Power industry dynamic knowledge base construction method and system based on large language model

Abstract

The invention discloses a power industry dynamic knowledge base construction method and system based on a large language model, and relates to the field of electric digital data processing, and the method comprises the steps: receiving a power field document imported by a user, and converting the power field document into an electronic document; calling a large language model to analyze the electronic document, extracting a terminology definition, chart position information and a document logic level in the electronic document, and generating a knowledge unit containing a plurality of contexts; connecting the knowledge units into directed knowledge links, calculating semantic similarity between adjacent knowledge units in the directed knowledge links, and segmenting the directed knowledge links to obtain a plurality of knowledge topics; receiving a knowledge retrieval instruction, and determining a target knowledge theme corresponding to the knowledge retrieval instruction; and extracting a corresponding target knowledge unit in the target knowledge theme, and generating a knowledge retrieval result based on a target context of the target knowledge unit. By implementing the method, knowledge base construction can be optimized, and the integrity of knowledge base data is ensured.

Description

Technical Field

[0001] The present application relates to the field of electrical digital data processing, and in particular to a method and system for constructing a dynamic knowledge base for the electric power industry based on a large language model. Background Art

[0002] With the deepening advancement of information technology in the power industry, power companies have accumulated a large amount of diverse documentation, including equipment parameters, technical standards, and operation and maintenance manuals. These documents contain important professional knowledge and are of great value in ensuring the safe and stable operation of power systems and improving operation and maintenance efficiency. Establishing a comprehensive power industry knowledge base to achieve efficient document management and knowledge reuse has become a key requirement for industry development.

[0003] In related technologies, the power industry's knowledge base is constructed using intelligent document processing systems. These systems employ deep learning models for document parsing, recognizing content in various formats, including text, tables, and graphics. They utilize graph databases to store relationships between documents and establish a multi-dimensional knowledge index. Regarding knowledge organization, the system uses natural language processing techniques to perform topic clustering and label extraction on document content, enabling automatic document classification. During retrieval, the system combines vector retrieval and keyword matching mechanisms, filtering relevant documents through semantic similarity calculations. Knowledge updates are semi-automated, with the system automatically extracting key information from new documents and updating the knowledge base after expert review.

[0004] However, because professional documents in the power sector often contain a large number of intertwined diagrams, specialized terminology, and multi-level logical relationships, the system can easily break context when parsing documents. For example, when processing multi-page tables or multi-condition operating procedures in equipment manuals, the system can only store them as independent fragments, making it difficult to restore the complete knowledge context during retrieval. Summary of the Invention

[0005] This application provides a method and system for constructing a dynamic knowledge base for the power industry based on a large language model, which is used to optimize the knowledge base construction and ensure the integrity of the knowledge base data.

[0006] In the first aspect, the present application provides a method for constructing a dynamic knowledge base for the electric power industry based on a large language model, which is applied to a data processing system. The method includes: receiving electric power field documents imported by users, converting the electric power field documents into electronic documents containing a unique document identifier; calling a large language model to parse the electronic documents, extracting professional term definitions, chart location information and document logical hierarchy in the electronic documents, and generating knowledge units containing multiple contexts; connecting knowledge units into directed knowledge links, and calculating the semantic similarity between adjacent knowledge units in the directed knowledge links; segmenting the directed knowledge links based on semantic similarity to obtain multiple knowledge topics; receiving knowledge retrieval instructions, and determining the target knowledge topic corresponding to the knowledge retrieval instructions; extracting the corresponding target knowledge units in the target knowledge topic, and generating knowledge retrieval results based on the target context of the target knowledge units.

[0007] In the above embodiment, the data processing system is able to parse documents in the power field, extract professional term definitions, chart locations and document logical levels, and generate knowledge units containing complete context; by constructing directed knowledge links and calculating semantic similarity, accurate association and topic clustering of knowledge units are achieved; during retrieval, the system can generate structured retrieval results based on the context of the knowledge units, effectively ensuring the integrity of the knowledge base data and the accuracy of the retrieval.

[0008] In combination with some embodiments of the first aspect, in some embodiments, the step of receiving the power field document imported by the user and converting the power field document into an electronic document containing a document unique identifier specifically includes: receiving the power field document imported by the user, extracting the document attribute information of the power field document, and obtaining the document type, creation time and version number; selecting a corresponding layout analysis model based on the document type, identifying the layout structure of the power field document including text paragraphs, tables, and graphics, and obtaining a document structure tree; generating a document unique identifier based on the document structure tree and the version number; and encapsulating the layout structure and the document unique identifier into an electronic document in a preset standard format.

[0009] In the above embodiment, the data processing system performs in-depth layout analysis on the imported documents, extracts document attribute information and identifies complex layout structures such as text, tables, and graphics; by generating unique identifiers and standard format conversion, the uniqueness and structured storage of documents in the knowledge base are ensured.

[0010] In combination with some embodiments of the first aspect, in some embodiments, before calling a large language model to parse an electronic document, extracting professional term definitions, chart location information, and document logical hierarchy in the electronic document, and generating a knowledge unit containing multiple contexts, the method also includes: constructing a training sample set based on electric power field documents; the training sample set includes professional term definitions, equipment operation steps, and fault handling processes; extracting professional terminology from electric power field documents and constructing an electric power field vocabulary; adding the electric power field vocabulary to a pre-trained language model, and using the training sample set to perform incremental training on the pre-trained language model to obtain a large language model.

[0011] In the above embodiment, the data processing system constructs a professional training sample set based on documents in the power field, which includes content such as term definitions, operation steps and fault handling processes; through incremental training of the pre-trained language model, the model is equipped with the ability to understand power professional knowledge, thereby improving the accuracy of document parsing and knowledge extraction.

[0012] In combination with some embodiments of the first aspect, in some embodiments, the step of extracting the corresponding target knowledge unit in the target knowledge subject and generating knowledge retrieval results based on the target context of the target knowledge unit specifically includes: extracting the corresponding target knowledge unit in the target knowledge subject, obtaining the document location information of the target knowledge unit, and constructing the target context for knowledge retrieval; according to the target context, extracting the graphic content and table content of the target location to generate a multimodal knowledge sequence; based on the document logical hierarchy of the target knowledge unit, arranging the multimodal knowledge sequence hierarchically to obtain a structured knowledge body; and generating knowledge retrieval results according to the hierarchical relationship of the structured knowledge body.

[0013] In the above embodiment, the data processing system organizes the retrieval results into multimodal knowledge, extracts graphic content and table content and generates a structured description; arranges the knowledge sequence in layers based on the document logical hierarchy, constructs a knowledge system with a clear hierarchy, and realizes the structured presentation of the retrieval results.

[0014] In combination with some embodiments of the first aspect, in some embodiments, the step of extracting the graphic content and table content of the target position according to the target context and generating a multimodal knowledge sequence specifically includes: locating and extracting the graphic element features of the target graphic and the cell layout features of the target table according to the chart position information in the target context; identifying the graphic components and connection relationships in the target graphic, and converting the graphic components and connection relationships into structured graphic descriptions; identifying the header hierarchy and data area of ​​the target table, extracting the association relationship between cells, and converting it into a structured table description; combining the structured graphic description and the structured table description with the corresponding text description blocks to generate a multimodal knowledge sequence.

[0015] In the above embodiment, the data processing system accurately identifies the component features and association relationships in the chart, converts complex graphic and tabular information into structured descriptions; through intelligent combination with text descriptions, it generates a complete multimodal knowledge sequence, ensuring the semantic integrity of the graphic and text information.

[0016] In combination with some embodiments of the first aspect, in some embodiments, after the steps of connecting knowledge units into directed knowledge links and calculating the semantic similarity between adjacent knowledge units in the directed knowledge links, the method also includes: calculating the out-degree value and in-degree value of each knowledge unit in the directed knowledge link to obtain the node weight; based on the node weight, constructing the transition probability matrix of the directed knowledge link, and calculating the page walk score; identifying the critical path in the directed knowledge link based on the node weight and the page walk score; based on the critical path, deleting redundant links whose node weight is lower than the preset weight threshold, and supplementing the missing association relationships to optimize the directed knowledge link.

[0017] In the above embodiment, the data processing system calculates the in-degree values ​​and weights of nodes in the knowledge link, constructs a transition probability matrix for importance analysis, and improves the connectivity and retrieval efficiency of the knowledge network by identifying critical paths and optimizing redundant links.

[0018] In combination with some embodiments of the first aspect, in some embodiments, the step of calculating the out-degree value and in-degree value of each knowledge unit in the directed knowledge link to obtain the node weight specifically includes: traversing each knowledge unit in the directed knowledge link, counting the number of edges pointing to other knowledge units, and obtaining the out-degree value; traversing each knowledge unit in the directed knowledge link, counting the number of edges from other knowledge units, and obtaining the in-degree value; performing weighted calculation based on the out-degree value and the in-degree value to obtain the node weight.

[0019] In the above embodiment, the data processing system calculates the node weight index by statistically analyzing the connection relationship between knowledge units; accurately evaluates the importance of knowledge units through weighted calculation of out-degree values ​​and in-degree values, providing a basis for knowledge network optimization.

[0020] In a second aspect, an embodiment of the present application provides a data processing system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the data processing system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when the computer program product is run on a data processing system, enables the data processing system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions, which, when executed on a data processing system, causes the data processing system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0023] It is understandable that the data processing system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of the present application. Therefore, the beneficial effects that can be achieved can be referenced to the beneficial effects of the corresponding methods and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By adopting a large language model-based power document parsing and knowledge link construction method, the system can intelligently identify professional terms, diagrams, and logical hierarchies in documents and organize them into knowledge units with complete context. By calculating the semantic similarity between knowledge units and constructing directed links, accurate knowledge association is achieved. This effectively solves the problems of document parsing fragmented context and inaccurate knowledge association in related technologies, thereby ensuring the integrity of knowledge base data and the semantic coherence of retrieval results.

[0025] 2. Due to the adoption of a multimodal knowledge extraction and organization method based on document location information, the system can accurately locate and extract the graphics and table content of the target location, convert it into a structured description and intelligently combine it with the text description; by hierarchically arranging the multimodal knowledge sequence, a clearly structured knowledge system is constructed; it effectively solves the problems of fragmentation of graphic and text information and fuzzy knowledge levels in related technologies, and thus realizes the multimodal fusion and structured presentation of retrieval results.

[0026] 3. Due to the adoption of a knowledge link optimization method based on node weight and page movement, the system can accurately evaluate the importance of knowledge units by calculating the node's in-degree value and transition probability matrix; by identifying critical paths and optimizing redundant links, the quality of the knowledge network is improved; and the problems of redundant knowledge links and inaccurate associations in related technologies are effectively solved, thereby achieving efficient connectivity and accurate retrieval of the knowledge network. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1This is a flow chart of a method for constructing a dynamic knowledge base for the power industry based on a large language model in an embodiment of the present application; Figure 2 This is another flowchart of the method for constructing a dynamic knowledge base for the power industry based on a large language model in an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of a physical device of the data processing system in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations of one or more of the listed items.

[0029] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] The operations and maintenance department of a power company needs to manage a large volume of power equipment documentation, including equipment manuals, operating procedures, and troubleshooting manuals. These documents come in a variety of formats and contain numerous specialized terms, diagrams, and operational procedures. When troubleshooting equipment failures, maintenance personnel need to quickly retrieve relevant knowledge. However, due to the fragmented content, inconsistent formats, and complex relationships between specialized knowledge, timely access to required information is often difficult, hindering troubleshooting efficiency. New employees, in particular, face challenges quickly grasping the knowledge structure of the vast document library, hindering work efficiency.

[0032] Related technologies can be used to store, manage, and retrieve documents in the power sector through keyword matching and simple text classification. However, this approach primarily relies on surface-level features of documents and fails to deeply understand their specialized content and knowledge relationships. The following describes a scenario using a related approach to building a dynamic knowledge base for the power industry based on a large language model.

[0033] Existing document management systems primarily rely on keyword matching for retrieval and simple text classification methods for document organization. For example, the document system used by a substation categorizes and stores equipment manuals by device type and document type, allowing retrieval based on title and keywords. When maintenance personnel need to query the protection setting plan for a specific circuit breaker model, the system can only return relevant documents based on keywords. It cannot understand the context of professional terms or connect related knowledge scattered across different documents, forcing users to manually browse through multiple documents to obtain complete information.

[0034] The method for constructing a dynamic knowledge base for the electric power industry based on a large language model, as described in the embodiments of this application, intelligently analyzes document content, constructs directed knowledge links, and calculates semantic similarity. This achieves deep knowledge association and precise retrieval, improving retrieval efficiency while ensuring the integrity and accuracy of knowledge transfer. The following describes scenarios in which the method for constructing a dynamic knowledge base for the electric power industry based on a large language model, as described in this application, is used.

[0035] After adopting this application solution, the document management system can intelligently parse the professional content of power documents. When a new equipment manual is uploaded, the system automatically identifies the document structure, extracts professional terms, diagrams and operating steps, and establishes knowledge associations. For example, when processing the setting plan of a certain type of relay protection device, the system can not only locate the specific parameter setting instructions, but also automatically associate the relevant principle descriptions, wiring diagrams and debugging steps to form a complete knowledge chain. Maintenance personnel can directly obtain structured knowledge content through retrieval and quickly understand the working principles and operating requirements of the equipment.

[0036] It can be seen that the method for constructing a dynamic knowledge base for the power industry based on a large language model in the embodiment of the present application can not only realize the dynamic update of the knowledge network, but also effectively solve the problems of knowledge fragmentation and inaccurate association, thereby realizing the systematic organization and intelligent service of knowledge in the power field.

[0037] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of a method for constructing a dynamic knowledge base for the power industry based on a large language model in an embodiment of the present application.

[0038] S101: Receive a power field document imported by a user, and convert the power field document into an electronic document containing a unique document identifier.

[0039] Among them, power field documents refer to technical documents related to the power system, including but not limited to equipment manuals, technical specifications, operating manuals, troubleshooting guides, etc.; document unique identifier refers to the string code used to uniquely distinguish different documents, which is generated through document attribute information and content characteristics; electronic documents are used to represent standardized digital format documents, which contain structured text content and layout information.

[0040] The data processing system executes this step when a user imports a power sector document through the document upload interface. Specifically, the system first reads the binary data stream of the imported document and identifies the document format type. It then extracts the document's metadata, including attributes such as creation time, author, and version number. It then uses the corresponding layout analysis model to process the document content, identifying layout elements such as text paragraphs, tables, and graphics, and constructing a document structure tree. Finally, it generates a unique identifier based on the document's attributes and structural features and encapsulates the parsed results into a standard electronic document format.

[0041] In some embodiments, document format conversion and unique identifier generation can be achieved in a variety of ways: Optionally, the data processing system can adopt a rule-based approach, first using regular expressions to extract key information such as document titles and chapters, then combining the document creation time and version number to generate a hash value as a unique identifier, and finally converting the content into Markdown format; Optionally, the data processing system can adopt a deep learning approach, using a pre-trained document understanding model to identify document structure and semantic information, generate a unique identifier based on the document feature vector, and output a structured representation in JSON format. It is understandable that other document parsing and identifier generation methods can also be used, which are not limited here.

[0042] In actual applications, you may encounter parsing errors caused by non-standard document formats. To address this, the data processing system first preprocesses and normalizes the document, including correcting non-standard typesetting, unifying character encoding, and handling special symbols. At the same time, the system maintains a library of format conversion rules and develops corresponding handling strategies for different types of exceptions. For example, for tables that span multiple pages, the system analyzes the correspondence between table borders and cells and merges them into a complete data structure. For non-standard chapter hierarchies, the system reconstructs the document hierarchy based on text formatting and indentation relationships.

[0043] S102: Invoke a large language model to parse the electronic document, extract professional term definitions, chart location information, and document logic hierarchy in the electronic document, and generate a knowledge unit containing multiple contexts.

[0044] Among them, the large language model refers to a pre-trained language model trained with knowledge in the power field, which is used to understand the content of professional documents; professional term definitions refer to the proper nouns in the power field and their explanations; chart position information is used to represent the spatial layout and reference relationships of graphics and tables in the document; the document logical hierarchy represents the hierarchical organizational structure of the content; and the knowledge unit refers to the smallest knowledge fragment with complete semantics, containing contextual information.

[0045] The data processing system performs this step after completing the electronic document conversion. Specifically, the data processing system first segments the electronic document into sentence-level text segments, while retaining structural information such as paragraphs and chapters. It then uses a domain-trained large language model to identify specialized terms in the text and extract their definitions. It then locates graphic elements in the document and analyzes their reference relationships with surrounding text. It then constructs a logical hierarchy of content based on the document structure. Finally, it combines the associated text segments, term definitions, and graphic references into knowledge units, recording their contextual information within the document.

[0046] In some embodiments, intelligent parsing of professional content can be achieved through a variety of methods: Optionally, the data processing system can employ a rule-based and template-based approach, using a predefined terminology dictionary and text pattern matching rules to identify professional content, determine the boundaries of term definitions through syntactic analysis, and construct a logical hierarchy based on document titles and section numbers. Optionally, the data processing system can employ a deep learning approach, using a sequence labeling model to identify professional terms and definitions, modeling the document's hierarchical structure through a graph neural network, and combining an attention mechanism to capture the relationship between diagrams and text. It is understood that other content understanding and knowledge extraction methods can also be employed, and are not limited here.

[0047] In practical applications, ambiguous professional terminology can lead to parsing errors. To address this, the data processing system maintains a domain terminology knowledge base, recording the various expressions and contexts of the term. When encountering an ambiguous term, the system analyzes the context of the sentence in which it appears, including co-occurring words, syntactic structure, and thematic background, and resolves the ambiguity through multi-feature fusion. For example, for the term "voltage," the system determines from the context whether it refers to an electrical parameter or the operating status of the device, and extracts the corresponding definition.

[0048] S103 , connecting the knowledge units into directed knowledge links, and calculating the semantic similarity between adjacent knowledge units in the directed knowledge links.

[0049] Among them, directed knowledge links represent directional association networks between knowledge units, which are used to describe the transmission and dependency of knowledge; semantic similarity refers to the degree of closeness in meaning between two knowledge units, which is calculated through distance measurement in vector space; adjacent knowledge units represent pairs of nodes directly connected in the knowledge link; knowledge associations are used to represent the logical relationships between different knowledge units, including sequence, inclusion relationship, reference relationship, etc.

[0050] The data processing system performs this step after extracting knowledge units. Specifically, it first analyzes explicit associations between knowledge units, including cross-references and term definition references within documents. It then establishes hierarchical relationships between knowledge units based on the document's logical hierarchy. It then uses a pre-trained language model to convert knowledge units into semantic vector representations. It then calculates the cosine similarity between adjacent knowledge unit vectors and sets a similarity threshold to filter valid connections. Finally, it constructs a directed knowledge link network based on explicit associations and semantic similarity.

[0051] In some embodiments, knowledge link construction can be achieved in a variety of ways: optionally, the data processing system can adopt a rule-based approach, using document structure and citation relationships to establish initial links, calculate the strength of association between nodes using traditional text similarity algorithms such as TF-IDF, and optimize the link structure based on heuristic rules; optionally, the data processing system can adopt a deep learning approach, using a bidirectional encoder to obtain the semantic representation of knowledge units, calculating the relevance score between nodes through an attention mechanism, and combining graph neural networks to learn the neighborhood characteristics of nodes. It is understandable that other knowledge association and similarity calculation methods can also be used, which are not limited here.

[0052] In actual applications, inaccurate semantic similarity calculations may lead to incorrect connections. To address this, the data processing system combines multiple similarity measurement methods, including vocabulary overlap, topic similarity, and structural similarity. The system first calculates similarity indicators for different dimensions and then uses weighted fusion to obtain a comprehensive score. At the same time, the system dynamically adjusts the similarity calculation strategy by considering the domain characteristics of the knowledge unit, such as the density of professional terminology and the type of technical parameters. For example, for device parameter descriptions, the system will place greater emphasis on matching numerical features; for operating procedure instructions, it will focus on the consistency of the step sequence.

[0053] S104: Segment the directed knowledge link based on semantic similarity to obtain multiple knowledge topics.

[0054] Among them, segmentation refers to dividing a group of closely related knowledge units into independent sub-graph structures; knowledge topics represent a collection of knowledge units with similar semantic content or common thematic features; semantic clustering is used to represent the automatic grouping process of knowledge units based on semantic similarity; and topic boundaries refer to the dividing lines between different knowledge topics, which are determined by the semantic similarity threshold.

[0055] The data processing system performs this step after constructing directed knowledge links. Specifically, the system first constructs a similarity matrix for knowledge units based on the calculated semantic similarity. It then uses a community discovery algorithm to detect closely connected subgroups within the knowledge links. It then analyzes the topic characteristics of each subgroup, including key term distribution, core concepts, and technical fields. It then optimizes topic boundaries and processes transitional knowledge units across topics. Finally, it generates descriptive tags for each knowledge topic and establishes relationships between topics.

[0056] In some embodiments, knowledge topic partitioning can be achieved in a variety of ways: optionally, the data processing system can adopt a density-based clustering method, using the DBSCAN algorithm to identify dense areas of knowledge units based on semantic similarity, forming knowledge topics through neighborhood expansion, and determining topic boundaries based on connectivity analysis; optionally, the data processing system can adopt a graph-based segmentation method, using a spectral clustering algorithm to convert knowledge links into Laplace matrices, obtaining an optimal segmentation scheme through eigenvalue decomposition, and optimizing topic partitioning in combination with modularity evaluation. It is understandable that other topic discovery and boundary optimization methods can also be used, which are not limited here.

[0057] In practical applications, the problem of inaccurate knowledge topic division may arise due to fuzzy boundaries. To address this, the data processing system adopts a multi-level topic division strategy. The system first identifies the core topic area at a high semantic similarity threshold, then gradually lowers the threshold to discover transition areas. For knowledge units located at the topic boundary, the system analyzes their association strength with multiple topics, allowing them to belong to multiple related topics simultaneously. For example, a knowledge unit describing the protection function of a device may be associated with both the device structure and the protection principle. The system retains this multi-topic attribute to reflect the integrity of the knowledge.

[0058] S105: Receive a knowledge retrieval instruction, and determine a target knowledge topic corresponding to the knowledge retrieval instruction.

[0059] Among them, the knowledge retrieval instruction represents the query request input by the user, which includes retrieval keywords and query conditions; the target knowledge topic refers to the set of knowledge units that are most relevant to the semantics of the retrieval instruction; the topic matching degree is used to indicate the degree of relevance between the retrieval instruction and the knowledge topic; the query intent represents the actual needs behind the user's retrieval request, which is identified through semantic understanding.

[0060] The data processing system executes this step upon receiving a user's search request. Specifically, the system first semantically understands the search instruction, identifying the core concepts and query intent. It then converts the search instruction into a vector representation and calculates the semantic similarity with each knowledge topic. It then sorts the knowledge topics based on topic matching, selecting the most similar topics. It then analyzes the relationships between topics and expands the scope of related topics. Finally, it integrates multiple features to determine the final set of target knowledge topics.

[0061] In some embodiments, retrieval intent understanding and topic matching can be achieved in a variety of ways: Optionally, the data processing system can adopt a rule-based approach, using keyword matching and syntactic analysis to extract query intent, select target topics through predefined topic mapping rules, and combine multiple matching conditions based on Boolean logic; Optionally, the data processing system can adopt a deep learning approach, using an intent classification model to understand retrieval requirements, calculating the relevance score between the query and the topic through a semantic retrieval model, and combining a multi-head attention mechanism to capture multiple aspects of the query. It is understandable that other retrieval understanding and topic matching methods can also be used, which are not limited here.

[0062] In practical applications, ambiguous search queries can lead to inaccurate topic matching. To address this, the data processing system implements an interactive query understanding strategy. When it detects ambiguous query intent, the system generates clarifying questions to guide the user in providing more information. Furthermore, the system maintains the user's search history and leverages contextual information to aid in understanding intent. For example, when a user queries "protection device parameter settings," the system determines the specific device type based on previous search records or prompts the user to select a specific protection type, thereby accurately locating the target knowledge topic.

[0063] S106: Extract the corresponding target knowledge unit in the target knowledge subject, and generate a knowledge retrieval result based on the target context of the target knowledge unit.

[0064] Among them, the target knowledge unit represents the knowledge fragment directly related to the retrieval requirements; the target context refers to the contextual information of the knowledge unit in the original document, including adjacent paragraphs, related charts, etc.; the knowledge retrieval result is used to represent the structured knowledge content returned by the system; the context association represents the semantic connection relationship between the knowledge unit and its surrounding content.

[0065] The data processing system performs this step after determining the target knowledge topic. Specifically, the data processing system first selects the knowledge units most relevant to the search intent from the target knowledge topic; then, based on the document location information of the knowledge units, it extracts their original context; then, it analyzes the logical relationships between knowledge units and reconstructs the organizational structure of the knowledge content; then, it fuses text descriptions, graphic content, and related information to generate a multimodal knowledge representation; finally, it adjusts the content presentation format and level of detail based on the user's search intent to form the final search results.

[0066] In some embodiments, knowledge retrieval results can be generated in a variety of ways: Alternatively, the data processing system can employ a template-based approach, using a predefined result organization template to sort knowledge units by importance and relevance, assembling contextual information through a rule engine, and generating hierarchical retrieval results; Alternatively, the data processing system can employ a generative approach, using a sequence-to-sequence model to reorganize and summarize knowledge content, integrating graphic and text information through a multimodal fusion network, and dynamically generating retrieval responses based on user intent. It is understood that other result generation and presentation methods may also be employed, and are not limited herein.

[0067] In actual applications, you may encounter the problem of overly lengthy context information, resulting in insufficiently refined search results. To address this situation, the data processing system will adopt an adaptive content screening strategy. The system first evaluates the importance of each piece of contextual content to understanding the target knowledge, and identifies key information through saliency analysis. For chart content, the system will extract local areas related to the search target to avoid completely introducing irrelevant content. For example, when retrieving the parameter settings of a certain protection function, the system will focus on retaining directly related content such as parameter definitions, value ranges, and typical configurations, while appropriately simplifying other functional characteristics of the device, thereby ensuring the accuracy and practicality of the search results.

[0068] Furthermore, with continued use of the system, the knowledge network can be continuously optimized and improved. For example, after repeatedly handling a certain type of fault, the system can supplement knowledge associations based on actual application scenarios, organically organizing information such as fault symptoms, diagnostic steps, and treatment methods. When a new fault case arises, the system automatically analyzes its association with existing knowledge and updates the knowledge network. Furthermore, the system can adaptively adjust the depth and breadth of knowledge presentation based on the user's expertise, providing basic guidance for junior maintenance personnel and in-depth principle analysis for professional technicians, thus enabling personalized customization of knowledge services.

[0069] The following is a more detailed description of the process of the method provided by this implementation. Figure 2 , which is another flow chart of the method for constructing a dynamic knowledge base for the power industry based on a large language model in an embodiment of the present application.

[0070] S201: Receive a power field document imported by a user, and convert the power field document into an electronic document containing a unique document identifier.

[0071] Referring to step S101 , the data processing system converts the power field document into an electronic document.

[0072] In some embodiments, the data processing system will perform attribute extraction and layout analysis on the imported document and convert it into an electronic document in a standard format. That is, the data processing system will receive the power field document imported by the user, extract the document attribute information of the power field document, and obtain the document type, creation time and version number; select the corresponding layout analysis model based on the document type, identify the layout structure of the power field document including text paragraphs, tables, and graphics, and obtain the document structure tree; generate a document unique identifier based on the document structure tree and version number; and encapsulate the layout structure and document unique identifier into an electronic document in a preset standard format.

[0073] Among them, power field documents refer to technical documents containing professional knowledge of the power system, including equipment manuals, regulations and specifications, etc.; document attribute information refers to the metadata set that describes the basic characteristics of the document; the layout analysis model is used to represent the layout recognition model for different types of documents; the document structure tree represents the hierarchical organizational structure of the document content; the layout structure refers to the spatial arrangement relationship of various content elements in the document; the document unique identifier is used to represent the unique code to distinguish different documents.

[0074] The data processing system executes this step when a user uploads a power sector document through the system interface. Specifically, the system first reads the uploaded document's binary stream, identifies the file format, and extracts metadata. It then loads the corresponding layout analysis model based on the document type and performs structured parsing on the document. It then uses a deep learning algorithm to identify text blocks, table areas, and graphic areas within the document, analyzing their positional and hierarchical relationships. The parsed results are then organized into a tree structure and a unique identifier is generated based on the structural features and version information. Finally, the document content is converted to a standard format and packaged into a unified electronic document.

[0075] In some embodiments, document parsing and conversion can be achieved in a variety of ways: Optionally, the data processing system can adopt a deep learning-based approach, first using a target detection model to detect different regional structures in the document page (header, footer, body text, pictures, formulas, etc.), and then using a corresponding recognition method for each region: the body uses OCR technology for text recognition, and the formula area uses a deep learning model such as UniMERNet for formula recognition, and finally all recognition results are fused to generate a document feature vector; Optionally, the data processing system can adopt layout analysis and rules to identify the document structure through a predefined template, extract regional information based on layout analysis, and use the SHA-256 algorithm to generate a document identifier. It is understandable that other document processing methods can also be used, which are not limited here.

[0076] In practical applications, complex document formats may lead to layout recognition errors. To address this, the data processing system employs a multi-model collaborative parsing strategy. First, a document type identifier is constructed to accurately determine the document's specific type and format characteristics. The most appropriate combination of layout analysis models is then selected to perform multi-level parsing of the document. Simultaneously, model performance is continuously optimized through rule verification and manual feedback. For example, for device manuals containing complex charts, the system utilizes both a table recognition model and a graphic analysis model to ensure accurate parsing of all types of content.

[0077] S202: Invoke a large language model to parse the electronic document, extract professional term definitions, chart location information, and document logic hierarchy in the electronic document, and generate a knowledge unit containing multiple contexts.

[0078] Referring to step S102 , the data processing system generates a knowledge unit.

[0079] In some embodiments, the data processing system constructs professional field training data and performs incremental model training, that is, the data processing system constructs a training sample set based on power field documents; the training sample set includes professional term definitions, equipment operation steps and fault handling processes; extracts professional terms from power field documents and constructs a power field vocabulary; adds the power field vocabulary to the pre-trained language model, and uses the training sample set to perform incremental training on the pre-trained language model to obtain a large language model.

[0080] Among them, the training sample set refers to the annotated data set used for model training; the professional term definition refers to the specific terms in the power field and their explanations; the power field vocabulary is used to represent the structured dictionary of power professional terms; the pre-trained language model refers to the basic model pre-trained on general corpus; incremental training refers to the process of domain adaptive training based on the pre-trained model.

[0081] The data processing system performs this step after completing the collection of power documents. Specifically, the system first extracts professional knowledge from the parsed power documents, including concept explanations, operational instructions, and fault cases. It then constructs a high-quality training corpus through text analysis and expert annotation. A large language model is then used to extract and construct a professional vocabulary from the document content, including information such as term definitions, usage scenarios, and associations. This vocabulary is then integrated into the pre-trained model's vocabulary system. Finally, the constructed training samples are used to conduct targeted training on the model, improving its understanding capabilities in the power sector.

[0082] In some embodiments, model training and optimization can be achieved through a variety of methods: Optionally, the data processing system can adopt a knowledge distillation-based approach, first training a large-scale professional domain model, then migrating domain knowledge into a lightweight model, improving model performance through multi-task learning, and regularly updating the model with new corpus; Optionally, the data processing system can adopt a few-shot learning-based approach to construct a domain knowledge prompt template, achieve rapid domain adaptation through prompt learning methods, and continuously optimize model performance in combination with active learning strategies. It is understood that other model training methods can also be used, and are not limited here.

[0083] In practical applications, uneven training sample quality can lead to unstable model performance. To address this, the data processing system employs a data management strategy with tiered quality control. First, a sample evaluation system is established to assess sample quality based on accuracy, completeness, and representativeness. Different training strategies are then applied to samples of varying quality levels, with high-quality samples used for core competency training and other samples for knowledge expansion. Simultaneously, adversarial training is used to enhance model robustness. For example, when processing equipment operating instructions, the system pays special attention to the accuracy of key steps and safety instructions to ensure the model can reliably understand and represent important information.

[0084] S203 : Connect the knowledge units into directed knowledge links, and calculate the semantic similarity between adjacent knowledge units in the directed knowledge links.

[0085] Referring to step S103 , the data processing system constructs a directed knowledge link.

[0086] S204: Calculate the out-degree value and in-degree value of each knowledge unit in the directed knowledge link to obtain the node weight.

[0087] Among them, the out-degree value indicates the number of connections from the knowledge unit to other knowledge units; the in-degree value refers to the number of connections from other knowledge units to the knowledge unit; the node weight is used to indicate the importance of the knowledge unit in the entire knowledge network; and the connection strength indicates the closeness of the association between knowledge units.

[0088] The data processing system performs this step after constructing a directed knowledge link. Specifically, the system first traverses each knowledge unit in the knowledge link and counts its outgoing and incoming connections. It then assigns a weight to each edge based on the semantic similarity of the connections. It then calculates the initial weight of the node by combining the outgoing and incoming degrees and the edge weights. The weights are then adjusted based on the depth and expertise of the knowledge unit. Finally, the final node weight is normalized.

[0089] In some embodiments, node weight calculation can be implemented in a variety of ways: Optionally, the data processing system can use a graph theory-based approach to iteratively calculate node centrality indicators, including closeness centrality, betweenness centrality, etc., and combine the weighted sum of multiple centrality indicators as the node weight; Optionally, the data processing system can use a random walk-based approach to calculate the node's PageRank value as the basic weight, and fine-tune the weight using expert knowledge rules to obtain a comprehensive score reflecting the importance of knowledge. It is understandable that other node importance assessment methods can also be used, which are not limited here.

[0090] In practical applications, imbalanced connections between knowledge units can lead to biased weight calculations. To address this, the data processing system employs a hierarchical weighted calculation strategy. The system first groups knowledge units by professional field and technical level, calculating base weights within each group. Weights are then adjusted across groups, taking into account the importance of each group. For example, for knowledge units related to equipment protection, the system increases their weight based on their criticality to safe operations; for knowledge units related to auxiliary instructions, the system decreases their weight accordingly, ensuring that node weights more accurately reflect the actual importance of the knowledge.

[0091] In some embodiments, the data processing system calculates the importance of nodes by analyzing the connection relationship between knowledge units, that is, the data processing system traverses each knowledge unit in the directed knowledge link, counts the number of edges pointing to other knowledge units, and obtains the out-degree value; traverses each knowledge unit in the directed knowledge link, counts the number of edges from other knowledge units, and obtains the in-degree value; and performs weighted calculation based on the out-degree value and the in-degree value to obtain the node weight.

[0092] Among them, the out-degree value indicates the number of connections from the knowledge unit to other units; the in-degree value refers to the number of connections from other units to the knowledge unit; the number of edges is used to indicate the number of direct associations between knowledge units; the weighted calculation refers to the weight calculation process that takes into account the importance of the connection; the node weight refers to a quantitative indicator of the importance of the knowledge unit in the entire network.

[0093] The data processing system performs this step after constructing directed knowledge links. Specifically, the system first traverses all nodes in the knowledge network, counting the number of connections from each node and recording this as the out-degree value. It then traverses the network again, counting the number of connections pointing to each node and recording this as the in-degree value. It then analyzes the semantic strength of the connections and assigns weight coefficients to different types of associations. It then calculates the initial weight of the node through a weighted summation of the in-degree values ​​and the connection weights. Finally, it performs normalization to obtain the final weight value that reflects the importance of the node.

[0094] In some embodiments, node weight calculation can be implemented in a variety of ways: Optionally, the data processing system can use a network topology-based approach to first calculate the basic centrality indicators of the node, including degree centrality, closeness centrality, and betweenness centrality, then obtain a comprehensive score by fusing multiple indicators, and finally adjust the weight distribution based on expert experience; Optionally, the data processing system can use a random walk-based approach to construct a transition probability matrix, obtain the steady-state distribution probability of the node through iterative calculation, and modify the probability value based on the semantic importance of the connection. It is understandable that other weight calculation methods can also be used, which are not limited here.

[0095] In practical applications, the problem of uneven distribution of node connections leading to biased weight calculations may arise. To address this, the data processing system adopts a hierarchical weighted calculation strategy. First, the network is stratified according to the type and level of the knowledge unit, and the basic weight is calculated within each layer. Then, the hierarchical weight coefficient is set, taking into account the importance differences between layers. Finally, the final result is obtained by combining the intra-layer weights and the hierarchical weights. For example, for an equipment fault diagnosis knowledge network, the system divides the layers according to the fault type, focusing on the weight calculation of nodes related to safety hazards to ensure that key knowledge receives a higher importance score.

[0096] S205: Based on the node weights, a transfer probability matrix of the directed knowledge link is constructed, and the page walk score is calculated.

[0097] Among them, the transition probability matrix represents the numerical representation of the jump probability between knowledge units; the page walk score refers to the node importance index calculated by the random walk algorithm; the jump probability is used to represent the possibility of transition from one knowledge unit to another; the damping factor represents the probability of interrupting the current path at any node and randomly jumping to other nodes.

[0098] The data processing system performs this step after obtaining the node weights. Specifically, the system first constructs an adjacency matrix based on the connections between knowledge units. It then calculates the transition probability between nodes based on the node weights and connection strengths. A damping factor is then introduced to account for the impact of random jumps. The steady-state distribution of the transition probability matrix is ​​then solved using a power iteration method. Finally, the steady-state distribution value is used as the page walk score for the knowledge unit, which is used to assess the node's importance in the global knowledge network.

[0099] In some embodiments, based on the resource consumption issues of knowledge graph technology in practical applications, the data processing system can choose different implementation solutions based on actual conditions: Alternatively, in scenarios with limited computing resources, a simple ranking method based on vector similarity can be used to directly calculate the semantic similarity between knowledge units as the basis for importance scoring; alternatively, when offline processing is possible and resources are sufficient, a complete knowledge graph solution can be used to obtain more accurate scoring results through complex graph calculations. The system will dynamically evaluate available resources and select an appropriate processing strategy to strike a balance between accuracy and efficiency.

[0100] In practical applications, data processing systems can also adopt a hierarchical processing strategy: for frequently accessed core knowledge subgraphs, knowledge graph calculations can be pre-computed and the results cached; for less frequently accessed edge knowledge, a lightweight solution with real-time calculations is employed. Simultaneously, the system continuously monitors resource usage and automatically adjusts processing strategies when necessary, such as downgrading to basic similarity calculations when resources are limited and resuming the full graph calculation process when resources are sufficient. This ensures that high-quality knowledge retrieval services are provided while ensuring system availability.

[0101] S206: Identify the key path in the directed knowledge link based on the node weight and the page walk score.

[0102] Among them, the critical path represents the most important knowledge transfer channel in the knowledge network; the path weight refers to the comprehensive weight score of all nodes on the path; the path relevance is used to represent the semantic coherence between adjacent knowledge units on the path; the critical node sequence represents the combination of core knowledge units that constitute the critical path.

[0103] The data processing system performs this step after obtaining the page walk score. Specifically, the data processing system first performs a weighted fusion of the node weight and the page walk score to obtain a comprehensive node importance score. It then sets a node screening threshold based on the comprehensive score to preliminarily determine a set of key nodes. It then analyzes the connections between key nodes to identify sequences of nodes with high semantic relevance. It then uses a dynamic programming algorithm to calculate the optimal path combination, taking into account path integrity and knowledge coverage. Finally, it optimizes the identified key paths to ensure path connectivity and knowledge transfer consistency.

[0104] In some embodiments, critical path identification can be achieved through a variety of methods: Optionally, the data processing system can use a shortest path-based approach, first assigning a weight based on node importance to each edge, then using the Dijkstra algorithm to calculate the optimal path between important nodes, and constructing a knowledge backbone network by merging paths; Optionally, the data processing system can use a flow maximization-based approach, treating the knowledge network as a flow network, identifying the main channels for knowledge transfer using a maximum flow algorithm, and then screening the final critical path based on node importance. It is understood that other path optimization algorithms can also be used, which are not limited here.

[0105] In practical applications, critical paths may encounter branches or loops, leading to confusion in knowledge transfer. To address this, the data processing system employs path simplification and loop handling strategies. The system first detects branch points and loop structures within the path and assesses their impact on knowledge transfer. For necessary branches, it retains the main paths with high semantic relevance. For loops, it analyzes the hierarchical relationships between nodes within the loop and breaks them while maintaining knowledge integrity. For example, when processing knowledge paths for equipment fault diagnosis, the system organizes the diagnostic process into a linear sequence of steps, with necessary branch points linked through cross-references.

[0106] S207. Based on the critical path, delete redundant links whose node weights are lower than a preset weight threshold, and supplement the missing association relationships to optimize the directed knowledge links.

[0107] Among them, redundant links represent redundant connections that do not contribute much to knowledge transfer; the preset weight threshold refers to the weight judgment standard used to screen effective connections; missing associations represent logical relationships between knowledge units that should exist but have not been identified; optimized knowledge links are used to represent high-quality knowledge networks that have been streamlined and supplemented.

[0108] The data processing system performs this step after identifying the critical path. Specifically, the system first screens the connections in the knowledge chain based on a preset weight threshold to identify low-value, redundant connections. It then analyzes the knowledge transfer patterns along the critical path and summarizes typical association patterns between knowledge units. Based on these association patterns, it searches the knowledge network for pairs of knowledge units that may exist but are not connected. These candidate associations are then verified for semantic similarity. Finally, redundant connections are deleted and new associations that have passed verification are added, achieving dynamic optimization of the knowledge chain.

[0109] In some embodiments, knowledge link optimization can be achieved through a variety of methods: Optionally, the data processing system can adopt a graph structure optimization approach, using a community detection algorithm to identify clustering characteristics of knowledge units, determine redundant links based on the connection density within and outside the cluster, and discover potential associations through transitive closure analysis; Optionally, the data processing system can adopt a knowledge reasoning approach to construct a semantic relationship model between knowledge units, discover implicit associations through rule reasoning and pattern matching, and determine the addition and deletion of connections in combination with confidence assessment. It is understood that other network optimization methods can also be used, which are not limited here.

[0110] In practical applications, improper supplementation of relationships can lead to knowledge contamination. To address this, the data processing system employs a multi-layered relationship verification mechanism. The system first constructs relationship verification rules based on the domain knowledge base, including the subordination of professional terms and the logical dependencies of technical concepts. It then verifies the rules for candidate relationships to ensure they conform to the domain knowledge system. Finally, it manually confirms important relationships through an expert feedback mechanism. For example, when supplementing relationships for equipment maintenance knowledge, the system rigorously verifies the sequential dependencies between maintenance steps to avoid introducing incorrect operational instructions.

[0111] S208 : Segment the directed knowledge link based on semantic similarity to obtain multiple knowledge topics.

[0112] Referring to step S104 , the data processing system determines a knowledge topic.

[0113] S209: Receive a knowledge retrieval instruction, and determine a target knowledge topic corresponding to the knowledge retrieval instruction.

[0114] Referring to step S105 , the data processing system determines the target knowledge topic based on the search.

[0115] S210: Extract the target knowledge unit corresponding to the target knowledge topic, obtain the document location information of the target knowledge unit, and construct a target context for knowledge retrieval.

[0116] Among them, the target knowledge unit represents a knowledge fragment that is highly relevant to the retrieval requirement; the document location information refers to the precise positioning of the knowledge unit in the original document, including page numbers, paragraph numbers, etc.; the target context is used to represent the contextual information of the knowledge unit, including adjacent content and reference relationships; the knowledge association strength represents the degree of relevance between the knowledge unit and the retrieval requirement.

[0117] The data processing system performs this step after determining the target knowledge topic. Specifically, the data processing system first selects the most relevant knowledge units within the target knowledge topic based on the search intent. It then locates the specific location in the original document based on the document's location information. It then extracts the contextual information of the knowledge unit, including preceding and following paragraphs, related diagrams, and citations. It then analyzes the relevance of the contextual content to the search requirements to determine the contextual scope. Finally, it organizes the knowledge unit and its effective context into structured basic retrieval data.

[0118] In some embodiments, target context construction can be achieved through a variety of methods: Optionally, the data processing system can use a syntactic analysis-based approach to identify linguistic dependencies between knowledge units, determine contextual references through coreference resolution, extract key information based on subject-verb relationships, and construct a semantically complete context representation; Optionally, the data processing system can use a topic coherence-based approach to calculate the semantic cohesion between paragraphs, identify the boundaries of topic flow, and divide the context range through hierarchical clustering methods to ensure the coherence of knowledge content. It is understandable that other context analysis methods can also be used, which are not limited here.

[0119] In practical applications, excessively large contexts can lead to increased retrieval noise. To address this, the data processing system employs an adaptive context-trimming strategy. The system first constructs a relevance assessment model based on the search intent, scoring the contextual content on a segment-by-segment basis. It then sets dynamic thresholds to retain highly relevant core content while maintaining necessary logical transitions and key definitions. For example, when searching for parameter settings for a protective device, the system prioritizes directly relevant content, such as parameter definitions and configuration steps, while appropriately simplifying lengthy descriptions of the device's principles.

[0120] S211. Extract the graphic content and table content of the target location according to the target context and generate a multimodal knowledge sequence.

[0121] Among them, graphic content refers to visual information such as schematic diagrams and structure diagrams in the document; table content refers to structured tables containing data and parameters; multimodal knowledge sequences are used to represent complete knowledge expressions that integrate text, graphics, and tables; content relevance represents the semantic correspondence between different modal information.

[0122] The data processing system performs this step after constructing the target context. Specifically, the data processing system first locates the graphics and tables referenced in the target context based on the document's location information. It then analyzes the components of the graphics, including primitives, annotations, and connections. It then parses the structural features of the tables, including the header hierarchy, data areas, and cell attributes. It then identifies the correspondence between text descriptions and chart content. Finally, it organizes the content of each modality according to its semantic association order to generate a structured knowledge sequence.

[0123] In some embodiments, multimodal knowledge sequence generation can be achieved in a variety of ways: Optionally, the data processing system can adopt an image processing-based approach, using a feature extraction algorithm to identify key components in the graphic, analyzing the graphic structure through edge detection and region segmentation, and extracting text information in the graphic in combination with optical character recognition to establish a corresponding relationship between the graphic and the text; Optionally, the data processing system can adopt a deep learning-based approach, using a multimodal pre-trained model to understand the semantics of the graphic, capturing the text-graphic association through a cross-modal attention mechanism, and constructing a coherent knowledge representation based on a sequence generation model. It is understandable that other multimodal content understanding methods can also be used, which are not limited here.

[0124] In practical applications, incomplete information extraction from diagrams can lead to knowledge gaps. To address this, the data processing system employs a hierarchical parsing and information complementation strategy. The system first performs a multi-level analysis of the diagram, extracting information from the overall layout down to the local details. It then uses relevant descriptions in the text description to supplement information that is difficult to directly identify in the diagram. Furthermore, it establishes an associative index for diagram elements. For example, for a complex circuit schematic, the system first identifies the main functional blocks, then analyzes the internal connections, and finally supplements detailed information such as device parameters and operating conditions through text descriptions.

[0125] In some embodiments, the data processing system will perform structured processing on the chart content in the document and integrate it with the text content, that is, the data processing system will locate and extract the graphic features of the target graphic and the cell layout features of the target table based on the chart position information in the target context; identify the graphic components and connection relationships in the target graphic, and convert the graphic components and connection relationships into a structured graphic description; identify the header hierarchy and data area of ​​the target table, extract the association relationship between cells, and convert it into a structured table description; combine the structured graphic description and the structured table description with the corresponding text description blocks to generate a multimodal knowledge sequence.

[0126] Among them, primitive features represent the basic components of graphics and their attributes; cell layout features refer to the spatial distribution and merging relationship of cells in a table; graphic components are used to represent functional structural units in graphics; structured graphic descriptions represent the standardized textual expression of graphic content; structured table descriptions refer to the conversion of table content into a machine-processable data structure; and multimodal knowledge sequences are used to represent comprehensive knowledge representations that integrate text, graphics, and tables.

[0127] The data processing system performs this step after obtaining the target context information. Specifically, the data processing system first accurately locates the target chart based on the location information, extracting the basic graphic elements and the grid structure of the table. It then analyzes the functional modules within the chart, identifying the connections between components and the signal flow. It then parses the table's hierarchical structure to determine the correspondence between the header range and the data area. It then extracts the cell merge information and data type, establishing logical connections between cells. Finally, it integrates the graphic semantics, table data, and related text descriptions into a unified knowledge representation.

[0128] In some embodiments, multimodal content parsing and fusion can be achieved in a variety of ways: Optionally, the data processing system can adopt a method based on a multimodal large language model to directly perform semantic understanding and feature extraction on the image content to achieve deep fusion of graphic and text information; Optionally, the data processing system can adopt a method based on deep learning to establish cross-modal correspondences through an attention mechanism and create structured descriptions based on a semantic generation model; Optionally, the data processing system can adopt a method based on deep learning to use a multimodal pre-trained model to understand the content of the chart, establish cross-modal correspondences through an attention mechanism, and create structured descriptions based on a semantic generation model. It is understandable that other multimodal processing methods can also be used, which are not limited here.

[0129] In practical applications, complex diagram structures can lead to incomplete information extraction. To address this, the data processing system employs a hierarchical and progressive parsing strategy. First, the diagram is analyzed holistically to identify key functional areas and layout features. Next, detailed information about each area, including component attributes and data relationships, is analyzed layer by layer. Textual descriptions are then used to supplement implicit information that is difficult to directly identify. For example, when parsing an electrical schematic, the system first identifies the main circuit modules, then analyzes component connections. Finally, textual descriptions are used to supplement component parameters and operating conditions, ensuring complete knowledge extraction.

[0130] S212. Based on the document logic hierarchy of the target knowledge unit, the multimodal knowledge sequence is arranged hierarchically to obtain a structured knowledge body.

[0131] Among them, the document logical hierarchy represents the hierarchical organizational structure of knowledge content; structured knowledge body refers to the complete knowledge representation reorganized according to the hierarchical relationship; hierarchical arrangement is used to represent the subordinate and parallel relationship of knowledge content; knowledge integrity represents the completeness of each component in the knowledge structure.

[0132] The data processing system performs this step after generating a multimodal knowledge sequence. Specifically, the data processing system first analyzes the hierarchical attributes of the target knowledge units to determine the primary and secondary relationships of the content. It then reconstructs the hierarchical framework of the knowledge sequence based on the document structure tree. Multimodal content is then mapped to corresponding hierarchies based on logical relationships. Cross-hierarchical references are then processed to ensure the coherence of knowledge transfer. Finally, the hierarchical structure is optimized to generate a knowledge system that is easy to understand and navigate.

[0133] In some embodiments, the construction of a structured knowledge body can be achieved through various approaches: Alternatively, the data processing system can employ a tree-based approach to construct a multi-level directory index, represent knowledge hierarchies through node relationships, and use a recursive algorithm to process nested structures, thereby achieving hierarchical organization of knowledge content. Alternatively, the data processing system can employ a semantic network-based approach to establish a concept hierarchy diagram, define the subordinate relationships between knowledge units through ontology relationships, and combine inference rules to construct a complete knowledge structure. It is understood that other knowledge organization methods can also be employed, and are not limited here.

[0134] In actual applications, you may encounter problems with structural chaos caused by unreasonable knowledge hierarchy division. To address this situation, the data processing system will adopt hierarchical optimization and structural normalization strategies. The system first designs a standardized hierarchical template based on the domain knowledge system, including fixed levels such as basic concepts, principle descriptions, and operating procedures; then classifies and maps the knowledge content to ensure that the knowledge granularity of each level is appropriate; at the same time, it handles hierarchical adjustments in special circumstances. For example, when organizing equipment maintenance knowledge, the system will arrange maintenance requirements, operating procedures, precautions, and other content according to a standard hierarchy. For important safety warning information, its level will be raised to highlight it.

[0135] S213. Generate knowledge retrieval results based on the hierarchical relationship of the structured knowledge body.

[0136] Among them, the knowledge retrieval results represent the final knowledge content returned by the system in response to user query requirements; the hierarchical relationship refers to the organizational structure and logical relationship of each part of the knowledge system; the result presentation form is used to represent the display and interaction methods of the retrieved content; the knowledge adaptability indicates the degree of match between the retrieval results and user needs.

[0137] The data processing system performs this step after constructing the structured knowledge body. Specifically, the data processing system first determines the presentation level and level of detail of the results based on the search intent; then extracts the corresponding level of content from the structured knowledge body; then adjusts the layout of the multimodal content to ensure the rationality of the visual presentation; then adds interactive navigation elements to support user knowledge exploration; and finally, personalizes and optimizes the search results based on user characteristics and application scenarios.

[0138] In some embodiments, search result generation can be achieved through a variety of methods: Optionally, the data processing system can employ a template-based rendering approach, designing result templates for different types of queries, filling corresponding locations with structured knowledge, achieving standardized content display through style customization, and adding related recommendations and quick navigation features. Optionally, the data processing system can employ a dynamic generation approach, analyzing user demand characteristics in real time, adaptively adjusting the organization of content, generating summaries through deep learning models, and combining visualization technology to enhance expressiveness. It is understood that other result generation methods can also be employed, and are not limited here.

[0139] In actual applications, you may encounter the problem that the search results are too professional and difficult to understand. To address this situation, the data processing system will adopt an intelligent interpretation and hierarchical display strategy. The system first identifies the professional terms and complex concepts in the results and automatically generates easy-to-understand explanations; then organizes the content according to different levels such as basic knowledge, professional knowledge, and in-depth knowledge; and provides knowledge path navigation to help users understand the relevant content step by step. For example, when a user inquires about the working principle of a certain protective device, the system will first display the basic functional description, and then gradually delve into the specific technical parameters and implementation mechanisms to ensure that users of different professional levels can obtain appropriate knowledge support.

[0140] In the embodiments of this application, due to the use of innovative technologies such as document parsing based on a large language model, knowledge link construction, and multimodal content fusion, it is possible to achieve intelligent processing and knowledge reorganization of documents in the power sector, including core functions such as professional terminology recognition, chart content parsing, and knowledge unit association. Through dynamically optimized knowledge links and adaptive retrieval services, the problems of knowledge fragmentation, low relevance, and inaccurate retrieval existing in traditional document management systems are effectively solved, thereby achieving systematic organization, accurate retrieval, and intelligent services for knowledge in the power sector, significantly improving knowledge management efficiency and service quality.

[0141] The data processing system in the embodiment of the present invention is described below from the perspective of hardware processing. Figure 3 , is a schematic diagram of a physical device structure of a data processing system in an embodiment of the present application.

[0142] It should be noted that Figure 3 The structure of the data processing system shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0143] like Figure 3 As shown, the data processing system includes a CPU 301, which can perform various appropriate actions and processes according to the programs stored in the ROM 302 or the programs loaded from the storage unit 308 into the RAM 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are connected to each other via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0144] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, push button switches, and the like; an output section 307 including a liquid crystal display (LCD), an audio output device, indicator lights, and the like; a storage section 308 including a hard disk and the like; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. Removable media 311, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 310 as needed, so that computer programs read from the removable media can be installed in the storage section 308 as needed.

[0145] In particular, according to an embodiment of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, an embodiment of the present invention includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 309 and / or installed from the removable medium 311. When the computer program is executed by the CPU 301, the various functions defined in the present invention are performed.

[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings.

[0147] Specifically, the data processing system of this embodiment includes a processor and a memory, and a computer program is stored in the memory. When the computer program is executed by the processor, the method for constructing a dynamic knowledge base for the power industry based on a large language model provided in the above embodiment is implemented.

[0148] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the data processing system described in the above embodiments, or may exist independently and not be incorporated into the data processing system. The above storage medium carries one or more computer programs, and when the one or more computer programs are executed by a processor of the data processing system, the data processing system implements the method for constructing a dynamic knowledge base for the power industry based on a large language model provided in the above embodiments.

[0149] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0150] As used in the above embodiments, the term “when” may be interpreted to mean “if” or “after” or “in response to determining that” or “in response to detecting that”, depending on the context. Similarly, the phrases “upon determining that” or “if (stated condition or event) is detected” may be interpreted to mean “if determining that” or “in response to determining that” or “upon detecting (stated condition or event)” or “in response to detecting (stated condition or event)”, depending on the context.

Claims

1. A method for constructing a dynamic knowledge base for the power industry based on a large language model, characterized in that: Applied to a data processing system, the method comprises: Receive a power domain document imported by a user, and convert the power domain document into an electronic document containing a unique document identifier; Invoking a large language model to parse the electronic document, extracting professional term definitions, chart location information, and document logical hierarchy in the electronic document, and generating a knowledge unit containing multiple contexts; Connecting the knowledge units into directed knowledge links, and calculating the semantic similarity between adjacent knowledge units in the directed knowledge links; Segmenting the directed knowledge link based on the semantic similarity to obtain multiple knowledge topics; receiving a knowledge retrieval instruction and determining a target knowledge topic corresponding to the knowledge retrieval instruction; The target knowledge unit corresponding to the target knowledge subject is extracted, and a knowledge retrieval result is generated based on the target context of the target knowledge unit.

2. The method according to claim 1, characterized in that The step of receiving a power domain document imported by a user and converting the power domain document into an electronic document containing a unique document identifier specifically includes: Receive the power field document imported by the user, extract the document attribute information of the power field document, and obtain the document type, creation time and version number; Selecting a corresponding layout analysis model based on the document type, identifying the layout structure of the power field document including text paragraphs, tables, and graphics, and obtaining a document structure tree; Generate a unique document identifier based on the document structure tree and the version number; The layout structure and the document unique identifier are packaged into an electronic document in a preset standard format.

3. The method according to claim 1, characterized in that Before the step of calling the large language model to parse the electronic document, extracting the definitions of professional terms, chart location information, and document logical hierarchy in the electronic document, and generating a knowledge unit containing multiple contexts, the method further includes: Constructing a training sample set based on the power field documents; the training sample set includes professional term definitions, equipment operation steps and fault handling processes; Extracting professional terms from the power field documents and constructing a power field vocabulary; The electric power field vocabulary is added to a pre-trained language model, and the pre-trained language model is incrementally trained using the training sample set to obtain the large language model.

4. The method according to claim 1, wherein The step of extracting the target knowledge unit corresponding to the target knowledge topic and generating a knowledge retrieval result based on the target context of the target knowledge unit specifically includes: Extracting the target knowledge unit corresponding to the target knowledge subject, obtaining the document location information of the target knowledge unit, and constructing the target context of knowledge retrieval; Extracting graphic content and table content at the target location according to the target context to generate a multimodal knowledge sequence; Based on the document logic hierarchy of the target knowledge unit, the multimodal knowledge sequence is arranged hierarchically to obtain a structured knowledge body; A knowledge retrieval result is generated according to the hierarchical relationship of the structured knowledge body.

5. The method according to claim 4, characterized in that The step of extracting the graphic content and table content of the target location according to the target context to generate a multimodal knowledge sequence specifically includes: Locating and extracting the graphic primitive features of the target graphic and the cell layout features of the target table according to the chart position information in the target context; Identifying graphic components and connection relationships in the target graphic, and converting the graphic components and connection relationships into a structured graphic description; Identify the header level and data area of ​​the target table, extract the relationship between cells, and convert it into a structured table description; The structured graphic description and the structured table description are combined with corresponding text description blocks to generate a multimodal knowledge sequence.

6. The method according to claim 1, wherein After the steps of connecting the knowledge units into directed knowledge links and calculating the semantic similarity between adjacent knowledge units in the directed knowledge links, the method further includes: Calculating the out-degree value and in-degree value of each knowledge unit in the directed knowledge link to obtain the node weight; Based on the node weights, a transition probability matrix of the directed knowledge link is constructed, and a page walk score is calculated; Identifying a key path in the directed knowledge link according to the node weight and the page walk score; Based on the critical path, redundant links with node weights lower than a preset weight threshold are deleted, and missing association relationships are supplemented to optimize the directed knowledge link.

7. The method according to claim 6, characterized in that The step of calculating the out-degree value and in-degree value of each knowledge unit in the directed knowledge link to obtain the node weight specifically includes: Traverse each knowledge unit in the directed knowledge link, count the number of edges pointing to other knowledge units, and obtain the out-degree value; Traverse each knowledge unit in the directed knowledge link, count the number of edges from other knowledge units, and obtain the in-degree value; A weighted calculation is performed based on the out-degree value and the in-degree value to obtain a node weight.

8. A data processing system, characterized in that: The data processing system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the data processing system to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on a data processing system, the data processing system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product is run on a data processing system, the computer program product causes the data processing system to execute the method according to any one of claims 1 to 7.

Images