A Knowledge Graph-Based Method and System for Asset Management Association Query and Visualization
By constructing an asset knowledge graph to detect abnormal assets and using the Tongyi Thousand Questions model to parse queries, combined with user exploration paths and attention scores, the passive response and lack of guidance issues of existing asset management systems are solved, realizing proactive discovery and intelligent guidance for asset management queries.
Patent Information
- Application Number
- CN202611518132.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing asset management systems cannot proactively detect abnormal assets and lack support for progressive exploration with multi-hop associations. Users need to repeatedly perform manual queries without receiving intelligent guidance.
By constructing an asset knowledge graph, detecting abnormal asset entity nodes, using the Tongyi Qianwen model to parse natural language queries, recording user exploration paths and calculating attention scores, and proactively pushing potential risk points and exploration tips.
It enables the system to proactively discover abnormal assets, lowers the threshold for querying graph databases, provides intelligent guidance for the next step of exploration, and forms a dual-track parallel exploration mechanism driven by user interests and system risks.
Smart Images

Figure CN123115265A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of knowledge graph and data visualization technology, specifically relating to a knowledge graph-based method and system for asset management association query and visualization. Background Technology
[0002] As enterprises deepen their digital transformation, the number of asset entities involved in asset management continues to grow. Asset types encompass diverse categories such as hardware equipment, software systems, contracts and documents, personnel and organizations, suppliers, and projects. These assets exhibit complex, multi-dimensional relationships involving ownership, association, dependence, and constraints. In daily enterprise operations, management tasks such as asset inventory, risk tracing, and impact analysis heavily rely on a comprehensive understanding and rapid retrieval of these asset relationships.
[0003] Currently, some solutions have incorporated knowledge graph technology for asset management. For example, Chinese patent CN121277911A discloses a metadata query method based on knowledge graphs, using the Tongyi Qianwen model to assist in the query; Chinese patent CN121547292A discloses a method and system for intelligently determining unknown threats based on multimodal knowledge graphs; other solutions involve data quality monitoring methods and digital asset change impact analysis methods based on knowledge graphs. However, all of the above-mentioned existing technologies have obvious limitations: the query mode is passive response, the system can only return the results of the user's explicit query, and cannot actively discover and push potential abnormal asset entity nodes such as isolated nodes, key hub nodes, and abnormally dense nodes; at the same time, users need to try to query multi-hop associations hop by hop themselves, the system cannot record the user's exploration path and actively prompt the next direction of exploration at the appropriate time, lacking the ability of progressive exploration and intelligent guidance.
[0004] Therefore, the existing asset management association query technology has the following problems: 1. The system cannot proactively discover and push abnormal assets, making it difficult for users to discover risks that have not been noticed; 2. It lacks support for progressive exploration of multi-hop associations, requiring users to repeatedly perform manual queries; 3. It cannot proactively guide the next query direction based on the user's exploration behavior. Summary of the Invention
[0005] (a) Technical problems to be solved
[0006] In response to the problems in related technologies, this invention provides a knowledge graph-based method and system for asset management association query and visualization to overcome the aforementioned technical problems existing in the existing related technologies.
[0007] (II) Technical Solution
[0008] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:
[0009] S1. Obtain multi-source asset management data, and extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source asset management data through knowledge extraction;
[0010] An asset knowledge graph is constructed based on the asset entities, the asset entity attributes, and the relationships between the asset entities.
[0011] S2. Detect abnormal asset entity nodes by using the graph structure features of each asset entity node in the asset knowledge graph to obtain a list of abnormal asset entities;
[0012] Push the list of abnormal asset entities to the user interface;
[0013] S3. Semantic parsing of the natural language query statement input by the user is performed using the Tongyi Qianwen model to obtain the graph query language statement;
[0014] The initial query results are obtained by performing queries in the asset knowledge graph using the graph query language statement, and then displayed through a visual interface.
[0015] S4. Based on the user's selection operation of asset entity nodes in the visualization interface, obtain the current focus node;
[0016] Based on the current focus node, a query is performed in the asset knowledge graph to obtain the next-hop related asset entity;
[0017] Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path;
[0018] S5. Based on the access frequency and dwell time of each asset entity node in the historical exploration path, calculate the user attention score.
[0019] The asset entity nodes whose user attention scores are higher than the attention threshold and have undisplayed associated paths are marked as nodes that can be further explored;
[0020] The visual interface outputs exploration prompts for the nodes that can be further explored.
[0021] This invention forms a complete closed loop through four dimensions: proactive discovery, natural language interaction, gradual exploration, and intelligent guidance. From the system's proactive problem discovery, user natural language query, and gradual exploration through clicked nodes to the system's intelligent guidance for the next step, it solves the technical problems of existing asset management query systems that are passive in response, unable to discover hidden connections, and lack exploration guidance.
[0022] Preferably, step S1 includes the following steps:
[0023] S11. Obtain multi-source asset management data, including asset master data and asset relationship data;
[0024] S12. Extract asset entities and asset entity attributes from the asset master data; extract relationships between asset entities from the asset association data;
[0025] S13. Construct an asset knowledge graph based on the asset entity, the asset entity attributes, and the relationships between the asset entities;
[0026] The above steps extract asset entities, attributes, and relationships from asset master data and asset association data, and construct an asset knowledge graph using a graph database, providing a structured data foundation for subsequent association queries, anomaly detection, and incremental exploration;
[0027] Preferably, step S2 includes the following steps:
[0028] S21. Calculate the graph structure features of each asset entity node in the asset knowledge graph, wherein the graph structure features include degree centrality, betweenness centrality and clustering coefficient;
[0029] S22. Detect abnormal asset entity nodes through the graph structure features to obtain the list of abnormal asset entities;
[0030] S23. Push the list of abnormal asset entities to the user interface;
[0031] The above steps actively detect isolated nodes, key hub nodes, and abnormally dense nodes by calculating the degree centrality, betweenness centrality, and clustering coefficient of each node, and push the list of abnormal assets to the user, thus realizing a shift from passive query to proactive recommendation.
[0032] Preferably, the detection of abnormal asset entity nodes through the graph structure features in step S22 includes the following steps:
[0033] S221. Determine asset entity nodes with degree centrality below the low degree centrality threshold as isolated nodes;
[0034] S222. Asset entity nodes with betweenness centrality higher than the high betweenness centrality threshold are identified as critical hub nodes.
[0035] S223. Asset entity nodes with clustering coefficients higher than the high clustering threshold are classified as abnormally dense nodes;
[0036] S224. The isolated node, the critical hub node, and the abnormally dense node are designated as the abnormal asset entity node.
[0037] The above steps use threshold determination of three types of graph structure features to make three types of abnormal assets explicit: isolated nodes, key hub nodes, and abnormally dense nodes. The three types of anomalies correspond to the data omission risk, key node failure risk, and over-coupling risk in the asset network, respectively, so that users can proactively know the potential risk points without initiating a query.
[0038] Preferably, step S3 includes the following steps:
[0039] S31. Receive the natural language query statement input by the user;
[0040] S32. Semantic parsing of the natural language query statement is performed using the Tongyi Qianwen model to obtain the graph query language statement; the training dataset of the Tongyi Qianwen model comes from the historical query logs of the enterprise asset management system.
[0041] S33. Use the graph query language statement to perform a query in the asset knowledge graph to obtain the initial query results;
[0042] S34. Display the initial query results through a visual interface;
[0043] The above steps automatically convert natural language query statements into graph query language statements through the Tongyi Qianwen model, which lowers the threshold for using graph database queries and enables business personnel without graph query language skills to directly use the system to perform asset association queries.
[0044] Preferably, step S32 includes the following steps:
[0045] S321. Generate semantic parsing prompt word templates based on the graph pattern information of the asset knowledge graph;
[0046] S322. Fill the query statement placeholder position in the semantic parsing prompt word template with the natural language query statement to obtain the complete prompt word;
[0047] S323. Input the complete prompt word into the general meaning thousand questions model, and perform semantic parsing through the general meaning thousand questions model to obtain the initial graph query language statement;
[0048] S324. Perform syntax validation on the initial graph query language statement. If the syntax of the initial graph query language statement is correct, use the initial graph query language statement as the graph query language statement.
[0049] S325. When the initial graph query language statement has a syntax error, the syntax error information and the initial graph query language statement are re-inputted into the general meaning question model.
[0050] The graph query language statement is obtained by modifying the general meaning thousand questions model; the modified graph query language statement is then used as the graph query language statement.
[0051] The above steps guide the Tongyi Qianwen model to perform semantic parsing through prompt word templates, and improve the accuracy and reliability of query statement generation through a closed-loop mechanism of syntax verification and error correction, avoiding query failures caused by syntax errors.
[0052] Preferably, step S4 includes the following steps:
[0053] S41. Receive the user's selection operation on the asset entity node in the visualization interface, and obtain the selected asset entity node.
[0054] S42. Set the selected asset entity node as the current focus node;
[0055] S43. Query the asset knowledge graph through the current focus node to obtain the next-hop related asset entity; the next-hop related asset entity does not include asset entity nodes that have appeared in the historical exploration path;
[0056] S44. Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path;
[0057] The above steps unfold the map step by step by the user clicking on nodes, realizing progressive exploration, while recording the user's historical exploration path, providing a data foundation for subsequent intelligent guidance;
[0058] Preferably, step S5 includes the following steps:
[0059] S51. Extract the access frequency and dwell time of each asset entity node in the user's historical exploration path;
[0060] S52. Calculate the user attention score based on the access frequency and the dwell time;
[0061] S53. Based on the user attention score and the display status of the associated path, select nodes that can be further explored from the asset entity nodes; the user attention score of the nodes that can be further explored is higher than the attention threshold and there are undisplayed associated paths.
[0062] S54. Output exploration prompts for the nodes that can be further explored through the visualization interface;
[0063] The above steps analyze users' historical exploration behavior, calculate the attention score of each node, and proactively prompt users that there are related directions that can be explored further, thus realizing an interactive upgrade from user-driven to system-guided intelligent guidance.
[0064] Preferably, the calculation of user attention score for each asset entity node in step S52 based on the access frequency and dwell time includes the following steps:
[0065] S521. Obtain the access frequency of each asset entity node in the historical exploration path;
[0066] S522. Obtain the dwell time of each asset entity node during each visit in the historical exploration path, and calculate the average dwell time of each asset entity node;
[0067] S523. The user attention score of each asset entity node is obtained by adding the product of the access frequency and the frequency weight and the product of the average dwell time and the duration weight.
[0068] The above steps calculate the user attention score by weighting the frequency of visits and the average dwell time, quantifying user behavior data into comparable numerical indicators, providing a quantitative basis for screening nodes that can be further explored, and making intelligent guidance based on objective data rather than fixed rules.
[0069] Preferably, a knowledge graph-based asset management association query and visualization system is used to implement the above-mentioned knowledge graph-based asset management association query and visualization method. The system includes a multi-source data acquisition module, a knowledge graph construction module, an anomaly detection module, a query parsing module, an interactive exploration module, and an intelligent guidance module.
[0070] The knowledge graph construction module is used to extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source data of asset management through knowledge extraction, and to construct an asset knowledge graph based on the asset entities, asset entity attributes, and relationships between asset entities.
[0071] An anomaly detection module is used to detect abnormal asset entity nodes through the graph structure features of each asset entity node in the asset knowledge graph, obtain a list of abnormal asset entities, and push the list of abnormal asset entities to the user interface.
[0072] The query parsing module is used to perform semantic parsing on the natural language query statement input by the user through the Tongyi Qianwen model to obtain a graph query language statement. The graph query language statement is then used to query the asset knowledge graph to obtain initial query results, which are then displayed through a visual interface.
[0073] The interactive exploration module is used to obtain the current focus node based on the user's selection operation of asset entity nodes in the visualization interface, to query the asset knowledge graph based on the current focus node, to obtain the next-hop related asset entity, to add the next-hop related asset entity to the visualization interface, and to record the user's historical exploration path.
[0074] The intelligent guidance module is used to calculate the user attention score based on the access frequency and dwell time of each asset entity node in the historical exploration path, mark the asset entity nodes with the user attention score higher than a preset threshold and with undisplayed related paths as nodes that can continue to be explored, and output the exploration prompt information of the nodes that can continue to be explored through the visualization interface.
[0075] (III) Beneficial Effects
[0076] The present invention has the following beneficial effects:
[0077] This invention proactively detects three types of abnormal assets—isolated nodes, key hub nodes, and abnormally dense nodes—by calculating the degree centrality, betweenness centrality, and clustering coefficient of asset entity nodes in an asset knowledge graph and pushes them to the user interface. This upgrades passive querying to proactive discovery, allowing users to know potential risk points without initiating a query.
[0078] This invention uses the Tongyi Qianwen model to parse user-input natural language queries into graph query language statements, enabling business personnel without graph query language skills to directly use the system for asset association queries, thus lowering the barrier to entry for graph database queries. At the same time, syntax validation and error correction improve the accuracy and reliability of query statement generation.
[0079] This invention records the historical exploration path of a user as they click on nodes to expand related entities. Based on the access frequency and dwell time of each node in the path, the system calculates the user's attention score. Nodes with attention scores higher than a threshold and with undisplayed related paths are marked as nodes that can be explored further, and exploration prompts are actively output. This enables the system to proactively guide the next exploration direction at appropriate times.
[0080] This invention organically combines proactive anomaly detection, natural language query parsing, progressive exploration, and intelligent guidance, forming a complete closed loop from the system proactively discovering problems to the user's natural language query, then to the progressive exploration of clicked nodes, and finally the system intelligently guiding the next step. This solves the problems of existing asset management query systems being passive in response, unable to discover hidden connections, and lacking exploration guidance.
[0081] This invention achieves a dual-track exploration guidance mechanism that combines user interest-driven and system risk-driven approaches through the synergistic effect between anomaly detection results and intelligent guidance mechanisms. This mechanism enables intelligent guidance to not only calculate attention scores based on users' historical exploration behavior but also integrate structural risk information proactively discovered by the system, thus combining user interest-driven and system risk-driven exploration guidance to form a dual-track parallel exploration guidance mechanism.
[0082] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0083] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.
[0084] Figure 1 This is a flowchart illustrating the knowledge graph-based asset management association query and visualization method of the present invention.
[0085] Figure 2 This is a hierarchical diagram of the overall system architecture of the knowledge graph-based asset management association query and visualization system of the present invention.
[0086] Figure 3 This is a schematic diagram of the modules of the knowledge graph-based asset management association query and visualization system of the present invention; Detailed Implementation
[0087] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.
[0088] Example 1:
[0089] Please see Figure 1 This invention discloses a knowledge graph-based method for asset management association query and visualization, comprising the following steps:
[0090] S1. Obtain multi-source asset management data, and extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source asset management data through knowledge extraction;
[0091] An asset knowledge graph is constructed based on the asset entities, the asset entity attributes, and the relationships between the asset entities.
[0092] S1 includes the following steps:
[0093] S11. Obtain multi-source asset management data, including asset master data and asset relationship data;
[0094] Specifically, in embodiment S11, the system obtains asset master data through the application programming interface of the enterprise asset management system. The asset master data includes fields such as asset code, asset name, asset type, department, purchase date, and original asset value. The system obtains asset association data through the asset association database. The asset association data includes the ownership relationship, dependency relationship, constraint relationship, and validity period of the association between assets. The system stores the above data in association with the asset code as the primary key to obtain multi-source asset management data. The multi-source asset management data obtained by the system comes from the enterprise's internal asset management system. The data is the enterprise's own asset information and does not involve personal information or sensitive data. Its collection and use comply with the enterprise's internal data management regulations.
[0095] The mapping rules from fields to asset entity attributes are as follows: the asset code field in the asset master data is mapped to the unique identifier of the asset entity; the asset name field is mapped to the asset entity name attribute; the asset type field is mapped to the asset entity type attribute; the department field is mapped to the asset entity's department attribute; the purchase date field is mapped to the asset entity's purchase time attribute; and the original asset value field is mapped to the asset entity's initial value attribute. Fields not listed in the above mapping rules are not extracted as asset entity attributes. The entity alignment rules are as follows: cross-data source entity alignment is performed using the asset code as the unique primary key of the asset entity. Records with the same asset code appearing in different data tables are merged into the same asset entity node. For data records lacking an asset code, matching is performed using the edit distance similarity of the asset name, with a similarity threshold set at 0.85. Records with a similarity higher than this threshold are considered as lacking an asset code. For the same asset entity, the similarity threshold is set based on the following: Statistical analysis of historical data records in the enterprise asset management system shows that the normalized edit distance similarity of different descriptions of the same asset entity due to non-standard naming or input errors is usually between 0.85 and 0.95; while the name similarity between different asset entities is usually below 0.6. Setting the similarity threshold to 0.85 can achieve a recall rate of over 95% while controlling the false match rate below 5%, achieving a balance between recall and precision. The judgment condition for relation extraction is: directly reading the predefined relation type field from the asset association data. The relation types include attribution, dependency, constraint, and reference relationships. The source asset code and target asset code in each association record are mapped to the relation start point and relation end point, respectively, and the relation type field is mapped to the relation type.
[0096] S12. Extract asset entities and asset entity attributes from the asset master data; extract relationships between asset entities from the asset association data;
[0097] Specifically, in embodiment S12, the system adopts a rule-based knowledge extraction method: identifying fields such as asset code, asset name, asset type, and department from the asset master data as asset entity attributes, and using the asset code as the unique identifier of the asset entity; identifying ownership, dependency, constraint, and reference relationships from the asset association data as relationships between asset entities;
[0098] The mapping rules from fields to asset entity attributes are as follows: the asset code field in the asset master data is mapped to the unique identifier of the asset entity; the asset name field is mapped to the asset entity name attribute; the asset type field is mapped to the asset entity type attribute; the department field is mapped to the asset entity's department attribute; the purchase date field is mapped to the asset entity's purchase time attribute; and the original asset value field is mapped to the asset entity's initial value attribute. Fields not listed in the above mapping rules are not extracted as asset entity attributes. The entity alignment rules are as follows: cross-data source entity alignment is performed using the asset code as the unique primary key of the asset entity. Records with the same asset code appearing in different data tables are merged into the same asset entity node. For data records lacking an asset code, matching is performed using the edit distance similarity of the asset name. The similarity threshold is set to 0.85; records with a similarity higher than this threshold are considered to point to the same asset entity. One asset entity; the judgment condition for relation extraction is: directly read the predefined relation type field from the asset association data. The relation type includes attribution, dependency, constraint and reference relationships. The source asset code and target asset code in each association record are mapped to the relation start point and relation end point, respectively. The relation type field is mapped to the relation type. The similarity threshold is set based on: the system performs statistical analysis on historical data records in the enterprise asset management system. Different descriptions of the same asset entity due to non-standard naming or input errors usually have a normalized edit distance similarity between 0.85 and 0.95. The name similarity between different asset entities is usually below 0.6. Setting the similarity threshold to 0.85 can achieve a recall rate of over 95% while controlling the false match rate below 5%, thus achieving a balance between recall and precision.
[0099] S13. Construct an asset knowledge graph based on the asset entity, the asset entity attributes, and the relationships between the asset entities;
[0100] Specifically, in embodiment S13, the system uses asset entities as nodes, asset entity attributes as node attributes, and relationships between asset entities as edges, and uses a graph database to store and construct an asset knowledge graph. The graph database contains three basic elements: asset entity nodes, asset entity attributes, and relationships between asset entities. The graph schema of the knowledge graph defines asset entity types, asset entity relationship types, and attribute fields contained in various asset entities. The graph database supports the retrieval and traversal of nodes and relationships using a graph query language.
[0101] The above steps extract asset entities, attributes, and relationships from multi-source asset management data and construct an asset knowledge graph using a graph database, providing a structured data foundation for subsequent relational queries and anomaly detection.
[0102] S2. Detect abnormal asset entity nodes by using the graph structure features of each asset entity node in the asset knowledge graph to obtain a list of abnormal asset entities;
[0103] Push the list of abnormal asset entities to the user interface;
[0104] S2 includes the following steps:
[0105] S21. Calculate the graph structure features of each asset entity node in the asset knowledge graph, wherein the graph structure features include degree centrality, betweenness centrality and clustering coefficient;
[0106] Specifically, in embodiment S21, the system calculates three types of graph structure features for each asset entity node in the asset knowledge graph: degree centrality is normalized degree centrality, calculated as CD(v)=kv / (N-1), where kv is the number of neighboring nodes of node v, and N is the total number of nodes in the asset knowledge graph; the normalized degree centrality compresses the absolute degree centrality of the node to the [0,1] interval, making the degree centrality in asset knowledge graphs of different sizes comparable; betweenness centrality is the proportion of the number of times the node appears in all shortest paths, reflecting the degree of the asset's bridging role in the asset network; the clustering coefficient is the proportion of the number of actual edges between the node's neighboring nodes to the number of fully connected edges, reflecting the degree of clustering among the neighboring nodes around the asset; the system generates a three-dimensional graph structure feature vector for each asset entity node that includes degree centrality, betweenness centrality, and clustering coefficient.
[0107] Degree centrality uses absolute degree centrality, and the calculation formula is as follows: ,in Let V be the number of neighboring nodes of node v, i.e., the number of asset entity nodes directly connected to node v; the betweenness centrality adopts the normalized betweenness centrality, and the calculation formula is as follows: Where N is the total number of nodes, The total number of shortest paths from node s to node t. The number of shortest paths passing through node v; the clustering coefficient uses the local clustering coefficient, calculated using the following formula: ,in, This represents the number of edges that actually exist between the neighboring nodes of node v;
[0108] The normalization factor for betweenness centrality Related to the total number of nodes N, the clustering coefficient and The actual topology derived from the asset knowledge graph;
[0109] S22. Detect abnormal asset entity nodes through the graph structure features to obtain the list of abnormal asset entities;
[0110] The step S22, which involves detecting abnormal asset entity nodes using the graph structure features, includes the following steps:
[0111] S221. Determine asset entity nodes with degree centrality below the low degree centrality threshold as isolated nodes;
[0112] The low degree centrality threshold is set based on the following: the degree centrality value corresponding to the 25th percentile of the degree centrality of all asset entity nodes is taken as the low degree centrality threshold. The reason for choosing the lower quartile is that in scale-free networks, the degree centrality distribution is highly right-skewed, and the mean is greatly affected by a few highly connected nodes, which cannot represent the normal connection level of most nodes. The lower quartile represents the low quantile of the node degree centrality distribution, and nodes below this value can be judged as isolated nodes. When the number of nodes in the asset knowledge graph is less than 100, the statistical significance of the lower quartile is insufficient, and 50% of the median is used as the low degree centrality threshold. Once the number of nodes accumulates to more than 100, the lower quartile strategy is switched.
[0113] S222. Asset entity nodes with betweenness centrality higher than the high betweenness centrality threshold are identified as critical hub nodes.
[0114] The high betweenness centrality threshold is set as follows: All asset entity nodes are sorted from highest to lowest normalized betweenness centrality values, and the normalized betweenness centrality value in the top 5% of the sorted results is taken as the high betweenness centrality threshold. The selection of the top 5% quantile is based on the fact that the normalized betweenness centrality range is [0,1]. In asset management networks, key nodes that act as bridges are often concentrated in the extreme head region of the betweenness centrality distribution, and the top 5% quantile can effectively cover the nodes in this head region. When the number of nodes in the asset knowledge graph is less than 100, a strategy of using the mean plus twice the standard deviation is used as an alternative. ,in The mean of the normalized betweenness centrality of all nodes. The standard deviation is used; once the number of nodes accumulates to more than 100, the strategy is switched to the top 5% quantile.
[0115] S223. Asset entity nodes with clustering coefficients higher than the high clustering threshold are classified as abnormally dense nodes;
[0116] The high clustering threshold is set based on the following: Local clustering coefficients of all asset entity nodes are sorted from highest to lowest, and the local clustering coefficients in the top 5% of the sorted results are used as the high clustering threshold. The selection of the top 5% quantile is based on the fact that local clustering coefficients range from [0,1]. Abnormally dense nodes correspond to a very small proportion of nodes at the head of the clustering coefficient distribution, and the top 5% quantile effectively covers extremely dense areas where clustering coefficients are significantly higher than ordinary nodes. When the number of nodes in the asset knowledge graph is less than 100, a strategy of using the mean plus twice the standard deviation is used as an alternative. ,in The mean of the local clustering coefficients for all nodes. The standard deviation is used; once the number of nodes accumulates to more than 100, the strategy is switched to the top 5% quantile.
[0117] S224. Designate the isolated node, the critical hub node, and the abnormally dense node as the abnormal asset entity node.
[0118] The above steps use three thresholds—degree centrality, betweenness centrality, and clustering coefficient—to make the three types of abnormal asset entity nodes—isolated nodes, key hub nodes, and abnormally dense nodes—explicit. These three types of anomalies correspond to the risks of data omission, key node failure, and over-coupling in the asset network, respectively, allowing users to proactively identify potential risk points without initiating a query.
[0119] S23. Push the list of abnormal asset entities to the user interface;
[0120] Specifically, in embodiment S23, the system organizes the detected abnormal asset entity nodes into an abnormal asset entity list, which includes the asset name, asset type, abnormality type and corresponding graph structure feature value of each abnormal node; the system actively pushes this list to the user through a visual interface without requiring the user to initiate a query;
[0121] The above steps analyze the topological characteristics of each node in the knowledge graph, proactively discover three types of abnormal asset entity nodes: isolated nodes, key hub nodes, and abnormally dense nodes, thus realizing a shift from passive querying to proactive recommendation.
[0122] S3. Semantic parsing of the natural language query statement input by the user is performed using the Tongyi Qianwen model to obtain the graph query language statement;
[0123] The initial query results are obtained by performing queries in the asset knowledge graph using the graph query language statement, and then displayed through a visual interface.
[0124] S3 includes the following steps:
[0125] S31. Receive the natural language query statement input by the user;
[0126] Specifically, in embodiment S31, the system receives a natural language query statement input by the user through the input box of the visual interface. The natural language query statement is an asset association query request described in Chinese natural language form.
[0127] S32. Semantic parsing of the natural language query statement is performed using the Tongyi Qianwen model to obtain the graph query language statement; the training dataset of the Tongyi Qianwen model comes from the historical query logs of the enterprise asset management system, and the historical query logs contain the natural language query statement input by the user and the corresponding standard graph query language statement annotations.
[0128] S32 includes the following steps:
[0129] S321. Generate semantic parsing prompt word templates based on the graph pattern information of the asset knowledge graph;
[0130] Specifically, in embodiment S321, the system obtains graph pattern information of the asset knowledge graph. The graph pattern information includes a list of asset entity types, a list of asset entity relationship types, and graph query language syntax specifications. The system constructs a semantic parsing prompt word template based on the graph pattern information. The prompt word template includes query statement placeholders and graph pattern description information, which are used to guide the Tongyi Qianwen model to generate graph query language statements that conform to the graph pattern.
[0131] The semantic parsing prompt template consists of three parts: the first part is a graph pattern description, used to inform the Tongyi Qianwen model of the graph pattern information of the asset knowledge graph, which includes a list of asset entity types and a list of asset entity relationship types; the second part is an output format constraint, used to constrain the Tongyi Qianwen model to only output graph query language statements and not output any explanatory text; the third part is a query statement placeholder, used to fill in the natural language query statement input by the user in S322; the asset entity types in the graph pattern information include equipment, systems, contracts, personnel, suppliers, and projects, and the asset entity relationship types include ownership relationships, dependency relationships, operation and maintenance relationships, constraint relationships, and reference relationships;
[0132] The full text of the semantic parsing prompt word template used in this embodiment is as follows:
[0133] "You are a graph database query language generation assistant. The current graph schema of the asset knowledge graph is as follows: asset entity types include equipment, systems, contracts, personnel, suppliers, and projects; asset entity relationship types include ownership, dependency, operation and maintenance, constraint, and reference relationships. The basic syntax rules of the graph query language are: MATCH (n:{entity type}) WHERE n.{attribute name}={attribute value} RETURN n; where {entity type}, {attribute name}, and {attribute value} are variables extracted from the user's natural language query statement, which you fill in by combining the graph schema information, and must be enclosed in backticks for Chinese tags and Chinese attribute names. Please convert the following natural language query statement into a graph query language statement that conforms to the above graph schema, outputting only the graph query language statement without any explanatory text. The natural language query statement is: {user-input natural language query statement}."
[0134] S322. Fill the query statement placeholder position in the semantic parsing prompt word template with the natural language query statement to obtain the complete prompt word;
[0135] S323. Input the complete prompt word into the general meaning thousand questions model, and perform semantic parsing through the general meaning thousand questions model to obtain the initial graph query language statement;
[0136] Specifically, in embodiment S323, the system calls the Tongyi Qianwen open-source model for semantic parsing. The Tongyi Qianwen model is a Tongyi Qianwen model based on the Transformer decoder architecture. Its network structure consists of multiple layers of stacked Transformer decoders. Each decoder layer includes a masked multi-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by residual connections and layer normalization operations. The Tongyi Qianwen model used in this embodiment is Qwen2.5-7B version, with 70 model parameters, 28 Transformer decoder layers, a hidden layer dimension of 3584, 28 heads of the multi-head self-attention mechanism, a fully connected layer dimension of 18944, a context window length of 32768 tags, and a vocabulary size of 152064. This model has been pre-trained on a general corpus for autoregressive language modeling and has general language understanding and generation capabilities.
[0137] The training data comes from the historical query logs of the enterprise asset management system. The system extracts user query records from the past 12 months from the historical query logs. Each historical query record contains the natural language query statement entered by the user, the query timestamp, and the query result. According to the query intent of each natural language query statement, it is labeled as the corresponding standard graph query language statement to form the initial labeled data. The initial labeled dataset contains 10,000 samples, covering common query intent types such as asset query, relationship query, path query, and condition query.
[0138] The training data is constructed as follows: The system extracts user query records from the past 12 months of historical query logs in the enterprise asset management system. Each historical query record includes the user's input natural language query statement, query timestamp, and query results. The system performs intent annotation on each natural language query statement, mapping it to the corresponding graph query language statement to form initial annotation data. Intent annotation adopts a semi-automated annotation process: First, the pre-trained graph query language generation model generates candidate graph query language statements for each natural language query statement. Then, the annotators manually review and correct the candidate statements, including the accuracy of node labels, the correctness of relation types, the completeness of query conditions, and the rationality of returned fields. After correction, a standard graph query language statement is formed as a supervision label. After annotation, the system performs data augmentation on the annotated data. Specifically, the asset entity name, asset type name, and relation type name in the natural language query statement are replaced with synonyms, expanding one annotated sample into multiple training samples with similar semantics but different expressions to enhance the model's generalization ability. The data-augmented samples are divided into training, validation, and test sets in an 8:1:1 ratio. After training, the system inputs the complete prompt words from the user into the Tongyi Qianwen model. The complete prompt words are composed of system prompt words and user queries concatenated in sequence. The system prompt words contain graph pattern descriptions and output format constraints, while the user queries are natural language query statements input by the user. After receiving the complete prompt words, the Tongyi Qianwen model performs autoregressive decoding based on the model parameters, generating each word in the output sequence sequentially. The generated word sequence does not contain delimiters or special markers and directly outputs the complete graph query language statement in plain text form. This statement is the initial graph query language statement.
[0139] The system performs data augmentation on the labeled dataset. Specifically, it replaces key entities such as asset names, asset types, and relationship types in natural language queries with synonyms, expanding one labeled sample into multiple training samples with similar semantics but different expressions to enhance the model's generalization ability. The augmented samples are divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for updating model parameters, the validation set is used for hyperparameter tuning and overfitting monitoring, and the test set is used for the final performance evaluation after model training.
[0140] The system employs the LoRA method for training, injecting a low-rank matrix into the linear transformation layer of the base model. During training, only the parameters of the low-rank matrix are updated, while the original parameters of the base model remain unchanged, thus reducing training costs while achieving domain adaptation. For frequently occurring professional terms in asset management queries, such as asset codes, full lifecycle status, and organizational hierarchy, the system constructs a dedicated vocabulary for the asset management domain. This vocabulary expansion includes asset codes, full lifecycle status enumeration values, and organizational hierarchy names. Randomly initialized embedding vectors are assigned to newly added terms in the model's word embedding layer, preventing the model from segmenting asset management terms into sub-words and preserving complete semantics. Simultaneously, the system performs specific enhancements on queries involving asset codes and status values during fine-tuning training. The training set is augmented with a higher proportion of query samples containing asset code exact matching and state value conditional filtering, enabling the model to learn the unique expression patterns of exact matching and conditional queries in asset management scenarios. Furthermore, during the training sample construction process, the system implements mask prediction training for asset codes: randomly replacing some characters of the asset codes in the training samples with mask markers, requiring the model to predict and complete the complete asset codes when generating graph query language statements, thereby enhancing the model's understanding of asset code formats. During fine-tuning, the system inputs natural language query statements from the training set into the Tongyi Qianwen model, and the model outputs corresponding graph query language statements. Using manually annotated standard graph query language statements as supervision signals, the cross-entropy loss function is employed to calculate the difference between the model output and the supervision signal. The formula for the cross-entropy loss function is: Where N is the length of the output sequence, and |V| is the vocabulary size. Let v be the value of the one-hot encoding corresponding to the real word at position t. The model predicts the probability that the t-th position is the v-th word in the vocabulary; the optimizer is AdamW with a learning rate of 2e-4, a weight decay of 0.01, 10 training epochs, and a batch size of 16; during training, the system calculates the gradient through backpropagation, updates only the parameters of the LoRA low-rank matrix, and keeps the basic model parameters unchanged.
[0141] After each training epoch, the system calculates the accuracy of the generated graph query language statements on the validation set. The accuracy is the proportion of samples whose generated graph query language statements are completely identical to the standard graph query language statements out of the total number of samples in the validation set. When the validation set accuracy reaches 90% or higher, the current model parameters are finally evaluated on the test set. If the test set accuracy also reaches 90% or higher, the current model parameters are used as the final model parameters for deployment. If the validation set accuracy does not improve for three consecutive training epochs, training is terminated early to avoid overfitting. The model parameters corresponding to the epoch with the highest validation set accuracy are evaluated on the test set. If the test set accuracy reaches 90% or higher, deployment is complete; if the test set accuracy does not reach 90%, the fine-tuning parameters are adjusted and training is re-executed. If the validation set accuracy still does not reach 90% after the maximum number of training epochs (10 epochs), the system adjusts the fine-tuning parameters and re-executes training.
[0142] When deploying the Tongyi Qianwen model after fine-tuning training, the system overlays a prefix tree constraint decoding mechanism during the model output decoding stage. This prefix tree constraint decoding mechanism works as follows: the system pre-extracts all asset entity type names, all asset entity relationship type names, and all unique asset codes from the graph schema of the asset knowledge graph, constructing a prefix tree data structure. In each decoding step of the Tongyi Qianwen model's autoregressive generation of graph query language statements, the system matches the currently generated lexical sequence with the prefix tree to determine the next lexical candidate set allowed for the current decoding position. If the currently generated lexical sequence matches a node in the prefix tree, then all sub-nodes of that node are... The corresponding word is used as a candidate word to be generated. The probability of words that are not in the candidate set in the original output probability distribution of the general knowledge graph model is set to zero to obtain the masked probability distribution. The output word of the current decoding step is generated by sampling from the masked probability distribution. This prefix tree constraint decoding mechanism ensures that the asset entity type, relation type and asset code involved in the graph query language statement generated by the model are all from the actual types and entities in the current asset knowledge graph. From the decoding stage, it fundamentally eliminates the illusion problem that may occur in the general large model in the field of asset management, such as the graph entity not existing or the relation type mismatch. It does not need to rely on the memory ability of the fine-tuned model to avoid the illusion output.
[0143] After the Tongyi Qianwen model is trained, the system will input the complete prompt words into the Tongyi Qianwen model for semantic parsing to obtain the initial graph query language statement;
[0144] S324. Perform syntax validation on the initial graph query language statement. If the syntax of the initial graph query language statement is correct, use the initial graph query language statement as the graph query language statement.
[0145] Specifically, in embodiment S324, the system performs grammatical verification on the initial graph query language statement. The grammatical verification includes keyword spelling check, grammatical structure integrity check, and graph pattern matching check. The specific rules for keyword spelling check are as follows: the initial graph query language statement is split into a sequence of lexical units, and each lexical unit is matched against the graph query language keyword list. Lexical units not in the keyword list are marked as unknown keywords and the error type is recorded as a spelling error. The specific rules for grammatical structure integrity check are as follows: the initial graph query language statement is checked to see if it contains a complete query clause structure, whether parentheses are matched in pairs, and whether there is an unclosed string or numeric constant at the end of the statement. If any check fails, it is determined to be a grammatical structure error. The specific rules for graph pattern matching check are as follows: the target asset entity type and the asset entity type and asset entity relationship type appearing in the query conditions are extracted from the initial graph query language statement. The extracted types are compared one by one with the asset entity type list and asset entity relationship type list defined in the graph pattern of the asset knowledge graph. If there is an entity type or relationship type that is not defined in the graph pattern, it is determined to be a graph pattern matching error. The error classification criteria are as follows: keyword spelling errors, grammatical structure errors, and graph pattern matching errors are mutually exclusive and are checked in the above order, with the higher priority error type being given priority. When the initial graph query language statement passes all the checks, the system directly uses it as the graph query language statement.
[0146] S325. When the initial graph query language statement has a syntax error, the syntax error information and the initial graph query language statement are re-inputted into the general meaning question model.
[0147] The graph query language statement is obtained by modifying the general meaning thousand questions model; the modified graph query language statement is then used as the graph query language statement.
[0148] Specifically, in embodiment S325, the system sets the maximum number of corrections to 3. When the initial graph query language statement fails the syntax check, the system concatenates the syntax error information with the initial graph query language statement and re-enters the general meaning question model for correction. After the correction is completed, the syntax check is performed again. If the corrected statement passes the check, it is output as the graph query language statement. If the correction count reaches 3 and the statement still fails the check, the system marks the currently generated graph query language statement as pending manual review and terminates the correction process. At the same time, it returns a prompt message to the user interface, informing the user that the current query statement cannot be automatically converted and suggesting that the query expression be adjusted or the administrator be contacted for processing.
[0149] After a correction fails, the system stores the initial graph query language statement, the three corrected statements, the syntax error information corresponding to each correction, and the timestamp in the failure record database. Administrators can view the failure records through the pending review list in the visual interface, manually correct them, and submit the correct graph query language statement. The system uses the manually corrected statement and the original natural language query statement as new labeled samples, and adds them incrementally to the training set periodically for fine-tuning the next round of the general semantic question model, so as to continuously optimize the model's semantic parsing ability. At the same time, the system counts the frequency distribution of various syntax error types in the failure records. When the frequency of a certain error type is significantly high, the error information and correction mode corresponding to that error type are added to the constraint description part of the prompt word template. In subsequent queries, stronger format constraints are imposed on the model output, reducing the probability of the same type of error from the input prompt level.
[0150] The above steps automatically convert natural language query statements into graph query language statements through the Tongyi Qianwen model, which lowers the threshold for using graph database queries and allows users to complete asset association queries without mastering graph query syntax. At the same time, through the closed-loop mechanism of syntax verification and error correction, the accuracy and reliability of query statement generation are improved, and query failures caused by syntax errors are avoided.
[0151] S33. Use the graph query language statement to perform a query in the asset knowledge graph to obtain the initial query results;
[0152] Specifically, in embodiment S33, the system submits the graph query language statement to the graph database for query execution. The graph database parses the graph query language statement and performs node and relationship retrieval and matching in the asset knowledge graph according to its semantics, and returns the matched asset entity nodes and related relationships as the initial query results.
[0153] S34. Display the initial query results through a visual interface;
[0154] Specifically, in embodiment S34, the system displays the initial query results in the form of a visual graph in the user interface. The visual graph is based on the target asset entity node and its one-hop related asset entity node, and is rendered and displayed using a preset force-oriented layout. The user interface supports interactive operations such as zooming, dragging, and node selection.
[0155] The above steps automatically convert users' natural language queries into graph query language statements through the Tongyi Qianwen model, which lowers the threshold for users to use graph query language and enables business personnel without graph query language skills to directly use the system to perform asset association queries.
[0156] S4. Based on the user's selection operation of asset entity nodes in the visualization interface, obtain the current focus node;
[0157] Based on the current focus node, a query is performed in the asset knowledge graph to obtain the next-hop related asset entity;
[0158] Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path;
[0159] S4 includes the following steps:
[0160] S41. Receive the user's selection operation on the asset entity node in the visualization interface, and obtain the selected asset entity node.
[0161] S42. Set the selected asset entity node as the current focus node;
[0162] S43. Query the asset knowledge graph through the current focus node to obtain the next-hop related asset entity; the next-hop related asset entity does not include asset entity nodes that have appeared in the historical exploration path;
[0163] Specifically, in embodiment S43, the system uses the current focus node as the starting point for the query and queries all neighboring nodes directly associated with the current focus node in the asset knowledge graph; the system compares the query results with the user's historical exploration path, removes nodes that have appeared in the historical exploration path, and uses the remaining nodes as the next-hop associated asset entities;
[0164] S44. Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path;
[0165] Specifically, in embodiment S44, the system adds the next-hop associated asset entity as a new node to the visualization interface, and establishes a connection edge between the new node and the current focus node; the system appends the node information and timestamp of this selection operation to the user's historical exploration path record, which includes all asset entity nodes that the user has visited and the order of visits;
[0166] The above steps unfold the map step by step by the user clicking on nodes, realizing progressive exploration, while recording the user's historical exploration path, providing a data foundation for subsequent intelligent guidance;
[0167] S5. Based on the access frequency and dwell time of each asset entity node in the historical exploration path, calculate the user attention score.
[0168] The asset entity nodes whose user attention scores are higher than the attention threshold and have undisplayed associated paths are marked as nodes that can be further explored;
[0169] The visual interface outputs exploration prompts for the nodes that can be further explored.
[0170] S5 includes the following steps:
[0171] S51. Extract the access frequency and dwell time of each asset entity node in the user's historical exploration path;
[0172] Specifically, in embodiment S51, the system extracts the access frequency and dwell time of each asset entity node from the recorded historical exploration path; the access frequency is the number of times the user clicks to view the node in the current query session, and the dwell time is the duration for which the user stays on the related view of the node after it is expanded;
[0173] S52. Calculate the user attention score based on the access frequency and the dwell time;
[0174] The calculation of user attention score for each asset entity node in S52 based on the access frequency and dwell time includes the following steps:
[0175] S521. Obtain the access frequency of each asset entity node in the historical exploration path;
[0176] S522. Obtain the dwell time of each asset entity node during each visit in the historical exploration path, and calculate the average dwell time of each asset entity node;
[0177] S523. The user attention score of each asset entity node is obtained by adding the product of the access frequency and the frequency weight and the product of the average dwell time and the duration weight.
[0178] Specifically, in embodiment S523, the system calculates the user attention score for each asset entity node by adding the product of access frequency and frequency weight and the product of average dwell time and duration weight. The system normalizes the access frequency and average dwell time separately before calculating the user attention score for each asset entity node. The calculation formula is as follows: ,in Let be the access frequency of the i-th asset entity node. Let be the average dwell time of the i-th asset entity node. and These are frequency weight and duration weight, respectively. The basis for setting these weights is as follows: In asset management related query scenarios, the average dwell time of a user on a specific asset entity node reflects the true level of attention more accurately than the access frequency. Access frequency may be inflated due to accidental clicks or rapid browsing, but dwell time directly reflects the user's need for in-depth viewing of the asset's related information. In this embodiment, the duration weight is greater than the frequency weight, i.e. =0.4、 =0.6; The system performs the above calculation on each asset entity node in the historical exploration path to obtain the user attention score of each asset entity node;
[0179] S53. Based on the user attention score and the display status of the associated path, select nodes that can be further explored from the asset entity nodes; the user attention score of the nodes that can be further explored is higher than the attention threshold and there are undisplayed associated paths.
[0180] Specifically, in embodiment S53, the system traverses all displayed asset entity nodes and filters asset entity nodes whose user attention score is higher than the attention threshold and which still have undisplayed next-hop related nodes in the asset knowledge graph as nodes that can continue to be explored. The attention threshold is determined by taking the upper quartile of the attention scores of all displayed nodes after sorting, that is, using the attention score corresponding to the 75th percentile as the attention threshold. The upper quartile is selected based on the fact that in the statistical distribution of attention scores, the upper quartile divides the scores into a higher attention part and a lower attention part. Nodes higher than this threshold represent the 25% of nodes with the highest attention, which meets the business requirement of focusing on a few high-value related paths in the asset management scenario. If the number of nodes in the user's historical exploration path is insufficient, the system uses the current highest attention score multiplied by 80% as the threshold, and automatically switches to the upper quartile threshold strategy after the exploration path is accumulated.
[0181] S54. Output exploration prompts for the nodes that can be further explored through the visualization interface;
[0182] Specifically, in embodiment S54, the system outputs exploration prompts for nodes that can be explored further in the visual interface by highlighting marks or floating prompt boxes. The prompts include the name of the node that can be explored further, the number of undisplayed associated paths, and the suggested next exploration direction.
[0183] The above steps analyze users' historical exploration behavior, calculate the attention score of each node, and proactively prompt users that there are related directions that can be explored further, thus realizing an interactive upgrade from user-driven to system-guided intelligent guidance.
[0184] Example 2:
[0185] Please see Figure 2 and Figure 3 The knowledge graph-based asset management association query and visualization system is used to implement the above-mentioned knowledge graph-based asset management association query and visualization method. It includes a multi-source data acquisition module, a knowledge graph construction module, an anomaly detection module, a query parsing module, an interactive exploration module, and an intelligent guidance module.
[0186] The multi-source data acquisition module is used to acquire multi-source data for asset management.
[0187] The knowledge graph construction module is used to extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source data of asset management through knowledge extraction, and to construct an asset knowledge graph based on the asset entities, asset entity attributes, and relationships between asset entities.
[0188] An anomaly detection module is used to detect abnormal asset entity nodes through the graph structure features of each asset entity node in the asset knowledge graph, obtain a list of abnormal asset entities, and push the list of abnormal asset entities to the user interface.
[0189] The query parsing module is used to perform semantic parsing on the natural language query statement input by the user through the Tongyi Qianwen model to obtain a graph query language statement. The graph query language statement is then used to query the asset knowledge graph to obtain initial query results, which are then displayed through a visual interface.
[0190] The interactive exploration module is used to obtain the current focus node based on the user's selection operation of asset entity nodes in the visualization interface, to query the asset knowledge graph based on the current focus node, to obtain the next-hop related asset entity, to add the next-hop related asset entity to the visualization interface, and to record the user's historical exploration path.
[0191] The intelligent guidance module is used to calculate the user attention score based on the access frequency and dwell time of each asset entity node in the historical exploration path, mark the asset entity nodes with the user attention score higher than a preset threshold and with undisplayed related paths as nodes that can continue to be explored, and output the exploration prompt information of the nodes that can continue to be explored through the visualization interface.
[0192] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0193] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.
Claims
1. A knowledge graph-based method for asset management association query and visualization, characterized in that, Includes the following steps: S1. Obtain multi-source asset management data, and extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source asset management data through knowledge extraction; An asset knowledge graph is constructed based on the asset entities, the asset entity attributes, and the relationships between the asset entities. S2. Detect abnormal asset entity nodes by using the graph structure features of each asset entity node in the asset knowledge graph to obtain a list of abnormal asset entities; Push the list of abnormal asset entities to the user interface; S3. Semantic parsing of the natural language query statement input by the user is performed using the Tongyi Qianwen model to obtain the graph query language statement; The initial query results are obtained by performing queries in the asset knowledge graph using the graph query language statement, and then displayed through a visual interface. S4. Based on the user's selection operation of asset entity nodes in the visualization interface, obtain the current focus node; Based on the current focus node, a query is performed in the asset knowledge graph to obtain the next-hop related asset entity; Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path; S5. Based on the access frequency and dwell time of each asset entity node in the historical exploration path, calculate the user attention score. The asset entity nodes whose user attention scores are higher than the attention threshold and have undisplayed associated paths are marked as nodes that can be further explored; The visualization interface outputs exploration prompts for the nodes that can be explored further.
2. The knowledge graph-based asset management association query and visualization method according to claim 1, characterized in that, S1 includes the following steps: S11. Obtain multi-source asset management data, including asset master data and asset relationship data; S12. Extract asset entities and asset entity attributes from the asset master data; extract relationships between asset entities from the asset association data; S13. Construct an asset knowledge graph based on the asset entity, the asset entity attributes, and the relationships between the asset entities.
3. The knowledge graph-based asset management association query and visualization method according to claim 1, characterized in that, S2 includes the following steps: S21. Calculate the graph structure features of each asset entity node in the asset knowledge graph, wherein the graph structure features include degree centrality, betweenness centrality and clustering coefficient; S22. Detect abnormal asset entity nodes through the graph structure features to obtain the list of abnormal asset entities; S23. Push the list of abnormal asset entities to the user interface.
4. The knowledge graph-based asset management association query and visualization method according to claim 3, characterized in that, The step S22, which involves detecting abnormal asset entity nodes using the graph structure features, includes the following steps: S221. Determine asset entity nodes with degree centrality below the low degree centrality threshold as isolated nodes; S222. Asset entity nodes with betweenness centrality higher than the high betweenness centrality threshold are identified as critical hub nodes. S223. Asset entity nodes with clustering coefficients higher than the high clustering threshold are classified as abnormally dense nodes; S224. The isolated node, the critical hub node, and the abnormally dense node are designated as the abnormal asset entity node.
5. The knowledge graph-based asset management association query and visualization method according to claim 1, characterized in that, S3 includes the following steps: S31. Receive the natural language query statement input by the user; S32. Semantic parsing of the natural language query statement is performed using the Tongyi Qianwen model to obtain the graph query language statement; the training dataset of the Tongyi Qianwen model comes from the historical query logs of the enterprise asset management system. S33. Use the graph query language statement to perform a query in the asset knowledge graph to obtain the initial query results; S34. Display the initial query results through a visual interface.
6. The knowledge graph-based asset management association query and visualization method according to claim 5, characterized in that, S32 includes the following steps: S321. Generate semantic parsing prompt word templates based on the graph pattern information of the asset knowledge graph; S322. Fill the query statement placeholder position in the semantic parsing prompt word template with the natural language query statement to obtain the complete prompt word; S323. Input the complete prompt word into the general meaning thousand questions model, and perform semantic parsing through the general meaning thousand questions model to obtain the initial graph query language statement; S324. Perform syntax validation on the initial graph query language statement. If the syntax of the initial graph query language statement is correct, use the initial graph query language statement as the graph query language statement. S325. When the initial graph query language statement has a syntax error, the syntax error information and the initial graph query language statement are re-inputted into the general meaning question model. The graph query language statement is obtained by modifying the general meaning of the thousand questions model; the modified graph query language statement is then used as the graph query language statement.
7. The knowledge graph-based asset management association query and visualization method according to claim 1, characterized in that, S4 includes the following steps: S41. Receive the user's selection operation on the asset entity node in the visualization interface, and obtain the selected asset entity node. S42. Set the selected asset entity node as the current focus node; S43. Query the asset knowledge graph through the current focus node to obtain the next-hop related asset entity; the next-hop related asset entity does not include asset entity nodes that have appeared in the historical exploration path; S44. Add the next-hop associated asset entity to the visualization interface and record the user's historical exploration path.
8. The knowledge graph-based asset management association query and visualization method according to claim 1, characterized in that, S5 includes the following steps: S51. Extract the access frequency and dwell time of each asset entity node in the user's historical exploration path; S52. Calculate the user attention score based on the access frequency and the dwell time; S53. Based on the user attention score and the display status of the associated path, select nodes that can be further explored from the asset entity nodes; the user attention score of the nodes that can be further explored is higher than the attention threshold and there are undisplayed associated paths. S54. Output exploration prompts for the nodes that can be further explored through the visualization interface.
9. The knowledge graph-based asset management association query and visualization method according to claim 8, characterized in that, The calculation of user attention score for each asset entity node in S52 based on the access frequency and dwell time includes the following steps: S521. Obtain the access frequency of each asset entity node in the historical exploration path; S522. Obtain the dwell time of each asset entity node during each visit in the historical exploration path, and calculate the average dwell time of each asset entity node; S523. The user attention score of each asset entity node is obtained by adding the product of the access frequency and the frequency weight and the product of the average dwell time and the duration weight.
10. A knowledge graph-based asset management association query and visualization system, characterized in that, The system is used to implement the knowledge graph-based asset management association query and visualization method as described in any one of claims 1 to 9, the system comprising: The multi-source data acquisition module is used to acquire multi-source data for asset management. The knowledge graph construction module is used to extract asset entities, asset entity attributes, and relationships between asset entities from the multi-source data of asset management through knowledge extraction, and to construct an asset knowledge graph based on the asset entities, asset entity attributes, and relationships between asset entities. An anomaly detection module is used to detect abnormal asset entity nodes through the graph structure features of each asset entity node in the asset knowledge graph, obtain a list of abnormal asset entities, and push the list of abnormal asset entities to the user interface. The query parsing module is used to perform semantic parsing on the natural language query statement input by the user through the Tongyi Qianwen model to obtain a graph query language statement. The graph query language statement is then used to query the asset knowledge graph to obtain initial query results, which are then displayed through a visual interface. The interactive exploration module is used to obtain the current focus node based on the user's selection operation of asset entity nodes in the visualization interface, to query the asset knowledge graph based on the current focus node, to obtain the next-hop related asset entity, to add the next-hop related asset entity to the visualization interface, and to record the user's historical exploration path. The intelligent guidance module is used to calculate the user attention score based on the access frequency and dwell time of each asset entity node in the historical exploration path, mark the asset entity nodes with the user attention score higher than a preset threshold and with undisplayed related paths as nodes that can continue to be explored, and output the exploration prompt information of the nodes that can continue to be explored through the visualization interface.
Citation Information
Patent Citations
Data asset management method and device, storage medium and program product
CN121277911A
Intelligent unknown threat determination method and system based on multi-modal knowledge graph
CN121547292A
