, for the entity "Chengdong Drainage Pumping Station", semantic features can represent the semantics of the entity in the current flood text; category features can represent that it belongs to water conservancy project entities; and location features can represent the starting position, ending position, or relative position of the entity in the text.

[0110] The second step is to calculate the semantic similarity between each entity node, establish candidate connection edges between entity nodes based on the semantic similarity, and construct a dynamic entity graph based on the entity nodes and candidate connection edges.

[0111] Candidate edges can represent graph connections used to indicate potential business relationships between two entity nodes. Dynamic entity graphs can represent graph structures where the number of entity nodes, the number of candidate edges, or the candidate connection relationships can change as the entity content in the current flood text changes.

[0112] Unlike methods using fixed adjacency relationships, the semantic features of each entity node change accordingly when the current flood text changes, and the candidate connection edges established can also change. For example, the same drainage pumping station may mainly form engineering linkage relationships with the pipeline network or storage facilities under normal operating conditions, but may also form new candidate associations with disaster events and emergency response targets under sudden flooding conditions. By determining candidate connection edges based on the current entity semantics, the dynamic entity graph can change with different flood texts and different operating conditions, reducing the limitations imposed by fixed adjacency structures on cross-sentence and dynamic relationship mining.

[0113] The third step involves calculating the association weights of each candidate connection edge under different relationship types based on the node representation of each entity node and the relationship type features corresponding to different relationship types through a multi-head relationship attention mechanism, and obtaining the corresponding relationship features based on the association weights.

[0114] A multi-head relational attention mechanism can be represented by using multiple parallel attention computation units to calculate the degree of association between entity nodes under different relation types from different feature subspaces, and then combining the multiple computation results. For example, eight attention heads can be used to process node features in parallel.

[0115] For relation types Entity node and entity nodes The association weights between them can be expressed as: ; in, Representing relation type Constrained entity nodes With entity nodes The correlation weight between them; Represents entity nodes The current corresponding node characteristics; Represents entity nodes The current corresponding node characteristics; Represents entity nodes The The node characteristics corresponding to each dynamically adjacent entity; Represents the node feature mapping parameter matrix; Representing relation type Corresponding relation type characteristics; Indicates the parameters for attention calculation; This indicates the transpose of the attention calculation parameters; Represents entity nodes The current set of dynamically adjacent entities; This represents the activation function of a linear unit with leakage correction; This indicates exponentiation.

[0116] For each candidate connection edge, the association weights corresponding to multiple relation types can be calculated separately. Then, the relevant node features and relation type features are aggregated according to their corresponding weights to form the relation features corresponding to that candidate connection edge. For example, the dynamic graph relation mining model can employ 256-dimensional hidden features, 8 attention heads, and a random deactivation coefficient of 0.2. Weight decay coefficient, The system is trained with a learning rate of 16, a batch size of 16, and 100 training epochs. Training can be stopped when the validation results no longer improve after 15 consecutive training epochs.

[0117] The fourth step is to classify the relation features to obtain the standardized business relations corresponding to each candidate connection edge.

[0118] Relationship classification can calculate the classification probability of different candidate relation categories based on the relation features corresponding to the candidate connection edges. For example, the Softmax classification function can be used to convert the output values ​​corresponding to each relation category into classification probabilities, and cross-entropy loss can be used for model training.

[0119] The relation classification loss can be expressed as: ; in, Represents the loss for relational classification; Indicates the number of relation categories involved in the classification; Indicates a relational category index; Indicates the first The actual label values ​​corresponding to each relationship category; The model predicts the first... The classification probability corresponding to each relation category.

[0120] In some embodiments, establishing candidate connection edges between entity nodes based on semantic similarity can be achieved through the following technical steps: The first step is to compare the semantic similarity between any two entity nodes with a preset similarity threshold.

[0121] Semantic similarity can be determined using cosine similarity. For entity nodes... and entity nodes Semantic similarity can be expressed as: ; in, Represents entity nodes and entity nodes Semantic similarity between them; Represents entity nodes The corresponding semantic feature vector; Represents entity nodes The corresponding semantic feature vector; Represents entity nodes The norm of the corresponding semantic feature vector; Represents entity nodes The norm of the corresponding semantic feature vector.

[0122] A preset similarity threshold represents the similarity boundary for determining whether two entity nodes have a candidate connection relationship. For example, the preset similarity threshold can be set to 0.65. In practical applications, this threshold can also be determined based on training data or the textual features of the target flood control area.

[0123] The second step is to determine the two corresponding entity nodes as dynamic adjacent entities when the semantic similarity reaches the preset similarity threshold.

[0124] In this context, a dynamic adjacency entity can represent two entity nodes in the current dynamic entity graph that are determined to have a candidate association relationship based on the current entity semantics. When the preset similarity threshold is reached, entity nodes can be... and entity nodes Add each other to the corresponding dynamic adjacent entity set.

[0125] For example, a flood control bulletin states "short-term heavy rainfall continues to intensify" at the beginning and "large-scale waterlogging occurs on main urban roads" at the end. Even if the two entities are not in the same sentence, when the corresponding semantic features meet the similarity condition, the two entity nodes corresponding to "short-term heavy rainfall" and "waterlogging on main urban roads" can be identified as dynamically adjacent entities.

[0126] The third step is to establish candidate connection edges between two dynamically adjacent entities to determine the adjacency relationship of the dynamic entity graph.

[0127] After establishing candidate edges, the corresponding connections can be written into the adjacency information of the dynamic entity graph. For two entity nodes whose semantic similarity does not reach a preset similarity threshold, no candidate edges need to be established. As the input flood text or working conditions change, the semantic similarity between entity nodes can be recalculated, and candidate edges can be re-determined. Through this dynamic adjacency determination, the graph structure can be made independent of a pre-fixed adjacency matrix, thus adaptively changing entity connection relationships according to different flood texts and working conditions.

[0128] In some embodiments, classifying relation features to obtain standardized business relationships corresponding to each candidate connection edge can be performed through the following technical steps: The first step is to classify the relation features corresponding to each candidate connection edge to obtain the relation classification probability corresponding to different relation types.

[0129] The relation features corresponding to candidate connection edges can be input into the relation classification output layer. Normalization classification is then applied to obtain the relation classification probabilities for different relation types. Each relation classification probability represents the likelihood that two entity nodes belong to the corresponding business relation category.

[0130] The second step is to determine the target relationship type corresponding to each candidate connection edge from the preset business relationship system based on the relationship classification probability.

[0131] The pre-defined business relationship system includes disaster causal relationships, spatial affiliation relationships, engineering linkage relationships, emergency response relationships, and attribute association relationships.

[0132] Among them, the causal relationship of disaster can represent the causal link between disaster-causing factors and the occurrence or change of flood disasters, including inducing and causing, aggravating and worsening, mitigating and weakening, and co-existing. For example, short-term heavy rainfall and road flooding can form an inducing and causing relationship; sustained high water levels can form an aggravating and worsening relationship with the risk of river flooding.

[0133] Spatial affiliation can represent the geographical location or spatial belonging relationship between flood-prone entities, including jurisdictional inclusion, geographical adjacency, upstream and downstream distribution, and low-lying affiliation. For example, a flood-prone area may belong to a certain administrative region, and a certain hydrological station may be located upstream of a certain river section.

[0134] Engineering linkages can represent the operation, scheduling, or hydraulic connections between different water conservancy projects, including coordinated scheduling, pipeline network connectivity, gate-pump linkage, and water storage and replenishment. For example, the Chengdong drainage pumping station and the Chengnan stormwater culvert can form a coordinated scheduling relationship; gates and pumping stations can form a gate-pump linkage relationship based on the current water level.

[0135] Emergency response relationships can represent the connection between disaster events and emergency resources or response actions, including rescue and relief efforts, resource allocation, personnel on duty, and early warning issuance. For example, road flooding can correspond to the emergency response of mobile water pumps, and a specific emergency situation can correspond to the allocation of flood control materials.

[0136] Attribute relationships can represent the matching connections between an entity's own attributes, operational status, warning level, or time conditions, including threshold correspondence, status binding, level matching, and time period coupling. For example, river water level can form a threshold correspondence with warning water level thresholds, and engineering operation status can form a time period coupling with corresponding scheduling time periods.

[0137] The third step is to define the target relationship type as the standardized business relationship between the corresponding entity nodes.

[0138] The target relation type can be selected based on the relation classification probability that satisfies the preset classification conditions. For cases where only one relation is allowed between a pair of entity nodes, the relation type with the highest classification probability can be selected; for cases where business rules allow multiple relations between a pair of entities, multiple relation types that satisfy the classification conditions can be retained. For example, the relation results corresponding to "short-term heavy rainfall, inducing flooding on main urban roads" can be obtained, as can the relation results corresponding to "Eastern drainage pumping station, coordinated scheduling, southern stormwater culvert". By mapping the relation classification results output by the model to the preset business relation system, relations obtained from different textual expressions can be represented using unified business relation names, improving the consistency of relation knowledge.

[0139] In some embodiments, constructing a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data can be achieved through the following technical steps: The first step is to determine the head entity and tail entity based on the two standardized flood entities associated with each standardized business relationship, and then combine the head entity, the standardized business relationship, and the tail entity to form a triple.

[0140] Among them, the head entity can represent a standardized flood entity that serves as the starting object of an entity relation; the tail entity can represent a standardized flood entity that serves as the object to which the relation points; and the triple can represent a data structure that expresses flood knowledge in a structured way according to the three components of the head entity, relation, and tail entity.

[0141] For example, a triplet can be formed: "short-term heavy rainfall, triggering, road flooding," to represent the catastrophic relationship between short-term heavy rainfall and road flooding; a triplet can be formed: "river level rise, triggering, sluice gate pre-discharge scheduling," to represent the connection between river level changes and engineering scheduling; a triplet can be formed: "mobile pumps, handling, road flooding"; and a triplet can be formed: "reservoir, undertaking, flood season peak shaving and storage tasks." By converting entities and relationships into triplets, business knowledge originally existing in natural language text can be transformed into a unified structural form.

[0142] The second step is to form multi-source heterogeneous knowledge based on multi-source flood data.

[0143] Among them, multi-source heterogeneous knowledge can be represented as knowledge extracted or transformed from flood text data, hydrological time series data, geographic information spatial data and engineering ledger data, and has heterogeneous characteristics due to differences in data source, data structure or representation form.

[0144] For example, the same drainage pumping station can appear as the pumping station name and scheduling description in flood text data, as drainage flow or operational status sequence in hydrological time series data, as pumping station coordinates or spatial objects in geographic information spatial data, and as fields such as project number, designed drainage capacity, and equipment status in engineering ledger data. All of these different data describe the same flood object, but their original forms of expression differ, thus forming multi-source heterogeneous knowledge.

[0145] The third step is to perform semantic alignment on the triples and multi-source heterogeneous knowledge to obtain semantically aligned knowledge.

[0146] Semantic alignment knowledge can represent knowledge that has established correspondences between knowledge from different sources and a unified, standardized flood entity. This step can match objects from different sources that have the same or similar business meanings, enabling multiple types of data on the same flood object to be associated with a unified entity.

[0147] The fourth step involves adaptively fusing the semantically aligned knowledge to obtain fused knowledge. This fused knowledge can represent knowledge formed by combining different sources of knowledge mapped to the same standardized flood entity according to their respective data characteristics.

[0148] The fifth step is to perform entity disambiguation on the fused knowledge to obtain disambiguated knowledge.

[0149] Disambiguation knowledge can represent the knowledge formed by distinguishing or merging homonymous, heteronymous, or duplicate entities in fused knowledge. Entity disambiguation can prevent the same actual flood object from becoming multiple duplicate entities due to different names, and can also prevent objects with the same name but located in different places or with different business attributes from being incorrectly merged.

[0150] The sixth step is to perform relation verification and temporal noise reduction on the disambiguation knowledge to obtain standardized fused knowledge, and then construct a flood knowledge graph based on the standardized fused knowledge.

[0151] Standardized fused knowledge can represent fused knowledge that meets the requirements of flood knowledge organization after entity unification, relation validity checks, and temporal logic processing. A flood knowledge graph can be constructed using a set of flood entities, a set of entity relationships, and a set of attributes corresponding to each entity.

[0152] Entity attributes can include numerical monitoring attributes, equipment operating status attributes, spatial location attributes, temporal variation attributes, and operating condition threshold attributes. For example, a drainage pumping station entity can have attributes such as design flow rate, current operating status, spatial location, and start-up and shutdown times; a flood-prone area entity can have attributes such as water depth, spatial extent, and risk threshold. For instance, the entity relationship structure of a flood knowledge graph can be referenced... Figure 2 As shown. See also Figure 2 The flood disaster knowledge graph can use meteorological entities, hydrological entities, geographical risk entities, dispatch object entities, monitoring and dispatch as knowledge nodes, and organize different flood business objects through corresponding relationships.

[0153] Meteorological entities can be associated with rainstorm processes and may include rainfall stations, weather radars, typhoons, and severe convective weather; hydrological entities may include floods, urban waterlogging, flash floods, dam breaks, debris flows, and water pollution; geographical risk entities may include flood-prone areas, landslide hazard points, hydrological stations, low-lying areas, sluice gates, administrative villages, pumping stations, and main transportation routes; and entities subject to dispatch may include blue alerts, dikes, dangerous sections of dikes, yellow alerts, drainage projects, orange alerts, emergency response teams, and material reserves.

[0154] Furthermore, different knowledge nodes in the flood disaster knowledge graph can be linked through information such as river channels, triggers, flow paths, locations, impacts, and schools. Combined with corresponding business information from monitoring and dispatching, a knowledge network is formed to represent different flood business objects and their relationships.

[0155] In one specific organizational approach, a flood knowledge graph can be organized into a data layer, ontology layer, knowledge layer, algorithm layer, and application layer. The data layer can include meteorological and hydrological data, pipeline infrastructure, topographic underlying surface data, water conservancy scheduling data, disaster records, and operation and maintenance emergency data. The ontology layer can set up five top-level ontology categories: disaster, environment, facilities, prevention and control, and scenario. The knowledge layer can form a structured triplet knowledge base, an entity attribute relationship network, and coupled knowledge of flood elements. The algorithm layer can be configured with semantic alignment, adaptive fusion, temporal correlation analysis, and knowledge reasoning processing. The application layer can form applications such as risk analysis, disaster cause analysis, drainage scheduling analysis, and hazard mitigation based on actual flood control operational needs.

[0156] In some embodiments, semantic alignment of triples and multi-source heterogeneous knowledge is performed to obtain semantically aligned knowledge, which can be achieved through the following technical steps: The first step is to obtain information on objects from different sources that correspond to the standardized flood entity from multi-source heterogeneous knowledge.

[0157] Information from different sources can represent information originating from different data modalities or different business systems, but potentially corresponding to the same actual flood-prone object. For example, for the same flood-prone location, water depth sequences can be obtained from hydrological time-series data, location and impact range can be obtained from geographic information spatial data, hazard point numbers and treatment records can be obtained from engineering ledger data, and disaster descriptions can be obtained from flood text data.

[0158] For object information from different sources, preliminary candidate matching can be performed by object name, spatial location, time information, or business fields to filter out combinations of object information that may correspond to the same standardized flood entity.

[0159] The second step is to extract the features corresponding to the information from different sources and determine the degree of semantic matching between the information from different sources.

[0160] For information from two different sources, feature vectors can be formed separately. and The degree of semantic matching can be determined using the following formula: ; in, Indicates the degree of semantic matching between information from objects of different sources; This represents the feature vector corresponding to the information of the first source object; This represents the feature vector corresponding to the information of the second source object; This indicates the number of feature dimensions involved in the matching of two feature vectors; Indicates the feature dimension index; Representing the eigenvector The Each feature component; Representing the eigenvector The Each feature component; Representing the eigenvector The average value of each characteristic component; Representing the eigenvector The average value of each characteristic component.

[0161] The closer the calculation result is to 1, the more consistent the two objects are in terms of semantic definition, indicator connotation, or feature changes; a lower calculation result indicates that there are significant semantic or business attribute differences between the two objects. For example, if one system uses "road water accumulation depth" as the field name and another system uses "flood-prone area water depth" as the field name, the corresponding features can be extracted after unifying the units and statistical time, and the semantics and business attributes can be used to determine whether the two correspond to the same flood indicator.

[0162] The third step is to map information from different sources corresponding to the same flood object to a unified, standardized flood entity based on the degree of semantic matching.

[0163] Semantic matching conditions can be set to perform unified entity mapping on object information from different sources that meet the conditions. During the mapping process, object name, spatial location, time range, or business attributes can be used for auxiliary judgment. For example, when the "Chengdong Drainage Pumping Station" in the engineering ledger has the same name and spatial location as the pumping station location in the geographic information spatial data, and the corresponding operation monitoring sequence corresponds to the same equipment number, all three types of information can be mapped to the standardized flood entity "Chengdong Drainage Pumping Station". For data with insufficient semantic matching or obvious conflicts in spatial or business attributes, unified mapping can be omitted, and these data can be retained as different object information awaiting further verification.

[0164] The fourth step is to establish the association between the triples and the mapped information from different source objects to obtain semantic alignment knowledge.

[0165] After completing the unified entity mapping, standardized flood entities in the original triples can be associated with object attributes from hydrological time-series data, geographic information spatial data, or engineering ledger data. For example, the two entities in the triple "Eastern drainage pumping station, coordinated scheduling, Southern stormwater culvert" can be associated with their corresponding engineering parameters, spatial locations, and operational statuses, respectively, thus forming semantically aligned knowledge that simultaneously contains relational knowledge and multi-source attribute information. Through this semantic alignment, the fragmentation of common-source knowledge caused by different business systems using different names, statistical methods, or data formats can be reduced, improving the consistency of multi-source flood knowledge.

[0166] In some embodiments, adaptive fusion of semantic alignment knowledge to obtain fused knowledge can be performed through the following technical steps: The first step is to obtain data from different sources that are mapped to the same standardized flood entity in the semantic alignment knowledge.

[0167] For example, for a flood-prone area, data such as real-time water depth monitoring data, manually reported water accumulation records, historical disaster ledgers, and spatial inundation range can be obtained. Since the collection methods, update frequencies, data accuracy, and completeness of data from different sources may vary, the information characteristics of each source data can be evaluated separately before fusion.

[0168] The second step is to determine the information entropy corresponding to data from different sources.

[0169] The entropy weight method can determine the degree of effective information based on the distribution characteristics of the data itself.

[0170] No. The information entropy of the source data can be represented as: ; in, Indicates the first Information entropy of the source data; Indicates the source data category index; Indicates the first The number of samples from the source data that participate in the information entropy calculation; This indicates the sample index in the source data; Indicates the first The first in the source data of the class The percentage of samples after normalization; Represents the natural logarithm operation.

[0171] During the calculation process, data from different sources can be normalized first to enable unified calculation of data with different dimensions, and then the information entropy can be calculated based on the normalized sample distribution.

[0172] The third step is to determine the corresponding fusion weights based on the information entropy of data from different sources.

[0173] No. The fusion weights corresponding to the source data of each class can be expressed as: ; in, Indicates the first The fusion weights corresponding to the source data of each category; Indicates the first Information entropy of the source data; This indicates the total number of data source categories participating in the current entity knowledge fusion; Index indicating the data source category participating in the fusion; Indicates the first Information entropy of the source data.

[0174] The fusion weights corresponding to the data from each source satisfy a normalization relationship, which can be used to control the degree of influence of different source data on the final fusion result.

[0175] Besides information entropy, when forming the features of the data to be calculated, the richness of information, the rationality of data dispersion, the real-time accuracy of monitoring, the frequency of time-series updates, the reliability of scenario adaptation, and the completeness of data can also be reflected. For source data with rich effective information, high monitoring accuracy, or relatively complete data, a relatively large fusion weight can be obtained; for data with high noise, delayed updates, or severe missing data, a relatively small fusion weight can be obtained.

[0176] The fourth step is to fuse the different source data mapped to the same standardized flood entity according to the fusion weights corresponding to the different source data to obtain fused knowledge.

[0177] For example, for the same water accumulation point, real-time monitoring water depth, manual inspection water depth, and historical experience data can be combined according to their respective fusion weights, giving the most timely and stable data a greater role in the fusion result. For numerical attributes, weighted calculations can be used to form fused attribute values; for status or textual attributes, attributes with higher confidence levels can be selected based on fusion weights and data consistency, while retaining data source information. By adaptively determining the fusion weights for data from different sources, the dilution effect of low-quality, outdated, or missing data on high-quality flood knowledge can be reduced, improving the reliability of the fused knowledge.

[0178] In some embodiments, entity disambiguation is performed on the fused knowledge to obtain disambiguated knowledge, which can be done through the following technical steps: The first step is to construct corresponding joint features based on the entity semantic features, entity category features, spatial location features, and business attribute features of the entity to be matched.

[0179] Here, the entities to be matched can represent two entities in the fused knowledge that have the same name, similar names, or may correspond to the same actual flood object. Entity disambiguation can refer to the process of distinguishing homonymous entities or merging homonymous entities based on entity information from multiple dimensions.

[0180] No. The joint features of the entities to be matched can be represented as: .in, Indicates the first The joint features of the entities to be matched; Indicates the first The semantic features of the entities to be matched; Indicates entity category characteristics; Indicates spatial location characteristics; Indicates business attribute characteristics, This indicates feature splicing.

[0181] For example, entity semantic features can be 768-dimensional semantic vectors output by a semantic coding model; entity category features can be categorized according to six types of flood entities; spatial location features can be normalized latitude and longitude or spatial coordinates; business attribute features can be encoded according to attributes such as engineering parameters, hazard type, and scheduling status.

[0182] The second step is to determine the overall similarity between the entities to be matched based on the joint features.

[0183] entity and entity The overall similarity between them can be expressed as: ; in, Representing entities and entity Overall similarity between them; Representing entities Corresponding joint features; Representing entities Corresponding joint features; This represents the inner product of two joint eigenvectors; and Let represent the norms of the two joint eigenvectors.

[0184] The third step is to determine the corresponding entities to be matched as homologous entities and perform entity fusion if the overall similarity meets the preset entity fusion conditions.

[0185] The preset entity fusion condition can be set by ensuring that the overall similarity is not less than the entity fusion threshold. For example, the entity fusion threshold... It can be set to 0.7, that is: ≥ .in, When two entities to be matched are identified as having the same origin, they can be determined to be entities with the same origin.

[0186] Entity fusion can include unifying entity names, merging entity attributes, re-associating existing triples, and removing duplicate entity nodes. For example, different business systems may use "Chengdong Drainage Station" and "Chengdong Drainage Pumping Station" to describe the same engineering facility. When the two entities are highly consistent in category, spatial location, and business attributes, and their overall similarity reaches a threshold, the two entities can be merged.

[0187] The fourth step is to retain the corresponding entities as different entities when the names of the entities to be matched are the same and the overall similarity does not meet the preset entity fusion conditions, thereby obtaining disambiguation knowledge.

[0188] For example, water conservancy projects named "Central Pumping Station" may exist in different regions. When the spatial location, engineering parameters, and business scope of two entities differ significantly, their overall similarity may be lower than the entity fusion threshold even if their text names are the same. In this case, the two entities can be retained separately, and each entity can continue to be associated with its corresponding attributes and business relationships.

[0189] By utilizing semantic, category, spatial, and business attributes simultaneously for entity disambiguation, erroneous merging caused by relying solely on entity names can be reduced, as can duplicate knowledge caused by using different names for the same entity.

[0190] In some embodiments, relation verification and temporal denoising are performed on the disambiguation knowledge to obtain standardized fused knowledge, which can be achieved through the following technical steps: The first step is to establish legality constraints between different entity types based on standardized business relationships.

[0191] Relationship legality constraints can represent a set of rules specifying what standardized business relationships are allowed between two entity categories. A constraint matrix between entity categories and allowed relationships can be established based on disaster causality, spatial affiliation, engineering linkage, emergency response, and attribute association.

[0192] For example, causal relationships such as inducing, causing, or exacerbating disasters are permissible between disaster-causing entities and disaster event entities; engineering linkage relationships such as coordinated scheduling, pipeline connection, or gate-pump linkage are permissible between water conservancy project entities; and emergency response relationships such as emergency rescue or material allocation are permissible between disaster event entities and emergency resource entities.

[0193] The second step is to use relational legality constraints to verify the legality of entity relations in the disambiguation knowledge and eliminate entity relations that do not meet the relational legality constraints.

[0194] For entities and entity Predictive relationship between The legality of a relationship can be expressed as: ; in, Indicates a valid identifier for the relationship; This indicates that the current entity relationship satisfies the relationship legality constraints; This indicates that the current entity relationship does not meet the relationship legality constraints; Representing entities and entity The relationship to be verified between them; Representing entities Entity category; Representing entities Entity category; Indicates entity category and entity category A set of standardized business relationships that are allowed to exist between them.

[0195] For example, if the model identifies a "material allocation" relationship between a disaster-causing entity and a flood control material entity, but the business relationship rules stipulate that material allocation should occur between a disaster event, a controlled and dispatched object, and emergency resources, then the relationship can be determined as not meeting the relationship legality constraints and removed.

[0196] The third step is to obtain the time information corresponding to each triplet and verify the temporal relationship between the triplets based on the time information.

[0197] Temporal denoising can identify and process knowledge that has time inversions or logical contradictions based on the temporal logic of the occurrence, development, and handling of flood disasters.

[0198] For example, the time when rainfall is triggered can be denoted as Record the time of the emergency as Record the project scheduling time as Under a normal timing relationship, the following can be satisfied: .in, Indicates the moment when rainfall is triggered; Indicates the moment the emergency occurred; Indicates the timing of the project scheduling.

[0199] Furthermore, timing conflicts can be represented as: ; in, Indicates a timing conflict flag; This indicates a temporal contradiction in the corresponding knowledge; This indicates that the corresponding knowledge satisfies the current temporal constraints.

[0200] For example, if knowledge records indicate that the emergency occurred before the triggering of rainfall, or if the engineering scheduling record time is earlier than the corresponding emergency formation time under the current business logic, this knowledge can be identified as having a timing conflict. For different business scenarios, different timing constraints can also be set according to the corresponding disaster evolution logic, not limited to the three types of time nodes mentioned above.

[0201] The fourth step is to remove or deduplicate triples that have relationship conflicts, temporal conflicts, or repetitions to obtain standardized fused knowledge.

[0202] Specifically, the process involves sequentially traversing the triples in the disambiguation knowledge, matching entity categories and relation types; eliminating triples with a relation validity flag of 0; reading the timestamps of the remaining triples and performing time-series checks; eliminating or marking triples with a time-series conflict flag of 1; and deduplicating duplicate triples where the head entity, relation, and tail entity are identical. For duplicate triples, retention can be determined based on time and confidence level. For example, more recent triples can be prioritized, or, if times are consistent, triples with higher relation classification confidence can be retained. After completing the above processing, the retained triples can be written into the flood knowledge graph as standardized fusion knowledge.

[0203] In some embodiments, incrementally updating the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph can be achieved through the following technical steps: The first step is to divide the knowledge in the flood knowledge graph into static inherent knowledge, dynamic temporal knowledge, and temporary working condition knowledge.

[0204] Static inherent knowledge can represent flood control spatial structures and basic engineering attributes that change with a low frequency over a relatively long period. Static inherent knowledge may include urban water system structures, river boundaries, pipeline parameters, flood control project locations, and administrative regions. In one implementation, it may also include fixed hazard locations. Because this type of knowledge changes with a low frequency, a relatively low-frequency update method can be used.

[0205] Dynamic time-series knowledge can represent rainfall, water conditions, and engineering operation information that continuously change over monitoring time. Dynamic time-series knowledge can include real-time rainfall, river levels, pipeline flow, and engineering operation status, as well as routine early warning levels that change with the status.

[0206] Temporary operational condition knowledge refers to knowledge generated by sudden flooding events or temporary flood control measures, which has a relatively short lifespan and high timeliness. Temporary operational condition knowledge can include extreme rainstorm emergencies, temporary sluice gate and pump scheduling, emergency rescue records, and material allocation information, as well as temporary flood control and management measures.

[0207] By classifying knowledge according to its change characteristics, different types of knowledge can be updated in a way that matches their frequency of change.

[0208] The second step is to determine the knowledge type corresponding to the newly added flood data, and then determine the corresponding update method based on the knowledge type. Static inherent knowledge is updated in batches, dynamic time-series knowledge is updated in streaming mode, and temporary working condition knowledge is updated in event-triggered mode.

[0209] Batch update can refer to updating a type of knowledge in a concentrated manner after a predetermined update cycle has been reached or a certain amount of new data has been accumulated. For example, static inherent knowledge can be updated in batches quarterly, or a batch update can be triggered after engineering renovations, changes in administrative regions, or adjustments to pipeline parameters.

[0210] Streaming updates can represent the continuous reception and processing of new data in the order it was generated. For example, dynamic time-series knowledge can be updated minute-by-minute, allowing real-time rainfall, river levels, pipeline flow, and engineering operation status to be continuously added to the map as monitoring data accumulates.

[0211] Event-triggered updates can be triggered when a corresponding flood event, dispatching action, or emergency response action is detected. For example, upon detecting extreme rainstorm risks, temporary sluice gate / pump dispatching instructions, or emergency rescue records, a corresponding temporary operational condition knowledge update can be initiated.

[0212] The third step involves entity extraction and relation mining of the newly added flood data to obtain new triples, and then performing validity checks and redundancy removal on the new triples.

[0213] Specifically, new flood data can be obtained from data sources such as flood control bulletins, dispatch documents, time-series monitoring databases, and disaster ledgers. For new text data, word segmentation, invalid word cleaning, and semantic encoding can be performed. Then, a pre-trained adaptive entity extraction model can be used to identify new standardized flood entities, and a dynamic graph relationship mining model can be used to identify new standardized business relationships, generating new triples according to the structure of head entity, relationship, and tail entity.

[0214] For newly added triples, the established business relationship rules can be used to perform legality verification, remove newly added triples whose entity categories and relationship types are obviously mismatched, and perform preliminary deduplication of triples that are duplicates of the currently added data.

[0215] By performing validity checks and deduplication before adding new knowledge to the graph, we can reduce the direct entry of erroneous and duplicate knowledge into the existing graph.

[0216] The fourth step is to match the newly added triples after verification with the existing knowledge in the flood knowledge graph to determine the new knowledge and the graph substructures related to the new knowledge.

[0217] Existing knowledge can represent entities, relationships, and attributes already stored in the flood knowledge graph before the incremental update. New knowledge can include new entities, relationships, or attributes that did not originally exist in the graph, or it can include state change information corresponding to existing entities or relationships.

[0218] The head and tail entities in a newly added triple can be matched with existing entities. If no matching entity exists, the corresponding entity can be identified as a newly added entity; if a matching entity already exists, but the corresponding attributes or relationships have changed, the corresponding content can be identified as an update of existing knowledge.

[0219] A graph substructure can represent a graph structure region centered on the entity nodes involved in the newly added knowledge, including adjacent entity nodes, related entity relationships, and corresponding local parameters. By determining the graph substructure, the scope of subsequent updates can be limited.

[0220] The fifth step involves updating the entity nodes, entity relationships, network parameters, and graph topology in the graph substructure according to the update method corresponding to the newly added knowledge, thereby obtaining a dynamic flood knowledge graph.

[0221] Entity node updates can include adding new entity nodes or updating the attributes of existing entity nodes; entity relationship updates can include adding new business relationships, modifying the status of existing relationships, or canceling invalid relationships; network parameter updates can make local adjustments to model parameters related to newly added knowledge; graph topology updates can redetermine the local node connection structure based on newly added entities and relationships.

[0222] After incremental updates, the adjacency status of entities and the temporal attributes of relationships related to the newly added knowledge can be refreshed simultaneously. New knowledge that conforms to business rules can be written into the flood knowledge graph, forming a dynamic flood knowledge graph that can be continuously updated with changes in rainfall, water levels, engineering conditions, emergencies, and scheduling.

[0223] For example, the structure for constructing and incrementally updating a flood knowledge graph can refer to... Figure 3 As shown. See also Figure 3 The structure can include an input layer, an entity relationship extraction layer, a rule constraint layer, and an output layer. The layers are linked according to the processing steps of multi-source flood data input, entity relationship extraction, knowledge graph construction, and incremental updates.

[0224] The input layer can include flood text data, hydrological time-series data, geographic information spatial data, and engineering ledger data, with each type of data serving as a parallel data source. Flood text data can be input into an adaptive entity extraction model; hydrological time-series data and geographic information spatial data can be used for relationship mining in a dynamic graph relationship mining model; and geographic information spatial data and engineering ledger data can be used for flood knowledge graph construction and incremental updates.

[0225] The entity relationship extraction layer can include an adaptive entity extraction model, a dynamic graph relationship mining model, and flood knowledge graph construction and incremental updates. The adaptive entity extraction model can include a domain semantic enhancement layer, a temporal feature extraction layer, and a label constraint decoding layer, used for domain semantic enhancement, bidirectional temporal feature extraction, and label constraint decoding, respectively, to obtain standardized flood entities. The dynamic graph relationship mining model can process based on standardized flood entities and can include node representation generation, candidate connection edge construction, and standardized business relationship determination, used to generate node representations corresponding to entity nodes, construct candidate connection edges between entity nodes, and determine standardized business relationships. The flood knowledge graph construction and incremental updates can process based on standardized flood entities, standardized business relationships, and relevant multi-source flood data, and can include triple transformation, graph construction, and incremental updates to construct a flood knowledge graph and update it according to newly added flood data.

[0226] The rule constraint layer can include entity annotation rules, business relationship system, and relationship verification and temporal denoising, each corresponding to different processing stages. Entity annotation rules can be used to constrain the entity recognition process in the adaptive entity extraction model; the business relationship system can be used to constrain the standardized business relationship determination process in the dynamic graph relationship mining model; and relationship verification and temporal denoising can be used to verify the legality of relationships and ensure temporal consistency of knowledge formed during the construction and incremental updating of the flood knowledge graph.

[0227] The output layer can include standardized flood entities, standardized business relationships, and a dynamic flood knowledge graph. Standardized flood entities can be further used to determine standardized business relationships, and these entities and relationships can be further used to construct the flood knowledge graph, which is then incrementally updated to obtain the dynamic flood knowledge graph.

[0228] In some embodiments, updating entity nodes, entity relationships, network parameters, and graph topology in the graph substructure can be performed through the following technical steps: The first step is to add the new entity nodes to the graph substructure based on the new knowledge, and update the entity information corresponding to the entity nodes that match the existing entity nodes.

[0229] The newly added entity node can represent a standardized flood entity identified in the newly added flood data that does not have a corresponding entity in the existing flood knowledge graph. For a newly added entity node, the entity name, entity category, spatial location, time attribute, and corresponding business attribute can be written.

[0230] For existing entity nodes, when new flood data reflects changes in their operational status, monitoring attributes, or working conditions, the corresponding entity information can be updated. For example, if a drainage pumping station was originally in standby mode, and new scheduling data indicates that it has been activated, then the operational status attribute of the corresponding entity node can be updated without creating a duplicate pumping station entity.

[0231] The second step is to freeze historical network parameters that are unrelated to the new knowledge and to incrementally update the network parameters that are related to the new knowledge.

[0232] Historical network parameters can represent model parameters that have been trained using historical flood knowledge and are not directly related to the newly added knowledge. Incremental updates can be performed using class-incremental learning or other local parameter update methods.

[0233] The global parameters of the model are represented as: .in, Represents the set of global parameters of the model participating in incremental training; This represents the set of historical network parameters that correspond to existing historical knowledge and remain frozen during this round of training. This represents the set of network parameters that are relevant to the new knowledge and can be updated during this training round.

[0234] The incremental training round index is uniformly denoted as... The network parameters related to the newly added knowledge can be updated as follows: ; ; in, Indicates the first The current values ​​of the network parameters to be updated in each incremental training round; This indicates the network parameter values ​​to be updated for the next training round; Indicates the first Historical network parameters from each training round; Indicates the incremental training learning rate; This represents the loss function corresponding to incremental training; This represents the operation of calculating the gradient of the loss function with respect to the network parameters to be updated. For example, It can be adopted By keeping the historical network parameters unchanged and only updating the gradients for parameters related to newly added knowledge, the parameter overlay caused by training with new knowledge on existing flood-water relationship patterns can be reduced.

[0235] The third step is to obtain historical triplet samples corresponding to historical flood knowledge, and then jointly train the historical triplet samples with the newly added samples corresponding to the newly added knowledge.

[0236] Among them, joint training can adopt the method of historical sample memory playback. In each round of incremental training, a portion of representative historical samples are selected from the historical triplet and combined with the currently added samples to form a training batch.

[0237] The combined loss of new and old knowledge can be expressed as: ; in, This indicates the loss from combined training of new and old knowledge; This indicates the loss corresponding to the newly added flood ternary sample; This represents the loss corresponding to historical cache triplet samples; This represents the balance weight between newly added knowledge and historical knowledge, and For example, A value of 0.7 can be used to make the incremental training fit the newly added flood knowledge relatively well, while limiting the significant degradation of the model performance corresponding to historical knowledge through historical sample loss.

[0238] For newly added relationships with a limited number of samples, such as extreme rainstorms, rare emergencies, or special gate pump linkages, feature interpolation enhancement can be performed on the newly added samples. The enhanced features of rare samples can be represented as: ; in, This represents the enhanced feature vector of rare samples; This represents the feature vector corresponding to the original rare relation sample; The feature vector of the reference sample that is adjacent to the original rare relation sample in the feature space; Denotes the interpolation coefficients, satisfying For example, It can be set to 0.5.

[0239] By interpolating features between the original rare relation samples and neighboring samples, the feature coverage corresponding to extreme conditions and small sample relations can be increased, reducing the relation recognition bias caused by severe imbalance in incremental training samples.

[0240] Furthermore, temporal location codes can be generated based on the time information corresponding to newly added knowledge, enabling entities and relationships in the graph to reflect different time conditions. The normalized knowledge timestamp is then recorded as... The dimension index of the temporal location encoding is denoted as The total dimension of the temporal location encoding is denoted as Then the timing position encoding can be represented as: ; ; in, Represents knowledge timestamps In the Temporal location encoding corresponding to each dimension; Represents knowledge timestamps In the Temporal location encoding corresponding to each dimension; Represents the sine function; Represents the cosine function; This represents a normalized knowledge timestamp; Indicates the time-series coding dimension index; This represents the total dimension of the temporal coding. For example, You can take 256.

[0241] The fourth step is to redetermine the adjacency relationships between entity nodes related to the newly added knowledge based on the network parameters after joint training.

[0242] After joint training is completed, the semantic features of the entity nodes involved in the newly added knowledge can be reacquired, and the semantic similarity between these entity nodes and their adjacent entity nodes can be recalculated according to the aforementioned semantic similarity calculation method.

[0243] When the recalculated semantic similarity changes, it is possible to re-determine whether the corresponding entity node belongs to a dynamically adjacent entity. For example, as rainfall continues to intensify, the semantic and business association between a drainage pumping station and a certain flooding event may be strengthened, thereby establishing new candidate connection relationships; when the emergency ends and normal operation resumes, the corresponding relationships can also be adjusted.

[0244] The fifth step is to update the corresponding entity relationships and graph topology based on the newly determined adjacency relationships to complete the update of the graph substructure.

[0245] The corresponding candidate connection edges can be refreshed based on the newly determined dynamic adjacency entity set, and then the updated entity relationships can be determined using a dynamic graph relationship mining model. Simultaneously, the temporal attributes of the relationship edges can be updated, and the newly added or updated compliance knowledge can be written into the dynamic flood knowledge graph. In one iterative control method, an early stopping mechanism can be adopted: incremental training stops when the F1 score on the validation set no longer improves after 15 consecutive training rounds, and the current network parameters and dynamic graph topology are saved.

[0246] By adding new entities for updates, updating local network parameters, jointly training historical samples, and redetermining adjacency relationships, we can absorb new flood knowledge while reducing the coverage of historical knowledge, and make the local graph structure change with the latest flood conditions.

[0247] In some embodiments, flood risk analysis results can be obtained by performing flood factor risk extrapolation based on a dynamic flood knowledge graph, which can be achieved through the following technical steps: The first step is to obtain the entity attributes corresponding to multiple standardized flood entities from the dynamic flood knowledge graph, and then construct flood risk analysis features based on the entity attributes.

[0248] Among them, flood risk analysis features can represent quantitative or coded features that reflect the occurrence, development, and influencing factors of floods. For example, attributes such as hourly rainfall, rainfall duration, rainfall intensity, and pre-flood soil moisture content can be obtained from disaster-causing factor entities; topographic slope, elevation, road surface hardening rate, and green space coverage can be obtained from spatial point entities or topographic data; pipeline connectivity, pipeline flow rate, pumping station drainage capacity, and project operation status can be obtained from water conservancy project entities; and water depth, spatial extent of water accumulation, and historical water accumulation records can be obtained from disaster event entities. For risk analysis features with different dimensions, normalization or standardization can be performed before inputting them into the model to ensure that each feature has a similar numerical scale, and the feature sequences corresponding to the flood evolution process can be organized chronologically.

[0249] The second step involves inputting the flood risk analysis features into a temporal attention neural network to explore the nonlinear coupling relationship between each flood risk analysis feature and the flood risk.

[0250] Temporal attention neural networks can represent models that assign corresponding attention weights based on the importance of different time positions or different input features, and learn the coupled features of multidimensional flood elements using the nonlinear mapping relationship of the neural network. For example, a temporal attention backpropagation neural network can be used.

[0251] Nonlinear coupling relationships can represent the non-simple linear superposition relationship formed when two or more flood risk analysis features act together on flood risk. For example, under the same rainfall intensity, different road surface hardening rates, different pipe network unobstructedness, and different pumping station drainage capacities may correspond to different water accumulation risks; the model can learn this type of combined relationship through multi-layer nonlinear transformations.

[0252] Model training can use mean squared error (MSE) as the loss function: ; in, Indicates the predicted losses from flood risk; This indicates the total number of samples used for training the current model; Indicates the training sample index; Indicates the first The model predicts flood risk values ​​for each training sample; Indicates the first The true value of the measured flood risk corresponding to each training sample.

[0253] The true value of measured flood risk can be determined based on measured disaster data, waterlogging monitoring data, and historical disaster statistics. During training, parameters can be iterated based on the error between the model's predicted values ​​and the measured true values, allowing the model to gradually fit the nonlinear coupling relationship between various flood risk analysis features and flood risk.

[0254] For example, historical samples covering different rainfall conditions such as light rain, moderate rain, heavy rain, rainstorm, torrential rain, and short-term heavy rainfall can be used for training, so that the model can learn the risk differences under different rainfall conditions, different underlying surface conditions, and different engineering operation states.

[0255] The third step is to determine the risk contribution weight of each flood-causing factor and the flood risk prediction results based on the nonlinear coupling relationship.

[0256] Risk contribution weights represent the relative impact of corresponding flood-causing factors on flood risk analysis results under current flood conditions. The risk contribution weights for each flood-causing factor can be determined based on time-series attention weights, internal model feature contributions, or normalized correlation weights.

[0257] The flood risk prediction results can represent the risk prediction information output by the temporal attention neural network based on the current flood risk analysis features. It can be represented by continuous risk values, normalized risk values, or risk states further determined based on continuous risk values.

[0258] For example, under the condition of 42 mm of cumulative rainfall in one hour, urban roads experiencing flooding with a maximum water depth of 0.35 m, the risk contribution weight corresponding to short-term rainfall intensity is 67.2%, the risk contribution weight corresponding to the hardened road surface is 18.1%, the risk contribution weight corresponding to the drainage capacity of the pipe network is 11.5%, and the combined risk contribution weight corresponding to other flood factors such as terrain slope and pump station scheduling is 3.2%. Therefore, it can be determined that short-term rainfall intensity has the greatest risk contribution under this condition.

[0259] Under different rainfall conditions, terrain conditions, and engineering operation states, the risk contribution weight of each disaster-causing factor can be recalculated. For example, when the drainage network is severely blocked, the risk contribution weight corresponding to the network patency may increase; after the pumping station starts discharging at high flow rates, the risk effect corresponding to the engineering drainage capacity may change.

[0260] The fourth step is to use the contribution weight of each risk as the contribution level of the corresponding flood-causing factor, and to obtain the flood risk analysis results based on the contribution level and the flood risk prediction results.

[0261] The results of flood risk analysis can simultaneously include the flood risk prediction results corresponding to the current or under-analyzed conditions of the target flood control area, as well as the contribution degree of different flood-causing factors, so that the risk analysis results can reflect both the flood risk status and the main influencing factors that form the risk status.

[0262] For example, for the above-mentioned case of 42 mm of cumulative rainfall in 1 hour, the current flood risk prediction results can be recorded in the flood risk analysis results, and the contribution of short-term rainfall intensity, hardened road surface, drainage capacity of pipe network and other flood factors can be recorded at the same time, so as to further determine the main disaster-causing factors.

[0263] In some risk analysis scenarios, the established entity relationships within a dynamic flood knowledge graph can be used to explain the formation process of flood risks. For example, the graph can create a chain of risks, such as heavy rainfall, insufficient soil infiltration, increased surface runoff, rising river levels, and water accumulation in low-lying areas. By combining monitoring data, multi-source data fusion results, water level prediction results, and graph relationships, further analysis can be conducted on information related to the current flood risk, such as the inundation area, inundation depth, flood arrival time, affected population, and risks to critical infrastructure.

[0264] By incorporating both flood risk prediction results and the contribution of flood-causing factors into the flood risk analysis results, the problem of outputting only a single risk value without reflecting the source of risk can be reduced. This allows the risk analysis results to simultaneously describe the risk status and the main influencing factors, thereby improving the accuracy and interpretability of flood risk analysis.

[0265] Furthermore, in the exemplary embodiments of this disclosure, a flood map construction system based on adaptive entity relationship extraction is also provided. (Refer to...) Figure 4 As shown, the flood map construction system 400 based on adaptive entity relationship extraction includes: a data acquisition module 410, an entity extraction module 420, a relationship mining module 430, a map construction module 440, an incremental update module 450, and a risk simulation module 460. Among them: The data acquisition module 410 can be used to acquire multi-source flood data of the target flood control area, including flood text data and hydrological time series data. The entity extraction module 420 can be used to input flood text data into the adaptive entity extraction model to obtain multiple standardized flood entities; The relationship mining module 430 can be used to construct a dynamic entity graph based on multiple standardized flood entities, input the dynamic entity graph into the dynamic graph relationship mining model, and obtain the standardized business relationships between entities; The graph construction module 440 can be used to construct a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data. The incremental update module 450 can be used to incrementally update the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph. The risk simulation module 460 can be used to perform risk simulation analysis of flood factors based on a dynamic flood knowledge graph, and obtain flood risk analysis results.

[0266] The specific details of each module in the flood map construction system based on adaptive entity relation extraction have been described in detail in the corresponding flood map construction method based on adaptive entity relation extraction, so they will not be repeated here.

[0267] It should be noted that although several modules or units of the flood map construction system based on adaptive entity relationship extraction have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0268] Furthermore, in an exemplary embodiment of this disclosure, an electronic device capable of implementing the above-described flood map construction method based on adaptive entity relationship extraction is also provided.

[0269] Those skilled in the art will understand that various aspects of this disclosure can be implemented as a system, method, or program product. Therefore, various aspects of this disclosure can be embodied in the following forms: a completely hardware embodiment, a completely software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0270] The following reference Figure 5 To describe an electronic device 500 according to such an embodiment of the present disclosure. Figure 5 The electronic device 500 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0271] like Figure 5 As shown, the electronic device 500 is presented in the form of a general-purpose computing device. The components of the electronic device 500 may include, but are not limited to: at least one processing unit 510, at least one storage unit 520, a bus 530 connecting different system components (including storage unit 520 and processing unit 510), and a display unit 540.

[0272] The storage unit stores program code that can be executed by the processing unit 510, causing the processing unit 510 to perform the steps described in the "Detailed Description" section of this specification according to various exemplary embodiments of this disclosure. The storage unit 520 may include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) unit 521 and / or a cache memory unit 522, and may further include a read-only memory unit (ROM) unit 523.

[0273] Storage unit 520 may also include a program / utility 524 having a set (at least one) program module 525, such program module 525 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0274] Bus 530 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0275] Electronic device 500 can also communicate with one or more external devices 570 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 500, and / or with any device that enables electronic device 500 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 550. Furthermore, electronic device 500 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 560. As shown, network adapter 560 communicates with other modules of electronic device 500 via bus 530. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0276] Through the description of the above embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal system, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0277] In exemplary embodiments of this disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible embodiments, various aspects of this disclosure may also be implemented as a program product including program code, which, when the program product is run on a terminal device, causes the terminal device to perform the steps according to various exemplary embodiments of this disclosure described in the "Detailed Description" section above.

[0278] The program product for implementing the above-described flood map construction method based on adaptive entity relation extraction, according to embodiments of the present disclosure, may employ a portable compact disk read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In the present disclosure, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0279] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0280] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0281] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0282] Program code for performing the operations of this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0283] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of this disclosure and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0284] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed 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 embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0285] 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 method for constructing flood maps based on adaptive entity relation extraction, characterized in that, include: Acquire multi-source flood data for the target flood control area, wherein the multi-source flood data includes flood text data and hydrological time-series data; The flood text data is input into an adaptive entity extraction model to obtain multiple standardized flood entities; A dynamic entity graph is constructed based on multiple standardized flood entities. The dynamic entity graph is then input into a dynamic graph relationship mining model to obtain standardized business relationships between entities. Construct a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data; The flood knowledge graph is incrementally updated based on the newly added flood data to obtain a dynamic flood knowledge graph. Based on the dynamic flood knowledge graph, the risk of flood factors is extrapolated and analyzed to obtain the flood risk analysis results.

2. The flood map construction method based on adaptive entity relation extraction according to claim 1, characterized in that, The process involves inputting the flood text data into an adaptive entity extraction model to obtain multiple standardized flood entities, including: The flood text data is segmented and positionally encoded, and the domain-enhanced features corresponding to proper nouns in the flood field are fused to obtain the text input features; The text input features are input into the domain semantic enhancement layer of the adaptive entity extraction model for semantic encoding to obtain domain semantic features. The domain semantic features are input into the temporal feature extraction layer for bidirectional temporal encoding to obtain bidirectional temporal features that integrate contextual information. The bidirectional temporal features are input into the label constraint decoding layer, and the label transfer is globally constrained and decoded according to the entity annotation rules to obtain multiple standardized flood entities.

3. The flood map construction method based on adaptive entity relation extraction according to claim 2, characterized in that, The global constraint decoding of the tag transfer according to the entity annotation rules yields multiple standardized flood entities, including: Obtain pre-established entity types and entity labeling rules in the flood domain. The entity types in the flood domain include disaster-causing factor entities, water conservancy project entities, disaster event entities, spatial location entities, emergency resource entities, and control and dispatch entities. The entity labeling rules include the entity's starting position, the entity's internal position, and the labels corresponding to non-entity positions. Determine the legitimate transfer relationships between tags based on the entity types in the flood control field and the entity labeling rules; Global constraints are applied to the candidate label sequences corresponding to the bidirectional temporal features based on the legal transition relationships; The target entity label sequence is determined from the candidate label sequence that satisfies the global constraints, and multiple standardized flood entities are obtained based on the target entity label sequence.

4. The flood map construction method based on adaptive entity relation extraction according to claim 1, characterized in that, The step of constructing a dynamic entity graph based on multiple standardized flood entities, and inputting the dynamic entity graph into a dynamic graph relationship mining model to obtain standardized business relationships between entities includes: Using multiple standardized flood entities as entity nodes, corresponding node representations are generated based on the semantic features, category features, and location features of each entity node; Calculate the semantic similarity between each entity node, establish candidate connection edges between the entity nodes based on the semantic similarity, and construct the dynamic entity graph based on the entity nodes and the candidate connection edges; Based on the node representation of each entity node and the relation type features corresponding to different relation types, the association weight of each candidate connection edge under different relation types is calculated through a multi-head relation attention mechanism, and the corresponding relation features are obtained based on the association weights. The relationship features are classified to obtain the standardized business relationships corresponding to each candidate connection edge.

5. The flood map construction method based on adaptive entity relation extraction according to claim 1, characterized in that, The construction of a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data includes: The head entity and tail entity are determined based on the two standardized flood entities associated with each of the standardized business relationships, and the head entity, the standardized business relationship, and the tail entity are combined to form a triple; Multi-source heterogeneous knowledge is formed based on the aforementioned multi-source flood data; Semantic alignment is performed on the triples and the multi-source heterogeneous knowledge to obtain semantically aligned knowledge. The semantic alignment knowledge is adaptively fused to obtain fused knowledge; Entity disambiguation is performed on the fused knowledge to obtain disambiguated knowledge; The disambiguation knowledge is subjected to relation verification and temporal denoising to obtain standardized fused knowledge, and the flood knowledge graph is constructed based on the standardized fused knowledge.

6. The flood map construction method based on adaptive entity relation extraction according to claim 5, characterized in that, The adaptive fusion of the semantic alignment knowledge to obtain fused knowledge includes: Obtain different source data from the semantic alignment knowledge that are mapped to the same standardized flood entity; Determine the information entropy corresponding to each of the different data sources; The corresponding fusion weights are determined based on the information entropy of the data from each of the different sources. Based on the fusion weights corresponding to the different sources of data, the different sources of data mapped to the same standardized flood entity are fused to obtain the fused knowledge.

7. The flood map construction method based on adaptive entity relation extraction according to claim 5, characterized in that, The entity disambiguation of the fused knowledge to obtain disambiguated knowledge includes: Construct corresponding joint features based on the entity semantic features, entity category features, spatial location features, and business attribute features of the entity to be matched; The overall similarity between entities to be matched is determined based on the joint features. If the overall similarity meets the preset entity fusion conditions, the corresponding entities to be matched are determined as homologous entities and entity fusion is performed. If the names of the entities to be matched are the same and the overall similarity does not meet the preset entity fusion conditions, the corresponding entities to be matched are retained as different entities, thus obtaining the disambiguation knowledge.

8. The flood map construction method based on adaptive entity relation extraction according to claim 1, characterized in that, The incremental update of the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph includes: The knowledge in the flood knowledge graph is divided into static inherent knowledge, dynamic temporal knowledge, and temporary working condition knowledge. The knowledge type corresponding to the newly added flood data is determined, and the corresponding update method is determined according to the knowledge type. The static inherent knowledge is updated in batches, the dynamic time-series knowledge is updated in streaming, and the temporary working condition knowledge is updated in an event-triggered manner. Entity extraction and relation mining are performed on the newly added flood data to obtain new triples, and the new triples are then validated for validity and deduplicated for redundancy. The newly added triples after verification are matched with the existing knowledge in the flood knowledge graph to determine the new knowledge and the graph substructures related to the new knowledge. According to the update method corresponding to the newly added knowledge, the entity nodes, entity relationships, network parameters and graph topology in the graph substructure are updated to obtain the dynamic flood knowledge graph.

9. The flood map construction method based on adaptive entity relation extraction according to claim 1, characterized in that, The flood risk analysis based on the dynamic flood knowledge graph yields the following results: Entity attributes corresponding to multiple standardized flood entities are obtained from the dynamic flood knowledge graph, and flood risk analysis features are constructed based on the entity attributes; The flood risk analysis features are input into a temporal attention neural network to explore the nonlinear coupling relationship between each flood risk analysis feature and flood risk. Based on the aforementioned nonlinear coupling relationship, the risk contribution weights of each flood-causing factor and the flood risk prediction results are determined. Each risk contribution weight is used as the contribution level of the corresponding flood-causing factor, and the flood risk analysis result is obtained based on the contribution level and the flood risk prediction result.

10. A flood map construction system based on adaptive entity relation extraction, used to implement the flood map construction method based on adaptive entity relation extraction as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition module is used to acquire multi-source flood data of the target flood control area, wherein the multi-source flood data includes flood text data and hydrological time series data; The entity extraction module is used to input the flood text data into the adaptive entity extraction model to obtain multiple standardized flood entities. The relationship mining module is used to construct a dynamic entity graph based on multiple standardized flood entities, input the dynamic entity graph into the dynamic graph relationship mining model, and obtain the standardized business relationships between entities. The graph construction module is used to construct a flood knowledge graph based on multiple standardized flood entities, standardized business relationships, and multi-source flood data. The incremental update module is used to incrementally update the flood knowledge graph based on newly added flood data to obtain a dynamic flood knowledge graph. The risk simulation module is used to perform risk simulation analysis of flood factors based on the dynamic flood knowledge graph, and obtain flood risk analysis results.