Electric power overhaul decision-making method and device based on knowledge domain mapping map retrieval generation

By constructing a knowledge domain mapping map for power maintenance and using a community detection algorithm to generate community summaries, the problem of insufficient intelligent decision support in the power maintenance system is solved, achieving efficient maintenance information processing and decision support, and improving the operation and maintenance efficiency and reliability of the power system.

CN120912168APending Publication Date: 2025-11-07BEIJING JIAOTONG UNIV +2
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202510857134.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-25
Publication Date
2025-11-07

AI Technical Summary

Technical Problem

The existing power maintenance system lacks intelligent decision support, resulting in low data integration efficiency, high risk of subjective misjudgment, and difficulty in achieving reasonable scheduling and real-time compliance of the maintenance information platform.

Method used

A knowledge domain mapping method is adopted, which constructs a power maintenance knowledge domain mapping map through deep semantic analysis using a large language model (LLM), generates hierarchical community summaries using the Leiden community detection algorithm, and achieves semantic matching of user queries through vectorization technology to generate power maintenance decision answers.

Benefits of technology

It enables automated knowledge extraction and graph construction from unstructured maintenance texts, providing highly reliable fault diagnosis, maintenance plan optimization, and risk warning, thereby reducing operation and maintenance costs and improving power supply reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120912168A_ABST
    Figure CN120912168A_ABST
Patent Text Reader

Abstract

The invention provides an electric power overhaul decision-making method and device based on knowledge domain mapping map retrieval generation. The method comprises the following steps: performing deep semantic analysis on an electric power overhaul file through a prompt project guide large language model LLM, and constructing an electric power overhaul knowledge domain mapping map; generating communities with hierarchical structures by adopting a community detection algorithm Leiden based on the electric power overhaul knowledge domain mapping map, and creating a community abstract in a report form for each community; and mapping an electric power overhaul query request initiated by a user to the embedding space of the electric power overhaul knowledge domain mapping map through a vectorization technology to realize semantic matching with the community abstract, and outputting an electric power overhaul decision answer result corresponding to the electric power overhaul query request. According to the method, an intelligent question-answering system in the field of electric power overhaul is constructed by fusing knowledge domain mapping map retrieval enhancement generation and LLM, automatic knowledge extraction and map construction of unstructured overhaul texts are realized, and high-reliability decision support is provided for fault diagnosis, overhaul plan optimization and risk early warning of electric power equipment.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power maintenance decision-making, and particularly relates to a power maintenance decision-making method and device based on knowledge domain mapping map retrieval generation. BACKGROUND

[0002] With the advancement of new power system construction and large-scale grid connection of renewable energy, the expansion of distribution network and the surge of multi-source heterogeneous data have greatly increased the complexity of distribution network operation. Artificial arrangement faces problems such as low data integration efficiency and high subjective misjudgment risk. At the same time, the existing technology has not fully covered the intelligent needs of the maintenance process, and the lack of maintenance information platform makes it difficult for the maintenance department to obtain technical support and to reasonably arrange the time of each project that needs to be maintained at the same time. Therefore, it is urgent to explore an intelligent decision-making method for distribution network maintenance operation mode.

[0003] Maintenance business data, as the core carrier of safe operation and maintenance of the power system, is created and maintained by field maintenance personnel, equipment management personnel and technical support team, and covers key information such as power safety operation procedures, equipment maintenance standards and emergency handling guidelines. However, although the maintenance files contain rich information, their complexity, professionalism and massiveness pose a serious challenge to traditional information retrieval methods. The main performances are as follows: ① The power maintenance documents are usually lengthy and terminologically intensive, and manual retrieval is time-consuming and inefficient. ② Due to the complexity of document expression, practitioners are prone to operational errors due to understanding bias. ③ The power industry is strictly regulated, and manual review cannot ensure real-time compliance.

[0004] Therefore, it is urgent to build an intelligent decision-making question and answer system with high professionalism for the maintenance system. SUMMARY

[0005] Embodiments of the present application provide a power maintenance decision-making method and device based on knowledge domain mapping map retrieval generation to effectively improve the efficiency of power maintenance answer decision-making.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions.

[0007] According to one aspect of the present application, a power maintenance decision-making method based on knowledge domain mapping map retrieval enhancement generation is provided, comprising:

[0008] The power maintenance files are analyzed in depth by the large language model LLM guided by the prompt engineering to construct a power maintenance knowledge domain mapping map;

[0009] Based on the power maintenance knowledge domain mapping map, a community detection algorithm Leiden is used to generate a hierarchical community, and a community summary in the form of a report is created for each community;

[0010] The user-initiated power maintenance query request is mapped to the embedding space of the power maintenance knowledge domain mapping map through vectorization technology, semantic matching with the community summary is realized, and a power maintenance decision answer result corresponding to the power maintenance query request is output.

[0011] Preferably, the power maintenance file is subjected to deep semantic analysis by the LLM guided by the prompting engineering to construct the power maintenance knowledge domain mapping map, comprising:

[0012] Source files related to power maintenance are collected, the source files are read and parsed, and saved as UTF-8 encoded txt files, the source files including unstructured text power maintenance reports and operation logs, the txt files are subjected to text segmentation to decompose long documents into slices, the slices are subjected to deep semantic analysis by the LLM guided by the prompting engineering using few-shot learning, step-by-step reasoning, parameterized templates, entity disambiguation and domain adaptation, key maintenance entities and the associated relationships between the maintenance entities are extracted, and the power maintenance knowledge domain mapping map is constructed.

[0013] Preferably, the power maintenance knowledge domain mapping map is subjected to community detection algorithm Leiden to generate hierarchical community structures, and a community summary in the form of a report is created for each community, comprising:

[0014] The nodes of the power maintenance knowledge domain mapping map are divided into multiple connected communities according to semantic or structural features by the community detection algorithm Leiden through node movement, partition refinement and network aggregation, each community is a node group, and the community corresponds to a semantic unit in the power maintenance knowledge;

[0015] Each edge in the community is sorted in descending order according to the combination degree of the source node and the target node, the source node, the target node, the edge itself and the description of the related declaration are sequentially added to the context window of the LLM according to the priority until the Token limit of the context window of the LLM is reached, a multi-level community summary in the form of a report is generated for each community, and a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation and important declarations is generated based on the community and the community summary according to the set prompt words.

[0016] Preferably, the user-initiated power maintenance query request is mapped to the embedding space of the power maintenance knowledge domain mapping map through vectorization technology, semantic matching with the community summary is realized, and a power maintenance decision answer result corresponding to the power maintenance query request is output, comprising:

[0017] Each community is regarded as a semantically consistent local subgraph, and each subgraph corresponding community summary is taken as the basis information source for subsequent question and answer generation. When a user initiates a power maintenance query request, the system first analyzes the query request, extracts multiple target entities from the user's query request, extracts the adjacent nodes, key relationships and attribute information of the target entities under the guidance of the power maintenance knowledge domain mapping map structure, and associates the corresponding original document fragments to construct a context window integrating structured graph and unstructured corpus, and obtains multiple community summaries corresponding to the query request;

[0018] The Nomic embedding model is used to vectorize and encode the text of the user query request and the corresponding multiple community summaries, the semantic similarity between the user query request vector and each community summary vector is calculated by using the cosine similarity algorithm, and the most relevant community summary is selected based on the set similarity threshold according to the semantic similarity;

[0019] The current problem and historical dialogue context of the user query request are combined, the most relevant community summary selected is taken as a prompt prefix, the generation semantic space of the LLM is limited, and the power maintenance decision answer result corresponding to the power maintenance query request is generated by using the LLM.

[0020] Preferably, the method further comprises: visualizing and explaining the power maintenance knowledge domain mapping map according to the community report by using the ForceAtlas2 algorithm, and the visualizing and explaining result comprises: topological structure, community clustering and dynamic interaction view.

[0021] According to another aspect of the present application, a device for retrieving and enhancing the generation of power maintenance decision based on knowledge domain mapping map is provided, comprising: a power maintenance knowledge domain mapping map construction module, a community and community summary generation module, and a power maintenance decision answer result acquisition module.

[0022] The power maintenance knowledge domain mapping map construction module is used to guide the large language model (LLM) to perform deep semantic analysis on the power maintenance file by using a prompt engineering, and construct a power maintenance knowledge domain mapping map.

[0023] The community and community summary generation module is used to generate hierarchical communities by using the community detection algorithm Leiden based on the power maintenance knowledge domain mapping map, and create a community summary in the form of a report for each community.

[0024] The power maintenance decision answer result acquisition module is used to map the power maintenance query request initiated by a user to the embedding space of the power maintenance knowledge domain mapping map by using vectorization technology, realize semantic matching with the community summary, and output the power maintenance decision answer result corresponding to the power maintenance query request.

[0025] Preferably, the power maintenance knowledge domain mapping map construction module is specifically used for collecting power maintenance related source files, reading and parsing the source files, saving as a UTF-8 encoded txt file, the source files including unstructured text power maintenance reports and operation logs, performing text segmentation on the txt file, decomposing long documents into slices, guiding the LLM to use few-shot learning, step-by-step reasoning, parameterized templates, entity disambiguation and domain adaptation to perform deep semantic analysis on the slices, extract key maintenance entities and the correlation between maintenance entities, and construct a power maintenance knowledge domain mapping map.

[0026] Preferably, the community and community summary generation module is specifically used for dividing nodes of the power maintenance knowledge domain mapping map into multiple connected communities according to semantic or structural features by using a community detection algorithm Leiden through node movement, partition refinement and network aggregation, each community being a node group, and the community corresponding to a semantic unit in the power maintenance knowledge.

[0027] For each edge in the community, the source node and the target node are sorted in descending order according to the combination degree, and the source node, the target node, the edge itself and the description of the related statement are added to the context window of the LLM in turn according to the priority until the Token limit of the context window of the LLM is reached, a multi-level report form community summary is generated for each community, and a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation and important statements is generated based on the community and the community summary according to the set prompt words.

[0028] Preferably, the power maintenance decision answering result acquisition module regards each community as a semantically consistent local subgraph, and regards the community summary corresponding to each subgraph as a basic information source for subsequent question answering, when a user initiates a query request for power maintenance, the system first analyzes the query request, extracts multiple target entities from the user's query request, extracts the adjacent nodes, key relationships and attribute information of the target entities under the guidance of the power maintenance knowledge domain mapping map structure, and associates the corresponding original document fragments to construct a context window integrating the structured graph and the unstructured corpus, and obtains multiple community summaries corresponding to the query request.

[0029] The text of the user query request and the multiple community summaries is vectorized and encoded by using the Nomic embedding model, the semantic similarity between the user query request vector and each community summary vector is calculated by using the cosine similarity algorithm, and the most relevant community summary is selected according to the semantic similarity based on the set similarity threshold.

[0030] In combination with the current problem of the user query request and the historical dialogue context, the most relevant community summary screened out is taken as a prompt prefix to limit the generation semantic space of the LLM, and the LLM is used to generate a power maintenance decision answer result corresponding to the power maintenance query request.

[0031] Preferably, the device further comprises a visual interpretation module configured to visually interpret the power maintenance knowledge domain mapping map according to the community report by using a ForceAtlas2 algorithm, and the visual interpretation result comprises a topological structure, community clustering and a dynamic interactive view.

[0032] As can be seen from the technical solutions provided by the above-mentioned embodiments of the present application, the embodiments of the present application construct an intelligent question and answer system in the power maintenance field by fusing the knowledge domain mapping map retrieval enhancement generation and the LLM, realize the automatic knowledge extraction and graph construction of unstructured maintenance text, provide high-reliability and interpretable decision support for power equipment fault diagnosis, maintenance plan optimization and risk warning, effectively reduce the operation and maintenance cost and improve the power supply reliability.

[0033] Additional aspects and advantages of the application will be set forth in part in the following description, and in part will become apparent to those skilled in the art upon examination of the following description, or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0035] Figure 1 A power maintenance decision method based on knowledge domain mapping map retrieval generation provided by the embodiments of the present application is provided.

[0036] Figure 2 A knowledge base preprocessing flow provided by the embodiments of the present application is provided.

[0037] Figure 3 A knowledge domain mapping map community report prompt word provided by the embodiments of the present application is provided.

[0038] Figure 4 A search workflow provided by the embodiments of the present application is provided.

[0039] Figure 5 A structure diagram of a power maintenance decision device based on knowledge domain mapping map retrieval enhancement generation provided by the embodiments of the present application is provided. DETAILED DESCRIPTION

[0040] Embodiments of the present application will be described in detail below with reference to the attached drawings, which show by way of example, embodiments in which the same or similar elements or elements having the same or similar functions are denoted by the same or similar reference numerals throughout the drawings. The embodiments described below with reference to the drawings are exemplary and are intended only for the purpose of explaining the present application, and cannot be construed as limiting the present application.

[0041] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the use of the phrase "comprising" in the specification of the present application means that the features, integers, steps, operations, elements and / or components described exist, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to the other element, or there can be intermediate elements. In addition, "connected" or "coupled" used herein can include wireless connection or coupling. The phrase "and / or" used herein includes any one of the associated listed items and all combinations thereof.

[0042] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as that generally understood by those skilled in the art to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have meanings consistent with those in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as such.

[0043] In order to facilitate the understanding of the embodiments of the present application, the following will be further explained with reference to the drawings by way of several specific examples, and each embodiment does not constitute a limitation on the embodiments of the present application.

[0044] The embodiments of the present application construct a power maintenance decision-making method combining Knowledge Graph Retrieval-Augmented Generation (KG-RAG) and LLM (Large Language Model).

[0045] The present application mainly includes three parts. First, the power standard text data is mapped to a semantic correlation graph; then, an open source knowledge domain mapping map RAG system is constructed for power maintenance; finally, the constructed question and answer model not only assists field personnel to quickly formulate maintenance schemes, but also provides decision-making reference based on historical experience for technical experts, and has a wide application prospect.

[0046] Embodiment One

[0047] The present application aims to propose a power maintenance decision-making scenario based on a knowledge domain mapping map retrieval enhanced generation of power maintenance decision-making method The processing flow is as shown in Figure 1 The processing steps are as follows:

[0048] Step S1: Constructing a power maintenance knowledge domain mapping map based on prompt engineering guided LLM.

[0049] The present application is based on knowledge domain mapping map retrieval enhanced generation of LLM, and uses LLM for power maintenance knowledge domain mapping map construction and decision-making question and answer. First, collect and preprocess the maintenance knowledge files to form different maintenance text slices. Then, build a prompt word guided LLM to extract key maintenance entities and maintenance relationships from the maintenance text slices to form a power maintenance knowledge domain mapping map. Secondly, divide the maintenance atlas associated node community hierarchical structure, and use LLM to summarize the intra-level or cross-level relationships to generate community content community abstract. Finally, humans use LLM to carry out global and local query maintenance query tasks.

[0050] First, data collection is carried out, and on the basis of data collection, maintenance file data transformation is carried out, the purpose of which is to extract operable knowledge units from unstructured data to support subsequent entity and relationship extraction tasks. For large model RAG technology, its processing of knowledge files is mostly formalized into txt format, therefore, the contents of common pdf format files such as papers and standards are extracted through pdfplumber, the source folder is read, and each page is parsed to finally save as a UTF-8 encoded txt file. Then, the power maintenance knowledge text in the above-mentioned txt format is segmented, and the long document is decomposed into small and easy-to-manage chunks. The purpose of the chunk is to disassemble long texts such as maintenance procedures, equipment manuals or fault records into short and clear parts, so as to facilitate the efficiency of subsequent knowledge domain mapping map construction and auxiliary question and answer process. The preset chunk length can be flexibly adjusted according to specific task requirements or technical constraints to adapt to the requirements of knowledge extraction and analysis in different scenarios, and the present application adopts 300Tokens.

[0051] After the power maintenance text preprocessing is completed, the power maintenance knowledge domain mapping map entity relationship extraction and graph construction based on prompt engineering guided LLM is carried out. The entity and relationship extraction of the text is essentially an abstract community abstract process for the maintenance field, through the deep semantic analysis of unstructured texts such as maintenance reports and operation logs, the key maintenance entities (such as transformers and circuit breakers) and their implied associated relationships (such as cause and influence) are extracted. These elements often exist in an implicit way in the original maintenance text.

[0052] Figure 2A knowledge base preprocessing flowchart is provided for an embodiment of the application. The prompt engineering guides the LLM to adopt technologies such as few-shot learning, step-by-step reasoning, parameterized templates, entity disambiguation and domain adaptation. Specifically, the few-shot learning strategy provides three typical power industry cases as paradigms, which not only limits the task boundary but also conveys domain knowledge; at the same time, through the step-by-step reasoning mechanism, entity recognition and relationship extraction are decoupled, and the model is forced to build an entity knowledge base before correlation analysis to avoid information confusion. The parameterized template design gives flexibility to the prompt word, and dynamic placeholders such as {entity_types} support cross-scene reuse, and the structured separator system (tuple / record / completion_delimiter) improves the analyzability. At the same time, for the problem of entity disambiguation, the descriptive anchoring strategy (such as "facilities serving 50,000 users") and the context binding mechanism are adopted, combined with the uppercase naming specification to strengthen the uniqueness of the entity. In particular, the introduction of the relationship strength quantification system (1-7 score) breaks through the traditional binary relationship judgment, and through the implicit scoring standard (such as personnel direct relationship 7 points, equipment support relationship 4 points), it realizes fine-grained relationship modeling, and finally forms a knowledge domain mapping map output with domain professionalism and structural specification.

[0053] Take the relevant power maintenance paragraph as an example for analysis. After processing the prompt word, the entity and relationship correspondence table is as shown in Figure 3 It can be seen that the LLM guided by the prompt engineering has certain domain adaptability and structured information extraction capability in the construction of the power maintenance knowledge domain mapping map:

[0054] 1) In the aspect of maintenance entity recognition, the LLM accurately captures explicit named entities (such as "Jordan Lee" personnel name), and identifies non-explicit domain features through context reasoning. For example, when classifying "High-Voltage Tower M7" as a FACILITY type, it automatically integrates key business attributes such as geographic location (near Redwood National Reserve) and operation and maintenance status (last maintenance time six months ago) in the description. Deep semantic understanding benefits from the descriptive anchoring strategy in the designed prompt engineering, which forces the model to build an entity description field, effectively realizing the structured extraction of entity attributes.

[0055] 2) In the relationship extraction dimension, the model breaks through the limitations of traditional co-occurrence relationship recognition and has certain fine-grained semantic association modeling capability. The "maintenance" relationship (strength 5) between PowerCare Ltd and High-Voltage Tower M7 and the "assignment detection" relationship (strength 4) between technical personnel form differentiated representation, verifying the effectiveness of the implicit scoring standard: the institutional responsibility relationship between organizations and facilities is given higher weight, while the temporary task association between personnel and facilities is scored lower. This quantitative evaluation system provides a calculable association degree basis for subsequent graph reasoning applications, enabling business scenarios such as maintenance priority determination and responsibility tracing to build decision-making models based on relationship strength values.

[0056] It is worth noting that the LLM exhibits precise disambiguation capability when dealing with composite entities. Although the original text uses "structure" as a reference, the model can accurately associate it to "High-Voltage Tower M7" through context binding, thanks to the capital naming specification constraint in the prompt strategy and the entity description generation mechanism. Meanwhile, the step-by-step reasoning design effectively avoids information confusion, i.e., the process of first establishing a complete entity library and then conducting association analysis, ensuring that the implicit employment relationship between "PowerCare Ltd" and the three technical personnel, although not explicitly stated in the text, can be extended and reasoned through the co-occurrence relationship between organizational entities and personnel entities.

[0057] Step S2: Based on the power maintenance knowledge domain mapping map, use the community detection algorithm Leiden to generate a hierarchical community, and create a community summary in the form of a report for each community.

[0058] After the construction of the power maintenance knowledge domain mapping map, the community detection algorithm Leiden is used to divide the power maintenance knowledge domain mapping map into several tightly connected communities, each of which is a node group. These communities correspond to semantic units in the power maintenance knowledge and can provide a structured basis for subsequent decision support.

[0059] Leiden algorithm is an improved algorithm for community detection, which optimizes community structure through three stages of node movement, partition refinement, and network aggregation.

[0060] 1) Node movement: The algorithm optimizes the quality function by moving individual nodes. Each node is moved to the community that maximizes the quality function. Leiden algorithm uses a fast local movement method that only accesses nodes whose neighbors have changed, thereby improving efficiency. Leiden algorithm can optimize multiple quality functions, among which the most commonly used are modularity and constant Potts model. Modularity measures the quality of community structure and is defined as:

[0061]

[0062] where: e c The number of edges within the community c. K c The sum of the degrees of all nodes in the community. m is the total number of edges in the network. γ is a resolution parameter used to control the granularity of the communities. Higher values of γ result in more, smaller communities.

[0063] 2) Refinement of partitions: In this phase, the algorithm refines the partitions, further dividing the communities into sub-communities. By randomly merging nodes into other communities, the algorithm ensures that each community is internally connected and optimizes the quality function. The refinement process allows for a certain degree of randomness to explore a wider space of partitions.

[0064] 3) Network aggregation: In this phase, the algorithm constructs a new aggregated network based on the refined partitions. Each community is treated as a node in the aggregated network, and the algorithm repeats the above process on the aggregated network until the quality function cannot be further optimized.

[0065] After generating the communities for the power maintenance based on the Leiden algorithm, a community summary in the form of a report is created for each community in the hierarchy.

[0066] These community summaries can be used independently as a way to understand the global structure and semantics of the dataset, and they can be used themselves to understand the corpus without specific queries. A template is created for each community that contains the community summary of various elements (such as nodes, edges, and related claims). A multi-level community summary is generated for each community. Users can browse the community summary at a high level to find topics of interest, and then read more detailed sub-community summaries at a lower level for more details. These community summaries can also be used as part of a graph-based index to answer global queries.

[0067] The process of generating a community summary is as follows: First, for each edge in the community, sort the source node and target node in descending order of their combined degree (i.e., their overall prominence in the network). Second, add the descriptions of the source node, target node, edge itself, and its related claims to the context window of the LLM in order of priority. Finally, continue adding until the Token limit of the context window is reached.

[0068] The prompt words revolve around information discovery integrity and structuring. The prompt words first clearly define the task goal, which is to generate a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation, and important statements. Second, the design logic of the report structure embodies the characteristics of hierarchy and standardization. The designed prompt words require the report to include a title, a community summary, an impact severity score, a score explanation, and detailed findings. The introduction of impact severity score and score explanation further enhances the quantitative analysis capability of the report, enabling power maintenance decision-makers to quickly assess the importance of the community through the community report. The design logic of the data reference rules emphasizes transparency and credibility. At the same time, the prompt words require that when referencing data in the report, the data set name and record ID are listed, and the number of records referenced is limited. This rule not only ensures the transparency of the report. Finally, considering the uncertainty of the output of LLM, a method is needed to standardize the output so that it is easier for programs to parse. JSON is a lightweight data exchange format that is easy for humans to read and write, and is easy for machines to parse and generate, making it suitable for large model standardized output. The prompt words require the report to be returned in JSON format, making it easy for users to share and use the report between different systems or tools.

[0069] Step S3: Visualize the power maintenance knowledge domain mapping map according to the community report using the ForceAtlas2 algorithm.

[0070] After the community report is generated, in the data preprocessing stage, the three types of structured data of nodes, relationships and community identification results extracted and sorted are first read, and the column structure is checked and standardized to adapt to the graph construction requirements.

[0071] In the visualization process of the power maintenance knowledge domain mapping map, considering the characteristics of the power maintenance knowledge domain mapping map, such as heterogeneous multi-level, tight clustering, uneven edge weight and large scale, the ForceAtlas2 algorithm is used as the main method of graph spatialization. ForceAtlas2 is a continuous force-directed layout algorithm, with good readability, interactive friendliness and excellent expression ability for community structure, widely used in complex network visualization.

[0072] Community-driven layout optimization: based on the community information divided by the Leiden algorithm, ForceAtlas2 enhances the attractive force weight of nodes within the community, so that device entities within the same community are closely clustered in space, forming visual clusters. At the same time, by adjusting the repulsive force parameter between communities, a reasonable distance between different community clusters is achieved, intuitively reflecting the modular semantic structure of power maintenance knowledge.

[0073] Visual mapping of community summary information: Based on the key entity attributes, fault correlation strength, and impact score extracted in the community summary, the importance of the device is mapped through the node size, the relationship strength is mapped through the edge transparency, and the community affiliation is distinguished by the node color, forming a multi-dimensional semantic encoding. The impact severity score generated in the community summary is further used to dynamically adjust the node gravity parameter, highlighting the core position of high-risk devices in the graph.

[0074] What does the visual explanation result include?

[0075] Topology: Present the device entity nodes (such as transformers, circuit breakers), fault correlation edges (such as "cause" and "impact"), and their hierarchical relationships, expressing the semantic correlation strength through spatial proximity.

[0076] Community clustering: Differentiate different community clusters by color coding, with each cluster corresponding to a specific fault mode or maintenance scenario (such as typhoon protection and insulation aging), and the boundaries are clearly distinguishable.

[0077] Dynamic interactive view: Support node hovering to display detailed attributes (such as device model and historical fault frequency), relationship edge to show weight values (such as fault propagation probability), and provide semantic induction through community summary pop-up windows (such as "high typhoon damage area along the coastal power transmission line").

[0078] The basic principle of the ForceAtlas2 algorithm is to simulate a physical system: the nodes in the graph are considered as particles with the same charge, and there is a repulsive force between them; while the edges are considered as springs, and there is an attractive force between the connected nodes. By simulating the combined effect of these forces, the nodes continuously move in space, gradually reaching a state of force balance, thus forming a two-dimensional layout in which structurally similar nodes are visually similar. This layout can effectively reveal the modular structure and local clustering characteristics of fault correlations between power devices.

[0079] In mathematical modeling, ForceAtlas2 mainly relies on the following two types of force calculation formulas:

[0080] 1) Attractive force, used for the pulling effect between connected nodes, defined as:

[0081] F a (n1,n2)=d(n1,n2) (2)

[0082] Where: d(n1,n2) represents the Euclidean distance between nodes, the attractive force is proportional to the distance, reflecting the edge's pulling effect.

[0083] 2) Repulsive force, used for the mutual repulsion between all nodes, considering the influence of node degree, its form is:

[0084]

[0085] where deg(n) denotes the degree of a node (number of connected edges), k r represents the repulsion constant set for the user. By introducing the "degree+1" adjustment mechanism, the algorithm effectively reduces the visual congestion caused by high-connectivity nodes to low-connectivity nodes, making "leaf nodes" closer to the central node, thereby alleviating the edge dispersion phenomenon of the graph.

[0086] In addition, ForceAtlas2 introduces a gravity mechanism to avoid isolated subgraphs drifting to the edge of the view, which is in the form of:

[0087] F g (n)=k g ·(deg(n)+1) (4)

[0088] To improve the performance of the algorithm in large networks, ForceAtlas2 enables the Barnes-Hut approximation method to reduce the time complexity of the repulsion force calculation from O(n 2 ) to O(nlogn). At the same time, the algorithm uses a local and global adaptive speed adjustment mechanism to dynamically adjust the node movement step size by evaluating the "swing amplitude" and "effective traction" of the node, ensuring that high-connectivity nodes have higher stability, while low-degree nodes can quickly converge, thereby balancing the layout quality and computational efficiency. ForceAtlas2 algorithm maps structural information to a physical force field, which not only effectively presents the fault correlation between devices in the power maintenance knowledge domain mapping map visualization, but also reveals the potential community structure through a compact and clear graphical layout, providing structured visual support for subsequent decision-making and knowledge retrieval.

[0089] In the visualization of the power maintenance knowledge domain mapping map, the interpretability of the ForceAtlas2 layout algorithm mainly lies in the following key aspects, which support the visual understanding and analysis of power maintenance decision-making from the perspectives of structure restoration, visual clustering, and semantic mapping:

[0090] 1) Structural similarity is transformed into spatial proximity. ForceAtlas2 automatically aggregates nodes that are structurally closely related in space by simulating attractive and repulsive forces in physical systems.

[0091] 2) Visualization of community structure. ForceAtlas2 is naturally suitable for revealing the modularity structure in the network, i.e., community clusters. Combined with the Gephi community detection algorithm and node coloring strategy, these structures can be presented in a color+space position dual manner, enhancing the semantic interpretation ability of the graph;

[0092] 3) Semantic expression of the relationship between edge weight and fault strength. ForceAtlas2 supports the adjustment of the influence of edge weight on attraction. The greater the edge weight, the stronger the attraction, and the closer the visual distance between nodes. This mechanism has the following interpretation effects in the power maintenance knowledge domain mapping map:

[0093] 4) Dynamic interaction supports "human-machine consensus modeling". Since ForceAtlas2 is a continuous layout algorithm, real-time dynamic visualization and parameter interaction adjustment can be performed in Gephi. Users can observe the structural evolution of the graph under different gravity / repulsion parameters, manually drag key nodes to observe their impact on the overall graph, adjust gravity and edge weight influence, and explore different clustering structures. This interactive modeling process not only improves the interpretability of the graph, but also provides a "consensus modeling" platform for power experts to transition from experience cognition to data understanding.

[0094] Step S4: Vectorize the user's power maintenance query text, filter out the text units of the relevant community summary through the semantic matching mechanism, and obtain the answer to the user's power maintenance query text using the local question answering mechanism.

[0095] The input data of this step is the natural language query request initiated by the user. The system maps the query semantics to the embedding space of the knowledge domain mapping map through vectorization technology, and realizes semantic matching with the community summary. The output data is an intelligent answer result generated by combining the semantic community summary and the context of the local knowledge domain mapping map. After the multi-source information is fused and reasoned by the large language model, a structured decision-making suggestion conforming to the power maintenance specification is formed.

[0096] After completing the community division based on the Leiden algorithm, the nodes in the power maintenance knowledge domain mapping map are divided into multiple "communities" according to semantic or structural features. Each community can be regarded as a locally consistent subgraph. For each subgraph, the system generates a corresponding community summary as a basic information source for subsequent question answering. When the user initiates a query, the system first analyzes the user's question, filters out the text units of the relevant community summary through the semantic matching mechanism, and generates a targeted answer using LLM. In this process, the user query and the community summary content are mapped to the same embedding space to achieve more accurate semantic matching and efficient similarity retrieval. Figure 4 A search workflow is provided for the embodiments of the present application.

[0097] The application adopts Nomic embedding model to realize vectorization of user query text, the model has been continuously evolving since 2024, and the representative versions include nomic-embed-text-v1, v1.5 and v2, which are excellent in processing short text semantic matching and long text context understanding tasks, and the performance exceeds mainstream models such as text-embedding-ada-002 of OpenAI. The Nomic model is based on the BERT (Bidirectional Encoder Representations from Transformers) architecture extension, the initial version supports 2048 Token context, and the latest version has been extended to 8192 tokens, which is suitable for long document vectorization tasks. The model training adopts a multi-stage strategy, first pre-training with a self-supervised mask language model target, then fine-tuning on a large-scale text pair dataset through contrastive learning, and the training corpus covers StackExchange, Quora question and answer pairs, Amazon reviews, news community summaries and the like, with a scale of about 235 million pairs. The model can efficiently convert any input text into a fixed-dimensional semantic vector, providing strong support for semantic retrieval and graph alignment.

[0098] After completing the community summary embedding, the power maintenance question and answer system further introduces a local question and answer mechanism to meet the precise query needs of users for entity-level knowledge such as specific equipment, fault types or treatment schemes. In the field of power maintenance, the power maintenance knowledge domain mapping map has a high degree of structural characteristics, there are clear "equipment-fault-treatment" links between entities, and a large amount of operation experience and logical rules are concentrated in the interaction of entities in the local subgraph, therefore, it is difficult to cover such relationship-oriented, reasoning-driven problems by simply relying on global semantic similarity retrieval.

[0099] The application adopts a local question and answer mechanism, and its working principle is as follows:

[0100] 1) Entity recognition and matching: the system first extracts potential target entities (such as equipment names, fault types, etc.) from the user query, and vectorizes the target entities and compares them with the existing entity nodes in the power maintenance knowledge domain mapping map to identify the most relevant entities.

[0101] 2) Structured context construction: under the guidance of the structure of the power maintenance knowledge domain mapping map, the system extracts the adjacent nodes, key relationships and attribute information of the entity, and associates the corresponding original document fragments to construct a context window that integrates structured graphs and unstructured corpus. The Nomic embedding model is used to vectorize the user query statement and the community summary text in the knowledge domain mapping map, the semantic similarity between the user query vector and each community summary vector is calculated by using the cosine similarity algorithm, and the relevant community summaries are extracted by setting a similarity threshold.

[0102] 3) Semantic generation and answer output: Combine the user's current question and historical dialogue context, use the community summary filtered in 2) as a prompt prefix, limit the semantic space of LLM generation, and call LLM to generate high-quality answers that meet semantic and structural logic.

[0103] The relevant indicators of the power maintenance knowledge domain mapping map include average degree, average weighted degree, network diameter, graph density, click times, and modularity, etc. Some key indicators are as follows.

[0104] Average degree: The average degree is the average value of the degrees of all nodes in the graph. The degree represents the number of incoming edges + outgoing edges connected to a node, which is specifically represented as:

[0105]

[0106] In the formula: |E| is the total number of edges in the graph; |V| is the total number of nodes in the graph.

[0107] Average weighted degree: It is the average value of the weighted degrees of all nodes in the graph, and its formula is:

[0108]

[0109] In the formula: V is the node set. N(v) is the node set adjacent to node v; w uv is the weight of the edge between nodes v and u.

[0110] Modularity: Modularity is used to measure the community structure of the network (i.e. how well the graph is divided into multiple "groups" or "clusters"). The evaluation method in the classic literature Modularity Report is adopted. The definition of its modularity Q is as follows:

[0111]

[0112] In the formula: Modularity Q is used to measure the "goodness" of community division; m is the total weight of the edges in the graph (in an unweighted graph, it is the number of edges); A ij is the weight of the edge between nodes i and j (0 or 1 in an unweighted graph); k i is the degree (or weighted degree) of node i; c i is the community number where node i is located; δ(c i ,c j ) is the Kronecker function, which is 1 when c i =c j , otherwise 0.

[0113] The Precision, Recall, and F1 scores are used to evaluate the quality of the large model answers.

[0114] The pre-training model such as BERT is used to embed each word of the generated text and the reference text into a vector. The cosine similarity between each generated word and the reference word is calculated.

[0115] Let the generated text token set be C={c1,c2,...,c m}, and the reference text token set be R={r1,r2,...,r n}, each token is obtained by BERT to get the vector e c ,e r . For each generated word, find the most similar word in the reference (the maximum cosine similarity), and take the average, which is Precision (see formula). For each reference word, find the most similar word in the generation, take the average, get Recall (see formula). Finally, return the F1 score (see formula) to comprehensively reflect the semantic matching.

[0116]

[0117] Embodiment two

[0118] The embodiment of the application provides a power maintenance decision device based on a knowledge domain mapping map retrieval generation, and the structure of the device is as shown in Figure 5 Fig. 1, which comprises a power maintenance knowledge domain mapping map construction module 51, a community and community summary generation module 52, a power maintenance decision answer result acquisition module 53 and a visual interpretation module 54.

[0119] The power maintenance knowledge domain mapping map construction module 51 is used to guide the large language model LLM to perform deep semantic analysis on the power maintenance file through the prompt engineering, and construct the power maintenance knowledge domain mapping map.

[0120] The community and community summary generation module 52 is used to generate hierarchical communities based on the power maintenance knowledge domain mapping map using the community detection algorithm Leiden, and create a community summary in the form of a report for each community.

[0121] The power maintenance decision answer result acquisition module 53 is used to map the power maintenance query request initiated by the user to the embedding space of the power maintenance knowledge domain mapping map through vectorization technology, realize semantic matching with the community summary, and output the power maintenance decision answer result corresponding to the power maintenance query request.

[0122] The visual interpretation module 54 is used to visually interpret the power maintenance knowledge domain mapping map according to the community report using the ForceAtlas2 algorithm, and the visual interpretation result comprises a topological structure, community clustering and a dynamic interactive view.

[0123] The power maintenance knowledge domain mapping map construction module is specifically configured to collect source files related to power maintenance, read and parse the source files, and save them as UTF-8 encoded txt files. The source files include power maintenance reports and operation logs of unstructured text. The txt files are text segmented, and long documents are decomposed into slices. The LLM is guided to perform deep semantic analysis on the slices by prompting engineering, using few-shot learning, step-by-step reasoning, parameterized templates, entity disambiguation, and domain adaptation. Key maintenance entities and the relationships between maintenance entities are extracted, and a power maintenance knowledge domain mapping map is constructed.

[0124] The community and community summary generation module is specifically configured to divide the nodes of the power maintenance knowledge domain mapping map into multiple connected communities according to semantic or structural features by using the community detection algorithm Leiden through node movement, partition refinement, and network aggregation. Each community is a node group. The community corresponds to a semantic unit in the power maintenance knowledge.

[0125] For each edge in the community, the source node and the target node are sorted in descending order according to the degree of combination. The source node, target node, edge itself, and related statement descriptions are added to the context window of the LLM in order of priority until the Token limit of the LLM context window is reached. A multi-level report form community summary is generated for each community. Based on the community and community summary, a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation, and important statements is generated based on the set prompt words.

[0126] The power maintenance decision answer result acquisition module is specifically configured to treat each community as a semantically consistent local subgraph and use the community summary corresponding to each subgraph as the basis for subsequent question and answer generation. When a user initiates a power maintenance query request, the system first analyzes the query request and extracts multiple target entities from the user's query request. Under the guidance of the power maintenance knowledge domain mapping map structure, the adjacent nodes, key relationships, and attribute information of the target entities are extracted and associated with the corresponding original document fragments to construct a context window that integrates structured graphs and unstructured corpora. Multiple community summaries corresponding to the query request are obtained.

[0127] The user query request and the corresponding multiple community summaries are vectorized and encoded using the Nomic embedding model. The semantic similarity between the user query request vector and each community summary vector is calculated using the cosine similarity algorithm. The most relevant community summary is selected based on the set similarity threshold according to the semantic similarity.

[0128] The current problem of the user query request is combined with the historical dialogue context, the most relevant community summary screened out is taken as a prompt prefix, the generation semantic space of the LLM is limited, and the power maintenance decision answer result corresponding to the power maintenance query request is generated by using the LLM.

[0129] The specific process of generating the power maintenance decision based on the knowledge domain mapping map retrieval enhancement by using the device of the embodiment of the present application is similar to the foregoing method embodiment, and details are not repeated here.

[0130] In summary, the embodiment of the present application constructs an intelligent question and answer system in the power maintenance field by fusing the knowledge domain mapping map retrieval enhancement generation and the LLM, significantly improves the efficiency and accuracy of knowledge acquisition and decision support. The automatic knowledge extraction and graph construction of unstructured maintenance text are realized, the problem of insufficient analysis of complex hierarchical logic and industry standard correlation in the traditional method is effectively solved through structured semantic modeling; the ability of cross-document reasoning and dynamic knowledge updating is enhanced by using the multi-granularity retrieval mechanism combined with the community summary generation technology, the dependence on artificial experience and the risk of information lag are greatly reduced. Compared with the prior art, the system improves the efficiency of knowledge domain mapping map construction while improving the question and answer accuracy, and significantly suppresses the model "illusion" phenomenon, provides high-reliability and interpretable decision support for power equipment fault diagnosis, maintenance plan optimization and risk warning, effectively reduces the operation and maintenance cost and improves the power supply reliability.

[0131] Those skilled in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or flows in the drawings are not necessarily necessary for implementing the present application.

[0132] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and the necessary general hardware platform. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the method described in each embodiment or some parts of the embodiments of the present application.

[0133] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0134] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for retrieving enhanced generated power maintenance decisions based on a knowledge domain mapping map, characterized by, The application comprises the following steps: A large language model (LLM) is guided to conduct deep semantic analysis on power maintenance files by a prompt engineering to construct a power maintenance knowledge domain mapping map; A community detection algorithm Leiden is used to generate hierarchical communities based on the power maintenance knowledge domain mapping map, and a community summary in the form of a report is created for each community; A user-initiated power maintenance query request is mapped to an embedding space of the power maintenance knowledge domain mapping map through vectorization technology, semantic matching with the community summary is realized, and a power maintenance decision answer result corresponding to the power maintenance query request is output.

2. The method of claim 1, wherein, The LLM is guided to conduct deep semantic analysis on the power maintenance files by the prompt engineering to construct the power maintenance knowledge domain mapping map, comprising the following steps: Power maintenance-related source files are collected, read and parsed, and saved as UTF-8 encoded txt files, the source files including unstructured text power maintenance reports and operation logs, the txt files are text segmented to decompose long documents into slices, the LLM is guided to conduct deep semantic analysis on the slices by a prompt engineering using few-shot learning, step-by-step reasoning, parameterized templates, entity disambiguation and domain adaptation, key maintenance entities and the association between the maintenance entities are extracted, and the power maintenance knowledge domain mapping map is constructed.

3. The method of claim 2, wherein, The community detection algorithm Leiden is used to generate hierarchical communities based on the power maintenance knowledge domain mapping map, and a community summary in the form of a report is created for each community, comprising the following steps: The community detection algorithm Leiden is used to divide nodes of the power maintenance knowledge domain mapping map into multiple connected communities according to semantic or structural features through node movement, partition refinement and network aggregation, each community is a node group, and the community corresponds to a semantic unit in the power maintenance knowledge; Each edge in the community is sorted in descending order according to the combination degree of the source node and the target node, the source node, the target node, the edge itself and the related declaration description are sequentially added to a context window of the LLM according to the priority until the Token limit of the context window of the LLM is reached, a multi-level community summary in the form of a report is generated for each community, and a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation and important declarations is generated based on the community and the community summary according to a set prompt word.

4. The method of claim 3, wherein, The user-initiated power maintenance query request is mapped to the embedding space of the power maintenance knowledge domain mapping map through the vectorization technology, semantic matching with the community summary is realized, and the power maintenance decision answer result corresponding to the power maintenance query request is output, comprising the following steps: Each community is regarded as a semantically consistent local subgraph, and each subgraph corresponding community summary is taken as the basis information source for subsequent question and answer generation. When a user initiates a power maintenance query request, the system first analyzes the query request, extracts multiple target entities from the user's query request, extracts the adjacent nodes, key relationships and attribute information of the target entities under the guidance of the power maintenance knowledge domain mapping map structure, and associates the corresponding original document fragments to construct a context window integrating structured graph and unstructured corpus, and obtains multiple community summaries corresponding to the query request. The Nomic embedding model is used to vectorize the text of the user query request and the corresponding multiple community summaries, the semantic similarity between the user query request vector and each community summary vector is calculated by using the cosine similarity algorithm, and the most relevant community summary is selected based on the set similarity threshold according to the semantic similarity. The current problem and historical dialogue context of the user query request are combined, the selected most relevant community summary is taken as a prompt prefix, the generation semantic space of the LLM is limited, and the power maintenance decision answer result corresponding to the power maintenance query request is generated by using the LLM.

5. The method of claim 4, wherein, The method further comprises: visualizing and interpreting the power maintenance knowledge domain mapping map according to the community report by using the ForceAtlas2 algorithm, and the visualizing and interpreting result comprises: a topological structure, community clustering and a dynamic interactive view.

6. An apparatus for retrieving generated power maintenance decision based on knowledge domain mapping map, characterized in that, It comprises: a power maintenance knowledge domain mapping map construction module, a community and community summary generation module and a power maintenance decision answer result acquisition module. The power maintenance knowledge domain mapping map construction module is used to guide the large language model (LLM) to perform deep semantic analysis on the power maintenance file by using a prompt engineering, and construct a power maintenance knowledge domain mapping map. The community and community summary generation module is used to generate hierarchical communities by using a community detection algorithm Leiden based on the power maintenance knowledge domain mapping map, and create a community summary in the form of a report for each community. The power maintenance decision answer result acquisition module is used to map the power maintenance query request initiated by the user to the embedding space of the power maintenance knowledge domain mapping map by using vectorization technology, realize semantic matching with the community summary, and output the power maintenance decision answer result corresponding to the power maintenance query request.

7. The device of claim 6, wherein: The power maintenance knowledge domain mapping map construction module is specifically used to collect source files related to power maintenance, read and analyze the source files, save them as UTF-8 encoded txt files, the source files include unstructured text power maintenance reports and operation logs, perform text segmentation on the txt files, decompose long documents into slices, guide the LLM to perform deep semantic analysis on the slices by using a few-shot learning, step-by-step reasoning, parameterized template, entity disambiguation and domain adaptation, extract key maintenance entities and the association relationships between the maintenance entities, and construct a power maintenance knowledge domain mapping map.

8. The device of claim 7, wherein: The community and community summary generation module is specifically configured to divide nodes of the power maintenance knowledge domain mapping map into a plurality of connected communities according to semantic or structural features by node movement, partition refinement and network aggregation through a community detection algorithm Leiden, each community being a node group, and the community corresponding to a semantic unit in the power maintenance knowledge; For each edge in the community, the source node and the target node are sorted in descending order according to the degree of combination, and the source node, the target node, the edge itself and the description of the related statement are added to the context window of the LLM in turn according to the priority until the Token limit of the context window of the LLM is reached, and a multi-level report form community summary is generated for each community, and a comprehensive report containing community key entities, legal compliance, technical capabilities, reputation and important statements is generated based on the community and the community summary according to a set prompt word.

9. The apparatus of claim 8, wherein: The power maintenance decision answer result acquisition module regards each community as a locally consistent subgraph, and uses the community summary corresponding to each subgraph as a basic information source for subsequent question and answer generation. When a user initiates a query request for power maintenance, the system first analyzes the query request, extracts a plurality of target entities from the query request, extracts the adjacent nodes, key relationships and attribute information of the target entities under the guidance of the power maintenance knowledge domain mapping map structure, and associates the corresponding original document fragments to construct a context window that integrates structured graphs and unstructured corpus, and obtains a plurality of community summaries corresponding to the query request; The Nomic embedding model is used to vectorize and encode the text of the user query request and the plurality of community summaries, the cosine similarity algorithm is used to calculate the semantic similarity between the user query request vector and each community summary vector, and the most relevant community summary is selected based on a set similarity threshold according to the semantic similarity; The current problem and historical dialogue context of the user query request are combined, the selected most relevant community summary is used as a prompt prefix, the generation semantic space of the LLM is limited, and the power maintenance decision answer result corresponding to the power maintenance query request is generated by using the LLM.

10. The apparatus of claim 7, wherein, The apparatus further comprises: A visual interpretation module is configured to visually interpret the power maintenance knowledge domain mapping map according to the community report by using the ForceAtlas2 algorithm, and the visual interpretation result includes a topological structure, community clustering and a dynamic interactive view.

Citation Information

Cited By

  • Device operation and maintenance intelligent question answering method and system based on knowledge graph context fusion

    CN122019733A

  • Intelligent question and answer method and system for equipment operation and maintenance based on context fusion on knowledge graph

    CN122019733B