Knowledge-driven underground space information retrieval method, system and equipment

Abstract

The present invention provides a knowledge-driven underground space information retrieval method, system and device, which includes: performing intent analysis on user instructions to identify key entity information; retrieving key entity information in the underground space entity knowledge graph, and traversing its characteristic attributes, association relationships and relationships with other entity nodes based on graph relationships to obtain entity information retrieval results; retrieving related entities from the domain knowledge graph based on the key entity information and traversing related entities and association relationships along the graph relationships to obtain graph-side local retrieval results; matching relevant community summaries based on the domain knowledge graph as context data based on the key entity information, calling a large language model to generate graph-side global retrieval results based on the context data; matching the vector representation in the semantic knowledge base based on the graph retrieval results to obtain semantic-side retrieval results; fusing the retrieval results, and inputting the fused retrieval results into the large language model to generate final retrieval results.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and underground space information management, and in particular to a knowledge-driven underground space information retrieval method, system and device. Background Art

[0002] Urban underground space is rich in resources, including underground pipelines, underground buildings, etc. Effective management and utilization of underground space geographic information is of great significance to the operation and maintenance of underground space.

[0003] The existing technologies mainly include the following underground space information management technologies: Urban underground pipeline management system based on geographic information system (GIS), which uses GIS technology to establish an underground pipeline database, record the spatial location and attribute information of the pipeline, and perform spatial analysis and query. Underground space pipeline network modeling technology based on solid geometry method, which designs the system concept model and data organization model of the underground space pipeline network based on solid geometry method and three-dimensional topological integration model, uses adjacency table to store topological relationships, and finally stores data in a raster-vector integrated data structure. Domain knowledge management based on knowledge graph, the main technical solution is to perform semi-automatic triple extraction through manually defined ontology model, use graph database to store entity relationships, and use large language model and semantic matching for knowledge query and reasoning.

[0004] Because underground space operations and maintenance involve complex engineering structures, multi-source heterogeneous data, and dynamic risk factors, existing knowledge management technologies, such as GIS-based management systems and modeling solutions, lack semantic understanding and knowledge reasoning capabilities. While they can effectively express and manage spatial information, these lack the ability to effectively manage and utilize underground pipeline knowledge. Semantic knowledge bases and knowledge graphs, while capable of storing and reasoning knowledge, lack the ability to express spatial information, making it difficult to effectively manage and utilize underground space geographic information. Furthermore, unstructured text, such as national standards, local regulations, and accident cases, is stored in a dispersed manner, lacking unified semantic associations. Furthermore, insufficient unstructured data extraction capabilities result in inefficient cross-document knowledge retrieval. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a knowledge-driven underground space information retrieval method, system and device that overcome the above problems or at least partially solve the above problems.

[0006] One aspect of the present invention provides a knowledge-driven underground space information retrieval method, the method comprising:

[0007] The large language model is used to identify and parse the key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects;

[0008] The spatiotemporal geographic object is retrieved in a preset underground space entity knowledge graph, and the characteristic attributes, association relationships, and relationships with other entity nodes of the spatiotemporal geographic object are traversed based on the knowledge graph relationship of the spatiotemporal geographic object to obtain entity information retrieval results; the underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space field through a knowledge graph;

[0009] Relevant entities are retrieved from a preset domain knowledge graph based on the key entity information, and all relevant entities and associations are traversed along the graph relationship chain to obtain local search results on the graph side; the domain knowledge graph includes multiple communities, each of which is formed by graph clustering of entity-relationship-entity triples extracted from preset underground space domain knowledge data and empirical knowledge data;

[0010] Match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data according to the key entity information, and call a large language model to generate a graph-side global search result based on the context data;

[0011] Use the local search results on the graph side and / or the global search results on the graph side as vector search input, match the corresponding text unit vector representation from the preset semantic knowledge base, and obtain the semantic search results;

[0012] Fusing entity information retrieval results, graph-side local retrieval results, graph-side global retrieval results, and semantic-side retrieval results to form a fused retrieval result;

[0013] The fused retrieval results are input into the large language model to generate the final retrieval results.

[0014] Another aspect of the present invention further provides a knowledge-driven underground space information retrieval system, the system comprising:

[0015] The instruction parsing module is used to call the large language model to identify and parse the key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects;

[0016] An entity information retrieval module is used to retrieve the spatiotemporal geographic object in a preset underground space entity knowledge graph, and traverse the characteristic attributes, association relationships, and relationships with other entity nodes of the spatiotemporal geographic object based on the knowledge graph relationship of the spatiotemporal geographic object to obtain entity information retrieval results; the underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space field through the knowledge graph;

[0017] A graph local retrieval module is used to retrieve relevant entities from a preset domain knowledge graph based on the key entity information, and traverse all relevant entities and associations along the graph relationship chain to obtain graph-side local retrieval results. The domain knowledge graph includes multiple communities, each of which is formed by graph clustering of entity-relationship-entity triples extracted from preset underground space domain knowledge data and empirical knowledge data.

[0018] A graph global search module is used to match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data based on the key entity information, and to call a large language model to generate graph-side global search results based on the context data;

[0019] The semantic retrieval module is used to use the local retrieval results on the graph side and / or the global retrieval results on the graph side as vector retrieval inputs, match the corresponding text unit vector representations from the preset semantic knowledge base, and obtain the semantic retrieval results;

[0020] A fusion module is used to fuse entity information retrieval results, local retrieval results on the graph side, global retrieval results on the graph side, and retrieval results on the semantic side to form a fused retrieval result;

[0021] The retrieval result generating unit is used to input the fused retrieval results into the large language model to generate the final retrieval results.

[0022] Another aspect of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the knowledge-driven underground space information retrieval method described above are implemented.

[0023] Another aspect of the present invention provides a computer program product having a computer program stored thereon, which implements the steps of the above-mentioned knowledge-driven underground space information retrieval method when executed by a processor.

[0024] The knowledge-driven underground space information retrieval method, system and equipment provided by the embodiments of the present invention combine the advantages of underground space entity knowledge graph, semantic knowledge base and domain knowledge graph. By constructing an information model driven by both normative guidelines and empirical knowledge, and designing a dual-channel retrieval framework of semantic knowledge base and knowledge graph, a balance between accuracy and interpretability of retrieval results is achieved, the depth and accuracy of knowledge reasoning of large language models are improved, and comprehensive data knowledge support and decision support are provided for underground space information retrieval and operation and maintenance.

[0025] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0027] Figure 1 A flowchart of a knowledge-driven underground space information retrieval method provided by an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of a pipeline entity knowledge graph provided by an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of the underground space domain knowledge graph proposed in an embodiment of the present invention;

[0030] Figure 4 This is a structural block diagram of the knowledge-driven underground space information retrieval system proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0032] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0033] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by those skilled in the art in the art to which this invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless specifically defined, should not be interpreted in an idealized or overly formal sense.

[0034] Figure 1 The flowchart of the knowledge-driven underground space information retrieval method according to one embodiment of the present invention is schematically shown. Figure 1 The knowledge-driven underground space information retrieval method of the embodiment of the present invention specifically includes the following steps:

[0035] S11. Calling a large language model to identify and parse key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects.

[0036] Specifically, when a user issues an instruction for an underground space operation and maintenance scenario, the current instruction intent is understood by calling a large language model. Optionally, the instruction intent includes general questions, specification query questions, fault analysis questions, scenario decision questions, etc. Key entity information includes spatiotemporal geographic objects (such as pipelines, pipe points, buildings, etc.) and other related entities (such as time, events, attributes, etc.) in the instruction. For example, for the instruction "What does the maintenance specification of pipeline PID-2007567 include", the spatiotemporal geographic object "Pipeline PID-2007567" is identified, and "Maintenance specification" is an entity word other than the spatiotemporal geographic object. The extracted entities include "pipeline" and "maintenance specification", and the instruction intent is parsed into a specification query class.

[0037] S12. Retrieve the spatiotemporal geographic object from a preset underground space entity knowledge graph, and traverse the spatiotemporal geographic object's characteristic attributes, associated relationships, and relationships with other entity nodes based on the spatiotemporal geographic object's knowledge graph relationships to obtain entity information retrieval results. The underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space domain through a knowledge graph. Feature attributes include spatial features, temporal features, and attribute features.

[0038] This example uses a pipeline as an example. The spatial feature is the absolute digital position of the pipeline, including longitude (X), latitude (Y), and depth (Z). These coordinates can be obtained through GPS or GIS systems and expressed using the WGS84 coordinate system.

[0039] Time features include a construction timestamp and a most recent detection time. The timestamp can be expressed using the ISO 8601 standard as T=YYYY-MM-DDTHH:MM:SSZ. For example, 2007-09-06T16:51:00 indicates 4:51:00 p.m. on September 16, 2007.

[0040] Attribute features include material, pipe diameter, length, and construction unit. These attributes can be stored and queried by establishing a database table or XML / JSON format.

[0041] Association relationships include temporal associations, spatial associations, and semantic associations. Temporal associations refer to establishing temporal associations between spatiotemporal geographic objects, such as pipeline construction time, maintenance time, and usage time, and using the event-state mechanism to express spatiotemporal changes. Spatial associations refer to establishing spatial associations between spatiotemporal geographic objects, such as the connection relationship between pipelines and pipe points, the intersection relationship between pipelines, the distance relationship between pipelines and buildings, and using spatial topological relationships to describe them. Semantic associations refer to establishing semantic associations between spatiotemporal geographic objects, and defining the logical relationships between concepts through ontology, such as the relationship between pipeline type and function, the relationship between building type and use, etc., including "inheritance" and "belonging to".

[0042] Specifically, after identifying spatiotemporal geographic objects from underground space operation and maintenance scenario instructions, the corresponding spatiotemporal geographic objects are retrieved from the underground space entity knowledge graph based on the identified spatiotemporal geographic objects. The four-dimensional features of the spatiotemporal geographic objects, namely spatial features, temporal features, attribute features, and association relationships, as well as their relationships with other entity nodes, are traversed along the graph relationships. For example, for the pipeline entity "PID-2007567," its spatial features (longitude, latitude, depth), temporal features (construction time, maintenance time), attribute information (material, diameter, length), and association relationships with other entities (such as pipeline points and buildings) are retrieved.

[0043] S13. Retrieve relevant entities from the preset domain knowledge graph based on the key entity information, and traverse all relevant entities and association relationships along the graph relationship chain to obtain local retrieval results on the graph side; the domain knowledge graph includes multiple communities, each community is formed by graph clustering of entity-relationship-entity triples extracted from the preset underground space domain knowledge data and experience knowledge data, and each community has a corresponding community summary.

[0044] S14. Match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data according to the key entity information, and call the large language model to generate graph-side global retrieval results based on the context data.

[0045] S15. Using the local search results and / or global search results from the atlas as vector search inputs, the corresponding text unit vector representations are matched from a preset semantic knowledge base to obtain semantic search results. The semantic knowledge base includes a preset domain knowledge base and an experience knowledge base. The semantic knowledge base is formed by semantically parsing pre-collected vertical domain data of underground space and generating structured text data.

[0046] In this embodiment, semantic side retrieval is to match the text unit embeddings in the domain knowledge base and the experience knowledge base according to the graph retrieval results, and recall the relevant domain knowledge text units and experience knowledge text units.

[0047] In the semantic retrieval stage, a vector embedding model is used to generate a question vector representation from the graph-side retrieval results as the vector retrieval input. The semantic similarity between the vector retrieval input and the text unit vector representation matched from the semantic knowledge base is calculated using cosine similarity. The specific formula is:

[0048] ,

[0049] in, is the search result on the graph side, is the text unit to be matched, and They are respectively the vector representation of the conversion of the search results on the graph side and the text unit vector representation of the conversion of the text unit to be matched. and is their modulus length.

[0050] Furthermore, the semantic retrieval results are reordered and the results are prioritized according to relevance, ensuring that the most relevant results are at the front. Optionally, the score threshold can be set to 0.5, which is used to set the similarity threshold for text unit screening, and only recall text fragments that exceed the set score. In a specific example, the weight ratio of the domain knowledge base to the experience knowledge base can be set to 7:3 by default, that is, the domain knowledge text units with the top 7 relevance in the domain knowledge base and the experience knowledge text units with the top 3 similarity in the experience knowledge base are recalled. Furthermore, the present invention can also dynamically adjust the knowledge weight according to the instruction intent analysis. For example, when a standard query type problem is identified, the weight ratio is adjusted to 9:1; the weight ratio of scenario decision type problems is adjusted to 5:5.

[0051] S16. Fusing the entity information retrieval results, the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results to form a fused retrieval result.

[0052] S17. Input the fused search results into the large language model to generate the final search results. Specifically, the final fused search results are input into the large language model for logical verification and evidence chain integration to generate a comprehensive, accurate, and logically robust answer. The generated answer can also include a reference to the specification number, associated case ID, and reasoning path map to support decision-making in underground space operation and maintenance scenarios.

[0053] The knowledge-driven underground space information retrieval method provided by the embodiment of the present invention combines the advantages of underground space entity knowledge graph, semantic knowledge base, and domain knowledge graph. By constructing an information model driven by both normative guidelines and empirical knowledge, and designing a dual-channel retrieval framework of semantic knowledge base and knowledge graph, it achieves a balance between accuracy and interpretability of retrieval results, improves the depth and accuracy of knowledge reasoning of large language models, and provides comprehensive data knowledge support and decision support for underground space information retrieval and operation and maintenance.

[0054] In the embodiment of the present invention, a semantic knowledge base and a domain knowledge graph are pre-built, and a semantic-graph dual-channel retrieval and recall mode is used to realize information retrieval. Through the joint retrieval enhancement of semantic retrieval and graph retrieval, combined with the dual drive of semantic rules and expert experience, a balance between accuracy and interpretability of retrieval results is achieved. Accuracy depends on the precise matching of the knowledge graph and the generalized semantic matching of the semantic knowledge base. The two complement each other to reduce missed detections; graph retrieval provides a structured relationship path, and semantic retrieval supplements the domain norms and historical case experience to form an explainable and traceable chain of evidence. Among them, the graph-side retrieval and recall include graph-side local retrieval of the domain knowledge graph and graph-side global retrieval.

[0055] Specifically, the steps of graph-side local search are: based on the identified key entity information, use the graph query statement to locate the relevant knowledge entities and text blocks in the domain knowledge graph, and then traverse all relevant entities and associations along the graph relationship chain to obtain the graph-side local search results. Furthermore, after obtaining the graph-side local search results, the retrieved results are sorted according to the confidence level, and a specified number (such as top 3-10) of answers with higher confidence levels are selected as the optimal graph-side local search results based on the sorting results. The graph local search confidence level is 10. The calculation formula is:

[0056] ,

[0057] Graph local retrieval confidence Indicates the confidence level of the direct association between the entity and the query, decaying based on the number of hops the entity has in the graph. 0.8 is the decay factor, controlling the rate at which confidence decreases; larger values ​​result in slower decay. n is the number of hops required to reach the target entity from the query origin. The confidence is reduced as the number of hops increases, but non-zero values ​​are retained.

[0058] Specifically, the global retrieval steps on the graph side are: using the community detection algorithm, matching a series of community summaries from the graph as context data from the domain knowledge graph according to the key entity information, calling the large language model to generate multiple answers based on the context data and generate an availability score for each answer, ranging from 0 to 1. The higher the value, the stronger the community reliability. The generated answers are sorted according to the availability score, and a specified number of answers with higher availability scores are selected as the global retrieval results on the graph side according to the sorting results.

[0059] In the embodiment of the present invention, the retrieval result fusion includes integrating the retrieval results obtained from each channel to form a complete knowledge set. The retrieval results specifically include:

[0060] Entity information retrieval results provide relevant feature information of entity objects;

[0061] Local search results on the graph side provide specific knowledge points and related relationships;

[0062] Global search results on the graph side provide a field overview and knowledge clustering;

[0063] Semantic search results provide traceable document evidence.

[0064] The specific implementation method includes establishing a comprehensive scoring model and a dynamic weight adjustment mechanism. The comprehensive scoring model for result ranking based on attention weights evaluates the semantic relevance, knowledge credibility, and domain adaptability of multi-source evidence through quantitative graph and semantic evaluation, and comprehensively ranks and integrates search results. Based on the dynamic weight adjustment mechanism, channel weights are dynamically adjusted according to instruction intent.

[0065] Specifically, in step S16, the entity information retrieval results, the graph-side local retrieval results, the graph-side global retrieval results, and the semantic-side retrieval results are fused to form a fused retrieval result, including:

[0066] A comprehensive scoring model based on attention weights is used to comprehensively score the local retrieval results on the atlas side, the global retrieval results on the atlas side, and the retrieval results on the semantic side. The comprehensive scoring model is as follows:

[0067] ,

[0068] in, is the comprehensive score of the semantic side retrieval results, is the comprehensive score of the local search results on the graph side, is the comprehensive score of the global search results on the graph side, The cosine similarity of the semantic search results ranges from 0 to 1. A higher value indicates a stronger similarity between the search results and the input search elements. The confidence level of the local search results on the graph side ranges from 0 to 1. A higher value indicates that the entity in the local search result is more directly related to the query. It is the usability score of the global search results on the graph side, ranging from 0 to 1. The higher the score, the stronger the relevance of the community summary to the input information. 、 、 These are the weights corresponding to the semantic retrieval channel, the graph-side local retrieval channel, and the graph-side global retrieval channel, respectively. For general problems, the default values ​​can be set to 0.5 for semantics, 0.3 for graph-local, and 0.2 for graph-global. The semantic retrieval channel represents the retrieval process of obtaining semantic retrieval results based on the semantic knowledge base, the graph-side local retrieval channel represents the retrieval process of obtaining graph-side local retrieval results based on the domain knowledge graph, and the graph-side global retrieval channel represents the retrieval process of obtaining graph-side global retrieval results based on the domain knowledge graph.

[0069] When splicing and fusing the entity information retrieval results with the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results, the splicing order of the three retrieval results is determined based on the comprehensive score ranking results of the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results.

[0070] For example, for the instruction "What does pipeline maintenance specification include?", local retrieval of the graph retrieves directly related entities / relationships A, such as "pipeline-pipeline point," with a confidence score of 0.64. Global retrieval of the graph retrieves the pipeline network community summary B, with a score of 0.7. Semantic knowledge base rules retrieve text fragment C, with a similarity of 0.9, and empirical knowledge fragment D, with a similarity of 0.6. The combined scores are: A - 0.64 * 0.3 = 0.192; B - 0.7 * 0.2 = 0.14; C - 0.9 * 0.5 = 0.45; D - 0.6 * 0.5 = 0.3, and the retrieval results are ranked as: C > D > A > B.

[0071] Specifically, the present invention pre-establishes a dynamic weight adjustment mechanism that dynamically adjusts channel weights based on instruction intent analysis. For example, for standard query questions, the semantic channel is weighted more heavily, adjusting the weights to 0.7 for semantics, 0.2 for local graph, and 0.1 for global graph. For scenario-based decision questions, the graph channel is weighted more heavily, adjusting the weights to 0.4 for semantics, 0.3 for local graph, and 0.3 for global graph.

[0072] In this embodiment, an underground space information model is designed to provide comprehensive data knowledge support and decision-making support for underground space operation and maintenance. The underground space information model includes an underground space entity knowledge graph (i.e., an underground space entity model), a semantic knowledge base (domain knowledge base, experience knowledge base), and a domain knowledge graph.

[0073] In this embodiment, the underground space entities in the urban underground space field, i.e. pipelines, buildings, etc., can be abstracted and modeled, and attributes such as time characteristics, spatial characteristics, attribute characteristics, and association relationships can be given to them to obtain spatiotemporal geographic objects with spatiotemporal semantic information. All spatiotemporal geographic objects can be used to construct an underground space ontology model with spatiotemporal semantic information, and the knowledge graph can be used to represent it to obtain an underground space entity knowledge graph, so as to realize graph-based entity retrieval and query, thereby realizing efficient management and application of underground space entities. The construction steps of the underground space entity knowledge graph specifically include:

[0074] 1. Abstract and model the underground space entities in the underground space field to obtain the spatiotemporal geographic object class framework.

[0075] 2. Create the class structure in the underground space ontology model based on the spatiotemporal geographic object class framework, and instantiate the class structure in the underground space ontology model into entities in the underground space ontology model. Each class structure represents a class of spatiotemporal geographic object models with common characteristics.

[0076] Specifically, based on the spatiotemporal geographic object class framework, we define the class structure within the ontology, such as pipeline class, point class, and building class. Each class structure represents a type of spatiotemporal geographic object and contains its common characteristic attributes and relationships. The specific class structure is then instantiated as an entity within the ontology model.

[0077] For example, a specific pipeline class structure is instantiated as an instance of the "Pipeline Class" and assigned a unique identifier, such as PID-2007567. This identifier can be generated using a hash function or a UUID (Universally Unique Identifier) ​​algorithm to ensure uniqueness and scalability. For example, the UUID generation formula is: UUID = Construction Timestamp ⊕ Node ID ⊕ Random Number.

[0078] 3. Assign characteristic attributes and association relationships to each entity in the underground space ontology model to abstract the entity into a spatiotemporal geographic object with spatiotemporal semantic information. The characteristic attributes include time characteristics, spatial characteristics and attribute characteristics.

[0079] Specifically, parametric modeling technology can be used to give underground space entities four-dimensional characteristics: space, time, attributes, and relationships, and abstract them into spatiotemporal geographic objects with spatiotemporal semantic information. Spatiotemporal geographic objects are the mapping objects of underground space entities in computers.

[0080] 4. Based on the class structure, characteristic attributes and association relationships in the underground space ontology model, the spatiotemporal geographic objects are converted into a node-edge-node triple structure and semantically stored through the knowledge graph to obtain the underground space entity knowledge graph. The nodes in the underground space entity knowledge graph represent the entities and characteristic attributes in the underground space ontology model, and the edges in the underground space entity knowledge graph represent the association relationships between spatiotemporal geographic objects and the attribute relationships between the entities corresponding to the spatiotemporal geographic objects themselves and the corresponding attribute features. Figure 2 , Figure 2 Schematic diagram of pipeline entity knowledge graph in an embodiment of the present invention. Furthermore, the nodes in the underground space entity knowledge graph can also represent entity words of other concept classes in the underground space ontology model, such as pipeline class.

[0081] Specifically, based on the class structure, characteristic attributes, and relationship definitions in the underground space ontology model, spatiotemporal geographic objects are converted into a node-edge-node triple structure, and semantic storage and query are achieved through a graph database. The specific method is as follows:

[0082] (1) Node mapping: Entity instances (such as pipeline PID-2007567) and concepts (such as pipeline class and time feature) in the ontology model are represented as nodes in the knowledge graph, and the nodes are connected by edges.

[0083] (2) Relationship mapping: The relationship between spatiotemporal geographic objects is expressed through the edges of the knowledge graph, and types and attributes are assigned to them. Specifically:

[0084] An example of representing time association: For pipeline A that undergoes spatiotemporal changes, create a node relationship of "pipeline A previous state - maintenance event - pipeline A later state". The edge type is the corresponding event, and the edge attribute can define information such as the operation unit of the maintenance time.

[0085] An example of expressing spatial associations: describing topological relationships through edge types. For example, the edge type of "pipeline A - connection - pipe point B" is connection, and edge attributes can define information such as interface type and connection distance.

[0086] An example of expressing semantic associations: logical relationships are expressed through edge types. For example, the edge type of "Pipeline A—Belongs to—Pressure Pipeline" is Belongs to.

[0087] In this embodiment, before constructing the semantic knowledge base and experiential knowledge graph, vertical domain data for underground space is collected. The collected multi-source, heterogeneous domain data is then processed into structured text data that can be semantically parsed. The vertical domain data for underground space consists of both underground space domain knowledge and experiential knowledge.

[0088] Underground space domain knowledge mainly includes the following data types:

[0089] (1) Regulations: These regulations include those related to the development, use, and management of underground space, and are an important basis for the operation and maintenance of underground space, such as the Measures for the Planning and Management of the Development and Utilization of Urban Underground Space.

[0090] (2) Standards and specifications: Through the National Standard Information Public Service Platform, relevant national and local standards are collected by searching with keywords such as "underground space", "geotechnical engineering", and "pipeline". These standards and specifications provide a technical basis for the planning, design, construction, and operation and maintenance of underground space, such as "Data Requirements for Urban Underground Space" (GB / T 42987-2023), "Technical Specifications for Three-Dimensional Modeling of Urban Underground Space" (GB / T 41447-2022), "Data Specifications for Urban Underground Space" (DB 3401 / T 230-2021), and "Technical Standards for Geotechnical Engineering Information Modeling in Tianjin" (DB / T 29-292-2021).

[0091] (3) Books and documents: Professional books and documents in the field of underground space were collected by accessing Wanfang Database and other professional databases. These materials usually include theoretical research and practical cases of underground space.

[0092] Experiential knowledge mainly includes the following data types:

[0093] (1) Operation and maintenance case history records: by collecting accident case reports; collecting underground space sensor monitoring logs, operation and maintenance records, etc.

[0094] (2) Decision-making experience collection: Design domain-adaptive templates to describe specific underground space operation and maintenance scenarios, guide experts to think about how to identify problems, retrieve specifications, match cases, calculate parameters, and generate solutions, so as to transform implicit expert experience into structured textual expressions that can be retrieved by large models.

[0095] In a specific example, we provide an example of experience collection in tunnel water leakage treatment scenarios:

[0096] "Inside a power compartment of a city's underground utility corridor, a temperature sensor detected an abnormally high temperature in a local cable. Further analysis of the root cause and development of a treatment plan are required. Please explain the following in order:

[0097] [Problem Identification] What factors might be causing the current abnormal phenomenon? (Such as short circuit or overload, cable aging, loose connectors, poor heat dissipation, etc.) Is it necessary to integrate other monitoring data (such as local humidity and current fluctuations) to assist in the diagnosis?

[0098]

Regulation Search

[0099] [Parameter calculation] involves the parameters and formulas that need to be calculated:

[0100]

Experience Collection

[0101] [Solution Generation] Please list the steps to be taken according to priority. This includes which indicators need to be monitored after the solution is implemented to verify its effectiveness.

[0102] Specifically, the steps for building a semantic knowledge base include:

[0103] 1. Convert underground space domain knowledge data and experience knowledge data into structured text units. Underground space domain knowledge data includes normative text data, diagram data in standard documents, and term-definition pairs related to the underground space field. Experience knowledge data includes historical operation and maintenance case data and decision-making experience data.

[0104] Furthermore, converting the underground space domain knowledge data and the experience knowledge data into structured text units includes:

[0105] For normative text data, key identifiers and chapter content structures are extracted from the text data. A pre-set multi-level regular expression parsing template is then used to segment the chapter content structure into regular structured text units. A hierarchical construction approach is employed for normative text data. The first layer extracts key identifiers such as the name title, publication date, implementation date, standard number, and the region to which the local standard belongs. The second layer dynamically loads a multi-level regular expression parsing template, segmenting the text units based on the "chapter-clause-sub-item" boundary to ensure that each unit contains complete semantic constraints.

[0106] For chart data in standard documents, multimodal large-scale model-based OCR technology is used to parse chart data into structured text data to generate chart structured text units. For chart data in national and local standards, such as the "Underground Space Data Classification Table" and the "Underground Space Data Metadata Basic Content Table," multimodal large-scale model-based OCR technology is used to parse them into structured text data, such as "Serial Number: 1, Entity Set: Spatiotemporal Reference, Metadata Item: Data Collection Time, Description: Data Collection Date and Beijing Time, Field Type: Time, Constraint: M (Required)," with each chart or table as the boundary for segmentation.

[0107] For term-definition pairs related to underground space, we segment each term to generate structured text units. We also extract term-definition pairs from national standards, local standards, and professional books to build a terminology database. For example, we extract terms and definitions such as "underground space," "geotechnical engineering," and "structural health monitoring," and segment each term to provide a foundation for subsequent knowledge retrieval and reasoning.

[0108] For historical operation and maintenance case data, we perform structured extraction, constructing a case event chain description in the form of "accident type-cause-phenomenon-disposition." This data is then linked to the urban underground space geocoding system to obtain the case's spatiotemporal tags. The combined data of the case event chain description and spatiotemporal tags for each case is then text-segmented to generate structured case text units. We also perform structured extraction on historical operation and maintenance case data, such as historical accident reports and operation and maintenance records, to construct an event chain in the form of "accident type-cause-phenomenon-disposition." This data is then linked to the urban underground space geocoding system, retaining spatiotemporal tags such as project location and occurrence time, and segmenting each individual case. For example, for a tunnel water seepage incident, we construct an event chain of "water seepage incident-geological structural change-roof water seepage-grouting reinforcement," and record the project location and time of the incident.

[0109] For decision-making experience data, we perform structured extraction, establish a logical chain description in the form of "problem identification - standard retrieval - case matching - parameter calculation - solution generation", and segment the logical chain description of each scenario to generate experience structured text units. The collected decision-making experience data is decomposed into a five-stage logical chain of "problem identification → standard retrieval → case matching → parameter calculation → solution generation", converted into structured data, and segmented according to each individual scenario.

[0110] 2. A vector embedding model is used to vectorize text units derived from underground space domain knowledge data to generate a domain knowledge base. A vector embedding model is also used to vectorize text units derived from empirical knowledge data to generate an empirical knowledge base. The domain knowledge base and the empirical knowledge base constitute a semantic knowledge base. The vector embedding model can generate high-quality text vector representations and supports semantic similarity calculations. The present invention uses a vector embedding model to process text units derived from underground space domain knowledge data and empirical knowledge data, storing them in a knowledge base in a vector-embedded manner, thereby constructing a domain knowledge base and an empirical knowledge base.

[0111] In this embodiment, the steps of constructing the domain knowledge graph specifically include:

[0112] 1. Extract entity-relationship-entity triples from structured text units to obtain subgraph segments containing entity lists and relationship lists, and construct the initial knowledge graph based on each subgraph segment.

[0113] The present invention optimizes the existing graph retrieval generation technology. First, the prompt word template is converted into Chinese and the underground space domain information is added. In a specific example, the prompt word template is as follows:

[0114] "Target Activities

[0115] You are an intelligent assistant that helps human analysts analyze claims about certain entities in text.

[0116] Target

[0117] Given a text, entity specification, and claim description that may be relevant to this activity, extract all entities that match the entity specification and all claims for these entities.

[0118] step

[0119] Extract Entities: Extracts all named entities that match a predefined entity specification. The entity specification can be a list of entity names or a list of entity types.

[0120] Extract claims: For each entity identified in step 1, extract all claims related to that entity. Claims need to conform to the specified claim description, and the entity should be the subject of the claim. For each claim, extract the following information:

[0121] Subject: The entity name of the subject of the declaration, capitalized. The subject entity is the entity that performs the behavior described in the declaration and must be one of the named entities identified in step 1.

[0122] Object: The name of the entity that the declaration is about, in uppercase. The object entity is the entity that reports on, processes, or is affected by the actions described in the declaration. If the object entity is unknown, use NONE.

[0123] Claim Type: The overall category of the claim, capitalized. This naming convention should be reproducible so that similar claims share the same claim type.

[0124] The status of the claim: TRUE (confirmed), FALSE (falsified), or SUSPECTED (pending verification).

[0125] Claim Description: Describe in detail the reasoning for the claim, including all relevant evidence and references.

[0126] Claim Date: The time range (start date, end date) of the claim. Date format is ISO-8601. If the claim was made on a single date rather than a date range, set the start and end dates to the same date. Returns NONE if the date is unknown.

[0127] Claim Source Text: List all original text citations relevant to the claim.

[0128] Format each declaration as:

[0129] (<subject entity>{tuple_delimiter}<object entity>{tuple_delimiter}<declaration type>{tuple_delimiter}<declaration status>{tuple_delimiter}<start date>{tuple_delimiter}<end date>{tuple_delimiter}<declaration description>{tuple_delimiter}<declaration source>)

[0130] Return output: Returns a list of all declarations in Chinese, using {record_delimiter} as the list separator.

[0131] Completion flag: Output {completion_delimiter} after completion.

[0132] Specific examples:

[0133] Physical Specifications: Underground Space Facilities

[0134] Statement description: Functional description related to underground space facilities

[0135] Text: Trunk integrated pipeline corridor. This is a comprehensive pipeline corridor primarily used to accommodate pipelines for urban trunk projects. Its primary function is to provide transportation services for urban municipal stations, meeting the normal flow of personnel and possessing complete ancillary facilities.

[0136] Output:

[0137] (Trunk utility tunnel is mainly used to accommodate urban trunk engineering pipelines, provide transportation services for urban municipal stations, meet the normal passage of personnel, and have complete ancillary facilities. It is mainly used to accommodate urban trunk engineering pipelines. Its main function is to provide transportation services for urban municipal stations, meet the normal passage of personnel, and have complete ancillary facilities.)

[0138] {completion_delimiter}

[0139] Actual data

[0140] Generate your response using the following input.

[0141] Entity Specifications: {entity_specs}

[0142] Claim description: {claim_description}

[0143] Text: {input_text}

[0144] Output:

[0145] Use graph retrieval generation technology to index and build code, call the general language model, and extract a large number of entity-relationship-entity triples from the structured text units according to the above prompt word template, such as "comprehensive pipeline corridor-including-trunk comprehensive pipeline corridor", "trunk comprehensive pipeline corridor-transportation service-urban municipal station", etc., to obtain sub-graph segments containing entity lists and relationship lists, and use these extracted information to build the initial knowledge graph, such as Figure 3 As shown, Figure 3A schematic diagram of the underground space domain knowledge graph in one embodiment of the present invention is shown.

[0146] 2. Use a community detection algorithm to classify the initial knowledge graph to obtain multiple target communities, and generate a community summary for each target community to obtain a domain knowledge graph.

[0147] The present invention uses community detection technology to classify and divide the initial knowledge graph according to topics.

[0148] First define multiple topics:

[0149] (1) Ontology information: entity definitions and classification tables in regulations / standards;

[0150] (2) Disease information: disease type, cause, and treatment measures in the accident report (sub-graph: accident → cause → phenomenon → treatment)

[0151] (3) Operation monitoring: parameters of operation and maintenance records, maintenance and disposal records, etc.;

[0152] (4) Design information: design parameters in the standard;

[0153] (5) Project implementation: construction technology in the case.

[0154] Among them, a community detection algorithm is used to classify the initial knowledge graph to obtain multiple target communities, and a community summary is generated for each target community, further including: using the community detection algorithm to cluster similar triplets in the semantic space in the initial knowledge graph to obtain underlying communities, clustering the underlying communities according to multiple preset underground space themes to obtain multiple target communities corresponding to different underground space themes; reading the entity, relationship and attribute data of the subgraph corresponding to each cluster, and passing them to a universal large language model, using a preset structured prompt word template to guide the universal large language model to identify entity relationships and restore subgraph logic, and generate a community summary of the target community according to the entity relationships and subgraph logic; using a vector embedding model to convert the community summary of each target community into a vector representation for realizing the retrieval of the domain knowledge graph.

[0155] This method uses a community detection algorithm to cluster similar triplets in the semantic space. It then reads the entity, relationship, and attribute data of each clustered subgraph and passes it to a general large language model. Using a structured prompt word template, it gradually guides the general large language model to identify entity relationships, restore subgraph logic, and generate community summaries. From the underlying communities derived from clustering to the five defined thematic communities, increasingly comprehensive community summaries are generated. A vector embedding model is used to convert each community summary into a vector representation.

[0156] In a specific example, the prompt word template in this step is as follows:

[0157] "## Role

[0158] You are a knowledge graph analysis expert. You need to generate concise and objective summaries based on the entities and relationships of the subgraph, highlighting the core topics and key information.

[0159] ## Input **[knowledge graph subgraph]**

[0160] ## Skill

[0161] 1. Entity Recognition

[0162] - Accurately identify entity names, descriptions, and attributes, ignoring empty descriptions or meaningless entities.

[0163] - If the entity name is a document / directory ID, directly associate it with its semantics (e.g., "GB / T 42987-2023§5.2" is considered as a normative clause).

[0164] ### 2. Relationship Analysis

[0165] - Accurately identify relationship information, including source entity name, relationship name, target entity name, and relationship description.

[0166] ###3. Restore subgraph logic

[0167] - Restore the graph structure based on the identified entities and relationships

[0168] ###4. Generate summary

[0169] - Determine the topic of the subgraph based on the frequently appearing entities and their core relationships. Optional topics include "ontology information", "disease information", "operation monitoring", "design information", and "project implementation".

[0170] - Use accurate, professional, and concise language to summarize the information presented in the sub-graph.

[0171] The knowledge-driven underground space information retrieval system proposed in the present invention has at least the following beneficial effects:

[0172] 1. Constructing an urban underground space information model: This approach maps underground space knowledge with geographic information, establishes a physical model, and forms a knowledge graph and semantic knowledge base for knowledge reasoning and semantic understanding, enabling effective management and utilization of underground space information.

[0173] 2. Multimodal knowledge extraction method: Use text recognition technology based on a multimodal large model to parse chart information into structured data, realizing multimodal data fusion in the underground space field.

[0174] 3. Decision logic explicitness framework: Design a technical solution for collecting and encoding expert experience, extract implicit decision logic and convert it into a structured thinking chain to achieve the accumulation of implicit experience knowledge.

[0175] 4. Hybrid Retrieval Enhancement Architecture: This architecture builds a queryable and reasonable underground space knowledge graph based on improved graph retrieval enhancement technology, addressing the inefficiency of manually constructed knowledge graphs. A multi-level semantic recall module is designed to enhance the combined retrieval of semantic and graph retrieval, enabling hierarchical retrieval of underground space knowledge and improving the comprehensiveness and accuracy of knowledge retrieval. By combining semantic rules with expert experience, this architecture achieves a balance between accuracy and interpretability in retrieval results, enhancing the depth of knowledge reasoning.

[0176] For simplicity of description, the method embodiments are described as a series of actions. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because certain steps can be performed in other orders or simultaneously according to the embodiments of the present invention. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0177] Another embodiment of the present invention further provides a knowledge-driven underground space information retrieval system, which includes a functional module for implementing any of the above-mentioned knowledge-driven underground space information retrieval methods. Figure 4 The structure block diagram of the knowledge-driven underground space information retrieval system according to another embodiment of the present invention is schematically shown. Figure 4 The knowledge-driven underground space information retrieval system of this embodiment specifically includes an instruction parsing module 401, an entity information retrieval module 402, a graph local retrieval module 403, a graph global retrieval module 404, a semantic retrieval module 405, a fusion module 406, and a retrieval result generation unit 407, wherein:

[0178] The instruction parsing module 401 is used to call the large language model to identify and parse the key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects;

[0179] The entity information retrieval module 402 is configured to retrieve the spatiotemporal geographic object in a preset underground space entity knowledge graph, and traverse the characteristic attributes, association relationships, and relationships with other entity nodes of the spatiotemporal geographic object based on the knowledge graph relationships of the spatiotemporal geographic object to obtain entity information retrieval results. The underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space field through a knowledge graph.

[0180] The graph local retrieval module 403 is used to retrieve relevant entities from the preset domain knowledge graph based on the key entity information, and traverse all relevant entities and associations along the graph relationship chain to obtain graph-side local retrieval results; the domain knowledge graph includes multiple communities, each of which is formed by graph clustering of entity-relationship-entity triples extracted from the preset underground space domain knowledge data and empirical knowledge data;

[0181] A graph global search module 404 is configured to match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data based on the key entity information, and to call a large language model to generate graph-side global search results based on the context data;

[0182] Semantic search module 405, used to use the local search results on the graph side and / or the global search results on the graph side as vector search input, match the corresponding text unit vector representation from the preset semantic knowledge base, and obtain the semantic search results;

[0183] Fusion module 406, for fusing entity information retrieval results, graph-side local retrieval results, graph-side global retrieval results, and semantic-side retrieval results to form a fused retrieval result;

[0184] The retrieval result generating unit 407 is used to input the fused retrieval result into the large language model to generate the final retrieval result.

[0185] In the embodiment of the present invention, the atlas local search module 403 is further used to sort the retrieved results according to the confidence after obtaining the atlas side local search results, and select a specified number of answers with higher confidence as the atlas side local search results according to the sorting results. The atlas local search confidence The calculation formula is:

[0186] ,

[0187] Among them, 0.8 is the decay base, which is used to control the confidence decrease rate, and n is the number of hops, which indicates the relationship level that needs to be crossed from the query starting point to the target entity.

[0188] In an embodiment of the present invention, the graph global retrieval module 404 is specifically used to call the large language model to generate multiple answers based on the context data and generate an availability score for each answer; sort the generated answers according to the availability scores, and select a specified number of answers with higher availability scores as graph-side global retrieval results based on the sorting results.

[0189] In the embodiment of the present invention, the fusion module 406 is specifically used to use an attention weight-based result comprehensive scoring model to comprehensively score the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results; when splicing and fusing the entity information retrieval results with the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results, the splicing order of the three retrieval results is determined based on the comprehensive score ranking results of the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results. The result comprehensive scoring model is as follows:

[0190] ,

[0191] in, is the comprehensive score of the semantic side retrieval results, is the comprehensive score of the local search results on the graph side, is the comprehensive score of the global search results on the graph side, The cosine similarity of the semantic search results ranges from 0 to 1. A higher value indicates a stronger similarity between the search results and the input search elements. The confidence level of the local search results on the graph side ranges from 0 to 1. A higher value indicates that the entity in the local search result is more directly related to the query. It is the usability score of the global search results on the graph side, ranging from 0 to 1. The higher the score, the stronger the relevance of the community summary to the input information. 、 、 These are the weights corresponding to the semantic retrieval channel, the graph-side local retrieval channel, and the graph-side global retrieval channel, respectively. For general problems, the default values ​​can be set to 0.5 for semantics, 0.3 for graph-local, and 0.2 for graph-global. The semantic retrieval channel represents the retrieval process of obtaining semantic retrieval results based on the semantic knowledge base, the graph-side local retrieval channel represents the retrieval process of obtaining graph-side local retrieval results based on the domain knowledge graph, and the graph-side global retrieval channel represents the retrieval process of obtaining graph-side global retrieval results based on the domain knowledge graph.

[0192] In the embodiment of the present invention, the fusion module 406 is further used to establish a dynamic weight adjustment mechanism to dynamically adjust the channel weight according to the instruction intention.

[0193] As for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0194] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0195] In addition, another embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor; when the computer program is executed by the processor, the steps of the knowledge-driven underground space information retrieval method described above are implemented.

[0196] In addition, another embodiment of the present invention further provides a computer program product, which stores a computer program. When the computer program is executed by a processor, the steps of the knowledge-driven underground space information retrieval method described above are implemented.

[0197] Those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not other features, the combination of features from different embodiments is intended to be within the scope of the present invention and to form different embodiments. For example, any of the claimed embodiments may be used in any combination.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A knowledge-driven underground space information retrieval method, characterized in that: The method comprises: The large language model is used to identify and parse the key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects; The spatiotemporal geographic object is retrieved in a preset underground space entity knowledge graph, and the characteristic attributes, association relationships, and relationships with other entity nodes of the spatiotemporal geographic object are traversed based on the knowledge graph relationship of the spatiotemporal geographic object to obtain entity information retrieval results; the underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space field through a knowledge graph; Relevant entities are retrieved from a preset domain knowledge graph based on the key entity information, and all relevant entities and associations are traversed along the graph relationship chain to obtain local search results on the graph side; the domain knowledge graph includes multiple communities, each of which is formed by graph clustering of entity-relationship-entity triples extracted from preset underground space domain knowledge data and empirical knowledge data; Match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data according to the key entity information, and call a large language model to generate a graph-side global search result based on the context data; Use the local search results on the graph side and / or the global search results on the graph side as vector search input, match the corresponding text unit vector representation from the preset semantic knowledge base, and obtain the semantic search results; Fusing entity information retrieval results, graph-side local retrieval results, graph-side global retrieval results, and semantic-side retrieval results to form a fused retrieval result; The fused retrieval results are input into the large language model to generate the final retrieval results.

2. The method according to claim 1, characterized in that After obtaining the local search results on the atlas side, the method further includes: Sort the retrieved results by confidence, and select a specified number of answers with higher confidence as the local search results on the atlas side. The confidence of the local search on the atlas is The calculation formula is: ， Among them, 0.8 is the decay base, which is used to control the confidence decrease rate, and n is the number of hops, which indicates the relationship level that needs to be crossed from the query starting point to the target entity.

3. The method according to claim 1, characterized in that The calling of the large language model to generate a graph-side global search result based on the context data includes: Invoking a large language model to generate a plurality of answers based on the context data and generating a usability score for each answer; The generated answers are sorted according to their availability scores, and a specified number of answers with higher availability scores are selected as global search results on the graph side based on the sorting results.

4. The method according to claim 1, wherein The entity information retrieval results, the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results are integrated to form a fusion retrieval result, including: A comprehensive scoring model based on attention weights is used to comprehensively score the local retrieval results on the atlas side, the global retrieval results on the atlas side, and the retrieval results on the semantic side. The comprehensive scoring model is as follows: ， in, is the comprehensive score of the semantic side retrieval results, is the comprehensive score of the local search results on the graph side, is the comprehensive score of the global search results on the graph side, is the cosine similarity of the semantic side retrieval results; is the confidence of the local search results on the atlas side; is the availability score of the global search results on the graph side, 、 、 are the weights corresponding to the semantic side retrieval channel, the graph side local retrieval channel, and the graph side global retrieval channel, respectively. The semantic side retrieval channel represents the retrieval process of obtaining semantic side retrieval results based on the semantic knowledge base, the graph side local retrieval channel represents the retrieval process of obtaining graph side local retrieval results based on the domain knowledge graph, and the graph side global retrieval channel represents the retrieval process of obtaining graph side global retrieval results based on the domain knowledge graph; When splicing and fusing the entity information retrieval results with the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results, the splicing order of the three retrieval results is determined based on the comprehensive score ranking results of the local retrieval results on the graph side, the global retrieval results on the graph side, and the semantic side retrieval results.

5. The method according to claim 4, characterized in that The method further comprises: Establish a dynamic weight adjustment mechanism to dynamically adjust channel weights according to instruction intent.

6. The method according to claim 1, characterized in that The steps of constructing the underground space entity knowledge graph include: Abstract and model the underground space entities in the underground space domain to obtain the spatiotemporal geographic object class framework; According to the spatiotemporal geographic object class framework, the class structure in the underground space ontology model is created, and the class structure in the underground space ontology model is instantiated as an entity in the underground space ontology model. Each class structure represents a class of spatiotemporal geographic objects with common characteristics. Each entity in the underground space ontology model is given characteristic attributes and association relationships to abstract the entity into a spatiotemporal geographic object with spatiotemporal semantic information. The characteristic attributes include time characteristics, spatial characteristics and attribute characteristics. Based on the class structure, characteristic attributes and association relationships in the underground space ontology model, the spatiotemporal geographic objects are converted into a node-edge-node triple structure and semantically stored through the knowledge graph to obtain the underground space entity knowledge graph. The nodes in the underground space entity knowledge graph represent the entities and characteristic attributes in the underground space ontology model, and the edges in the underground space entity knowledge graph represent the association relationships between spatiotemporal geographic objects and the attribute relationships between the entities corresponding to the spatiotemporal geographic objects themselves and the corresponding attribute features.

7. The method according to claim 1, characterized in that The steps of constructing the semantic knowledge base include: Convert underground space domain knowledge data and experience knowledge data into structured text units respectively; A vector embedding model is used to vectorize the text units obtained from the underground space domain knowledge data to generate a domain knowledge base, and a vector embedding model is used to vectorize the text units obtained from the experience knowledge data to generate an experience knowledge base. The domain knowledge base and the experience knowledge base constitute a semantic knowledge base.

8. The method according to claim 7, characterized in that Underground space domain knowledge data includes normative text data, diagram data in standard documents, and term-definition pairs involved in the underground space field. Experience knowledge data includes historical operation and maintenance case data and decision-making experience data. The converting of the underground space domain knowledge data and the experience knowledge data into structured text units includes: For standard text data, the standard key identification information and chapter content structure are extracted from the text data, and the chapter content structure is segmented into text using a preset multi-level regular parsing template to generate regular structured text units; For chart data in standard documents, OCR technology based on multimodal large models is used to parse chart data into structured text data to generate chart structured text units; For the term-definition pairs involved in the field of underground space, the text segmentation is performed according to each term to generate term-structured text units; For historical operation and maintenance case data, we perform structured extraction on the data and establish a case event chain description in the form of "accident type-cause-phenomenon-disposition". We then link it to the urban underground space geocoding system to obtain the case's spatiotemporal labels. We then perform text segmentation on the combined data of the case event chain description and spatiotemporal labels for each case to generate case structured text units. For decision-making experience data, structured extraction is performed on the decision-making experience data, and a logical chain description in the form of "problem identification-standard retrieval-case matching-parameter calculation-solution generation" is established. The logical chain description of each scenario is divided into text to generate experience structured text units.

9. The method according to claim 7, characterized in that The steps of constructing the domain knowledge graph include: Extract entity-relationship-entity triples from structured text units to obtain subgraph segments containing entity lists and relationship lists, and construct an initial knowledge graph based on each subgraph segment; A community detection algorithm is used to classify the initial knowledge graph to obtain multiple target communities, and a community summary is generated for each target community to obtain a domain knowledge graph.

10. The method according to claim 9, characterized in that A community detection algorithm is used to classify the initial knowledge graph to obtain multiple target communities, and a community summary is generated for each target community, including: Using a community detection algorithm, similar triplets in the semantic space in the initial knowledge graph are clustered to obtain underlying communities. The underlying communities are clustered according to multiple preset underground space themes to obtain multiple target communities corresponding to different underground space themes. Read the entity, relationship, and attribute data of each cluster's corresponding subgraph and pass it to the general language model. Use the preset structured prompt word template to guide the general language model to identify entity relationships and restore subgraph logic. Then, generate a community summary of the target community based on the entity relationships and subgraph logic. A vector embedding model is used to convert the community summary of each target community into a vector representation.

11. A knowledge-driven underground space information retrieval system, characterized in that: The system comprises: The instruction parsing module is used to call the large language model to identify and parse the key entity information of the user's underground space operation and maintenance scenario instructions to understand the instruction intent of the current instruction. The key entity information includes spatiotemporal geographic objects and entity words other than the spatiotemporal geographic objects; An entity information retrieval module is used to retrieve the spatiotemporal geographic object in a preset underground space entity knowledge graph, and traverse the characteristic attributes, association relationships, and relationships with other entity nodes of the spatiotemporal geographic object based on the knowledge graph relationship of the spatiotemporal geographic object to obtain entity information retrieval results; the underground space entity knowledge graph is an entity model formed by semantically storing spatiotemporal geographic objects with spatiotemporal semantic information in the underground space field through the knowledge graph; A graph local retrieval module is used to retrieve relevant entities from a preset domain knowledge graph based on the key entity information, and traverse all relevant entities and associations along the graph relationship chain to obtain graph-side local retrieval results. The domain knowledge graph includes multiple communities, each of which is formed by graph clustering of entity-relationship-entity triples extracted from preset underground space domain knowledge data and empirical knowledge data. A graph global search module is used to match relevant community summaries from the community summaries of each community in the domain knowledge graph as context data based on the key entity information, and to call a large language model to generate graph-side global search results based on the context data; The semantic retrieval module is used to use the local retrieval results on the graph side and / or the global retrieval results on the graph side as vector retrieval inputs, match the corresponding text unit vector representations from the preset semantic knowledge base, and obtain the semantic retrieval results; A fusion module is used to fuse entity information retrieval results, local retrieval results on the graph side, global retrieval results on the graph side, and retrieval results on the semantic side to form a fused retrieval result; The retrieval result generating unit is used to input the fused retrieval results into the large language model to generate the final retrieval results.

12. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 10 are implemented.

13. A computer program product, characterized in that The computer program product stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images