A power grid drawing intelligent review method and system based on a knowledge graph

Abstract

The application relates to the technical field of image data processing, and discloses a power grid drawing intelligent review method and system based on a knowledge graph, wherein the method comprises the following steps: driving a multi-modal large language model based on a field-specific prompt engineering, performing quality evaluation and conflict detection on power grid field knowledge triples extracted from structured text data, automatically resolving the conflicting knowledge to obtain a candidate knowledge set; reviewing and confirming the candidate knowledge set through a man-machine collaborative verification mechanism to construct a power grid design specification knowledge graph; identifying to-be-reviewed elements in to-be-reviewed power grid drawings through the multi-modal large language model, querying the power grid design specification knowledge graph according to generated structured query statements, judging the compliance of the to-be-reviewed power grid drawings based on obtained query data packets, and generating a review report. The application can improve the efficiency and reliability of automatic drawing review, and ensure the explainability and traceability of the review results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, in particular to an intelligent grid drawing review method and system based on a knowledge graph. BACKGROUND

[0002] The compliance review of grid engineering design drawings is a key link to ensure the safe and stable operation of the power system. The traditional review method highly depends on the professional knowledge and experience of designers and auditors, and the drawing content is compared by manually checking a large number of design specification documents, but the manual review method has problems such as low review efficiency, being easily affected by human factors, and being difficult to ensure the consistency of the review results. Moreover, with the continuous expansion of the grid scale and the increasing complexity of design specifications, the design specification documents have diverse sources, different formats, and frequent updates, and manual review is difficult to keep up with the specification update rhythm and is prone to outdated review. At the same time, there are conflicts between different sources and different versions of specifications, and it is difficult to quickly and accurately determine the priority of the specifications during manual review, which may lead to incorrect review results. In addition, it is difficult to leave complete review basis traceability records during the manual review process, and once the review results are disputed, it is difficult to trace back the review basis, affecting the credibility of the review results. Therefore, the existing drawing review technology has problems of low review efficiency and insufficient reliability, which cannot meet the requirements of the grid industry for drawing review.

[0003] The foregoing narrative is to provide general background information and does not necessarily constitute the prior art. SUMMARY

[0004] The embodiments of the present application provide an intelligent grid drawing review method and system based on a knowledge graph, which can improve the efficiency and reliability of automatic drawing review, ensure the explainability and traceability of the review results, and meet the requirements of the grid industry for drawing review.

[0005] Therefore, the embodiments of the present application provide an intelligent grid drawing review method based on a knowledge graph, which includes:

[0006] Performing a preprocessing operation on multi-source heterogeneous grid design specification documents to generate standardized structured text data;

[0007] Driving a multi-modal large language model based on a domain-specific prompt engineering to extract grid domain knowledge triples including confidence, original text location, and source metadata from the structured text data;

[0008] Performing quality evaluation and conflict detection on the grid domain knowledge triples, automatically resolving conflicting knowledge according to a preset priority rule to obtain a candidate knowledge set;

[0009] Reviewing and confirming the candidate knowledge set through a man-machine collaborative verification mechanism to construct a grid design specification knowledge graph;

[0010] In response to the drawing review task, a to-be-reviewed element in the to-be-reviewed power grid drawing is identified by the multi-modal large language model, and a structured query statement is generated;

[0011] The power grid design specification knowledge graph is queried based on the structured query statement to obtain a query data package, and the query data package includes compliance knowledge and corresponding metadata;

[0012] Based on the query data package, a compliance judgment is made on the to-be-reviewed power grid drawing, and a review report with original specification clause references is generated.

[0013] Further, in some embodiments of the present application, the preprocessing operation is performed on the multi-source heterogeneous power grid design specification documents to generate standardized structured text data, including:

[0014] Obtain multi-source heterogeneous power grid design specification documents in different formats, and establish a mapping relationship between the unique identifier of each specification document and the document format type;

[0015] The text content of the scanned specification document is extracted using optical character recognition technology, and the boundary information of chapter titles, text paragraphs, table regions and formula regions in the specification document is identified and determined by a layout analysis algorithm;

[0016] The extracted text content is hierarchically divided according to the chapter titles, the divided text content is aligned and classified according to paragraph semantics, the table regions are parsed and converted to key-value pair format, and redundant information is removed, to generate standardized structured text data.

[0017] Further, in some embodiments of the present application, the multi-modal large language model is driven based on the domain-specific prompt engineering to extract power grid domain knowledge triples including confidence, original text location and source metadata from the structured text data, including:

[0018] A power grid domain-specific prompt template is constructed, the prompt template includes a power grid professional terminology dictionary, a knowledge triple mandatory output format definition and a thinking chain reasoning guide sentence, the power grid professional terminology dictionary contains polysemous term interpretation, and the thinking chain reasoning guide sentence is used to indicate the logical association of the multi-modal large language model in parsing specification clauses step by step;

[0019] The structured text data is split into text segments by chapters, each text segment is spliced with the prompt template and input into the multi-modal large language model, the multi-modal large language model is driven to identify core entities and logical relationships in specification clauses through thinking chain reasoning to generate preliminary knowledge triples, and the core entities include subjects and objects.

[0020] corresponding to the preliminary knowledge triplets, chapter numbers and page information to complete the source metadata, binding the confidence of the multi-modal large language model output with the preliminary knowledge triplets, source metadata and original text location to form the power grid domain knowledge triplets.

[0021] Further, in some embodiments of the present application, the quality evaluation and conflict detection of the power grid domain knowledge triplets are performed, and the candidate knowledge set is obtained by automatically resolving the conflicting knowledge according to a preset priority rule, including:

[0022] A preset confidence threshold is set, and the power grid domain knowledge triplets with a confidence lower than the preset confidence threshold are removed, and the high-confidence knowledge triplets are retained;

[0023] The high-confidence knowledge triplets are compared by a knowledge matching algorithm, and when the subject and the relationship of two high-confidence knowledge triplets are completely consistent and the object is different, it is determined that there is a knowledge conflict between the two high-confidence knowledge triplets;

[0024] The preset priority rule is called to process the high-confidence knowledge triplets with knowledge conflicts, the knowledge triplets meeting the preset priority rule are marked as candidate effective knowledge, and the remaining knowledge triplets with knowledge conflicts are marked as to-be-expert-confirmed knowledge, and the candidate effective knowledge and the to-be-expert-confirmed knowledge are merged to obtain the candidate knowledge set.

[0025] Further, in some embodiments of the present application, the candidate knowledge set is audited and confirmed by a man-machine cooperative verification mechanism, and a power grid design specification knowledge graph is constructed, including:

[0026] Based on the active learning mechanism, the historical verification records of domain experts are analyzed, the knowledge domains with an extraction error rate of the multi-modal large language model higher than a preset error threshold are identified, and the candidate knowledge in the knowledge domains is preferentially pushed to an expert verification interface;

[0027] The knowledge triplets with the same source and consistent relationship type in the candidate knowledge set are clustered and processed to generate a batch verification task, and an operation entry is provided on the expert verification interface;

[0028] Operation data of the domain experts confirming or adjusting the candidate effective knowledge through the expert verification interface are obtained;

[0029] Based on the operation data, the knowledge triplets that pass the audit are written into the power grid design specification knowledge graph, and the effective time, the abolition time and the version number of each knowledge triplet are recorded.

[0030] Further, in some embodiments of the present application, in response to the drawing review task, the multi-modal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed, and a structured query statement is generated, including:

[0031] In response to the drawing review task, the image data of the power grid drawing to be reviewed is input into the multi-modal large language model, and the multi-modal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed; the elements to be reviewed include power grid equipment, equipment parameters and spatial relationships;

[0032] The multi-modal large language model automatically generates a structured query statement in accordance with the elements to be reviewed, which conforms to the power grid design specification knowledge graph query syntax, and the structured query statement includes a subject to be checked and a relationship to be checked.

[0033] Further, in some embodiments of the present application, the power grid design specification knowledge graph is queried based on the structured query statement to obtain a query data package, and the query data package includes compliance knowledge and corresponding metadata, including:

[0034] After the power grid design specification knowledge graph receives the structured query statement, the compliance knowledge currently in the effective state is matched;

[0035] The metadata corresponding to the compliance knowledge in the effective state is extracted, and the compliance knowledge and the corresponding metadata are packaged to form the query data package; the metadata includes specification name, standard number and specific clause number.

[0036] Further, in some embodiments of the present application, the power grid drawing to be reviewed is judged based on the query data package to generate a review report with original specification clause references, including:

[0037] The requirement value of the compliance knowledge is extracted from the query data package, and the actual value of the equipment parameter in the power grid drawing to be reviewed is compared with the requirement value to generate a corresponding compliance determination result;

[0038] The compliance determination result, the actual value of the equipment parameter, the requirement value and the corresponding metadata are integrated to generate the review report; the review report includes a conclusion, a problem description, a drawing value, a requirement value and original specification clause reference information.

[0039] Further, in some embodiments of the present application, the power grid design specification knowledge graph adopts a three-level hierarchical architecture including an ontology layer, a data layer and a service layer. The ontology layer is used to define the core concepts of the power grid field and their association, and to construct the power grid field knowledge system. The data layer is used to store knowledge triples and version information using a time series database for multi-dimensional retrieval. The service layer is used to provide knowledge query and incremental update services through standardized interfaces, and to realize integration with the power grid drawing review system.

[0040] Correspondingly, the embodiments of the present application provide a power grid drawing intelligent review system based on a knowledge graph, comprising:

[0041] a document preprocessing module configured to perform preprocessing operations on multi-source heterogeneous power grid design specification documents to generate standardized structured text data;

[0042] a knowledge extraction module configured to extract power grid field knowledge triples including confidence, original text location and source metadata from the structured text data based on a domain-specific prompt engineering driven multi-modal large language model;

[0043] a knowledge quality evaluation and conflict detection module configured to evaluate the quality of the power grid field knowledge triples and detect conflicts, automatically resolve conflicting knowledge according to a preset priority rule, and obtain a candidate knowledge set;

[0044] an intelligent human-computer collaborative verification module configured to verify and confirm the candidate knowledge set through a human-computer collaborative verification mechanism, and to construct a power grid design specification knowledge graph;

[0045] a scenario understanding and query generation module configured to identify unreviewed elements in an unreviewed power grid drawing through the multi-modal large language model and generate a structured query statement in response to a drawing review task;

[0046] a compliance judgment and report generation module configured to query the power grid design specification knowledge graph based on the structured query statement to obtain a query data package, the query data package including compliance knowledge and corresponding metadata; and to perform compliance judgment on the unreviewed power grid drawing based on the query data package and generate a review report with original specification clause references.

[0047] The application provides a power grid drawing intelligent review method and system based on a knowledge graph. First, a standardized structured text is generated by preprocessing multi-source heterogeneous power grid design specification documents, solving the basic problem of different formats and different sources of specifications being difficult to unify processing, and providing regular data support for subsequent knowledge extraction. Then, based on the field-specific prompt engineering, a multi-modal large language model extracts power grid knowledge triples containing confidence, original text location and source metadata, which not only realizes the structured conversion of specification knowledge, but also retains the association between knowledge and original specifications through metadata. After the knowledge graph is constructed through quality evaluation, conflict resolution and human-computer collaborative verification, the conflicting knowledge is processed by priority rules, and combined with expert review to ensure the accuracy and authority of the knowledge in the graph. Finally, in response to the review task, the multi-modal large language model identifies the drawing elements to generate a query statement, queries the graph to obtain compliant knowledge, completes the compliance judgment and generates a report with original specification references, realizing the closed loop from specification knowledge structured construction to drawing intelligent review. Therefore, the application replaces a large amount of manual sorting and comparison work through automatic processing, improves the review efficiency and accuracy, and enhances the reliability and explainability of the review results due to the authority of the knowledge graph and the explicit reference of the original specification in the defect report, thereby promoting the intelligent upgrading of power grid drawing review. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0049] Figure 1 is an application environment of the power grid drawing intelligent review method based on the knowledge graph provided by the embodiments of the application;

[0050] Figure 2 is a flowchart of the power grid drawing intelligent review method based on the knowledge graph provided by the embodiments of the application;

[0051] Figure 3 is a structural schematic diagram of the power grid drawing intelligent review system based on the knowledge graph provided by the embodiments of the application;

[0052] Figure 4 is an architecture diagram of the MLLM-driven human-computer collaborative knowledge graph construction pipeline provided by the embodiments of the application;

[0053] Figure 5 is a flowchart of the knowledge graph enhanced MLLM reasoning and traceability provided by the embodiments of the application;

[0054] Figure 6 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] The exemplary embodiments will be described in detail with reference to the accompanying drawings. In the following description, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments are not meant to represent all embodiments consistent with the present application. Rather, they are merely examples of systems and methods consistent with some aspects of the present application, as detailed in the appended claims.

[0056] It should be noted that, in this document, by the term "comprising" or "including" or any other variant thereof, it is intended to encompass non-exclusive inclusion, such that processes, methods, articles, or apparatuses that comprise a list of elements are not limited to those elements, but can also include other elements not expressly listed, or inherent to such processes, methods, articles, or apparatuses. Without further limitation, an element defined by the phrase "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element. Moreover, components, features, elements with the same name in different embodiments of the present application can have the same meaning or different meanings, and the specific meaning thereof should be determined in the light of its explanation in the specific embodiment or further in conjunction with the context in the specific embodiment.

[0057] It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0058] In the following description, the suffixes "module", "part", or "unit" used for an element are merely intended for facilitating explanation of the present application, and by itself do not have a specific meaning. Therefore, "module", "part", or "unit" can be mixedly used.

[0059] In order to solve the above technical problems and break through the defects of the prior art, the embodiments of the present application provide a power grid drawing intelligent review method and system based on a knowledge graph, which can improve the efficiency and reliability of automatic review of drawings, while ensuring the explainability and traceability of the review results.

[0060] Figure 1 is an application environment diagram of the power grid drawing intelligent review method based on a knowledge graph in an embodiment. Refer to Figure 1The knowledge graph-based power grid drawing intelligent review method should be based on a knowledge graph-based power grid drawing intelligent review system. The knowledge graph-based power grid drawing intelligent review system includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network, and the terminal 110 can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, a tablet computer, a notebook computer, etc. The server 120 can be implemented by an independent server or a server cluster composed of multiple servers. The server 120 is configured to execute the above-mentioned knowledge graph-based power grid drawing intelligent review method, which includes: performing a preprocessing operation on a multi-source heterogeneous power grid design specification document to generate standardized structured text data; driving a multi-modal large language model based on a domain-specific prompt to extract power grid domain knowledge triples including confidence, original text location and source metadata from the structured text data; performing quality assessment and conflict detection on the power grid domain knowledge triples, automatically resolving conflicting knowledge according to a preset priority rule to obtain a candidate knowledge set; verifying and confirming the candidate knowledge set through a man-machine cooperative verification mechanism to construct a power grid design specification knowledge graph; in response to a drawing review task, identifying the to-be-reviewed elements in the to-be-reviewed power grid drawing through a multi-modal large language model to generate a structured query statement; querying the power grid design specification knowledge graph based on the structured query statement to obtain a query data package, the query data package including compliance knowledge and corresponding metadata; and performing compliance judgment on the to-be-reviewed power grid drawing based on the query data package to generate a review report with original specification clause references.

[0061] Please refer to Figure 2 , Figure 2 is a flowchart of the knowledge graph-based power grid drawing intelligent review method provided by an embodiment of the present application. The embodiment mainly takes the application of the knowledge graph-based power grid drawing intelligent review method in a computer device as an example to illustrate, and the knowledge graph-based power grid drawing intelligent review method provided by an embodiment of the present application can specifically include the following steps:

[0062] S1. Performing a preprocessing operation on a multi-source heterogeneous power grid design specification document to generate standardized structured text data;

[0063] Specifically, for step S1, first, the power grid design specification documents of different formats and sources are preprocessed. The preprocessing operation includes parsing, format conversion, content extraction, etc. of the document, and the purpose is to convert these heterogeneous documents into standardized structured text data. For example, the OCR technology can be used to extract the text content in the scanned document, the section title, the text paragraph, the table area, etc. in the document are recognized through the layout analysis algorithm, and they are converted into a unified text format.

[0064] S2. Extract power grid domain knowledge triples including confidence, original text location, and source metadata from structured text data based on domain-specific prompt engineering for driving multimodal large language models;

[0065] Specifically, for step S2, domain-specific prompt engineering is used to guide multimodal large language models (MLLMs) to extract key power grid domain knowledge from structured text data. This knowledge is represented in the form of triples, including confidence, original text location, and source metadata. For example, by designing specific prompt templates, MLLMs can identify key entities, relationships, and attributes in the text and combine them into knowledge triples, while recording their locations and sources in the original text.

[0066] S3. Quality assessment and conflict detection of power grid domain knowledge triples, automatic resolution of conflicting knowledge according to preset priority rules, and generation of a candidate knowledge set;

[0067] Specifically, for step S3, the quality and consistency of the extracted knowledge triples are evaluated, and potential knowledge conflicts are detected and resolved. Through preset priority rules, the system can automatically resolve conflicts and generate a candidate knowledge set. For example, when there are contradictions between knowledge triples from different sources, the system can select the most reliable knowledge as a candidate based on factors such as the reliability of the knowledge source and the age of the version.

[0068] S4. Review and confirm the candidate knowledge set through a human-machine collaborative verification mechanism to build a power grid design specification knowledge graph;

[0069] Specifically, for step S4, the candidate knowledge set needs to be reviewed by a human-machine collaborative verification mechanism to ensure its accuracy and reliability. Domain experts use a special verification interface to confirm or correct the candidate knowledge. Finally, the reviewed knowledge is integrated into the power grid design specification knowledge graph. This process combines artificial intelligence and machine intelligence, ensuring the accuracy of the knowledge and improving the efficiency of the knowledge graph construction.

[0070] S5. In response to the drawing review task, identify the elements to be reviewed in the power grid drawings to be reviewed through the multimodal large language model, and generate structured query statements;

[0071] Specifically, for step S5, in the drawing review task, a multi-modal large language model (MLLM) is used to identify key elements in the drawing and convert them into structured query statements. These query statements are used to find relevant compliance knowledge in the power grid design specification knowledge graph. For example, the MLLM can identify elements such as device types, parameters, connection relationships, etc. in the drawing and generate query statements that conform to the knowledge graph query syntax.

[0072] S6. Query the power grid design specification knowledge graph based on the structured query statement to obtain a query data package, the query data package including compliance knowledge and its corresponding metadata;

[0073] Specifically, for step S6, the structured query statement is used to retrieve relevant compliance knowledge and its metadata in the power grid design specification knowledge graph. The query data package contains information such as specification requirements, standard clauses, etc. related to the elements to be reviewed. This step provides necessary data support for subsequent compliance judgment.

[0074] S7. Based on the query data package, perform compliance judgment on the power grid drawing to be reviewed, and generate a review report with original specification clause references;

[0075] Specifically, for step S7, finally, based on the compliance knowledge in the query data package, the power grid drawing is subjected to compliance judgment. The judgment result will generate a review report containing compliance evaluation of each element in the drawing, and reference to the relevant original specification clauses as the basis. This step ensures the accuracy and traceability of the review results.

[0076] This embodiment realizes efficient conversion from multi-source heterogeneous power grid design specification documents to structured knowledge graph, and uses a multi-modal large language model (MLLM) to realize intelligent review of power grid drawings, improving the efficiency and accuracy of power grid design drawing review, and reducing the work intensity and error rate of manual drawing review. By constructing the power grid design specification knowledge graph, high-credibility and interpretable review results can be provided, providing strong technical support for quality assurance of power grid engineering design.

[0077] Further, in some embodiments, step S1 "performing preprocessing operations on multi-source heterogeneous power grid design specification documents to generate standardized structured text data" can specifically include:

[0078] S11. Obtain multi-source heterogeneous power grid design specification documents in different formats, and establish a mapping relationship between the unique identifier of each specification document and the document format type;

[0079] Specifically, for step S11, in the preprocessing stage of the grid design specification document, first, documents from different sources and formats need to be processed. These documents may include national standards, industry standards, enterprise standards, etc., and the formats may be PDF, Word, HTML, or image scans, etc. In order to efficiently manage and process these documents, the system needs to assign a unique identifier to each document and establish a mapping relationship between the identifier and the document format type. This step ensures that in the subsequent processing process, the system can quickly identify and call each document, and apply the corresponding processing method according to its format. In practical application, the document management system (DMS) can be used to store and manage these multi-source heterogeneous documents. By using a universally unique identifier (UUID) to generate a globally unique identifier for each document, accuracy and efficiency can be ensured in large-scale document processing scenarios. In addition, metadata management technology can associate the format type, source information, version information, etc. of the document with the unique identifier to form a complete document information library.

[0080] S12. Adopting optical character recognition technology to extract the text content of the scanned specification document, and identifying and determining the boundary information of chapter titles, text paragraphs, table regions and formula regions in the specification document through layout analysis algorithm;

[0081] Specifically, for step S12, for grid design specification documents in image form (such as scanned PDF files or drawings), optical character recognition (OCR) technology is a key means to extract the text content therein. OCR technology can recognize and convert the text in the image into editable text format. At the same time, the layout analysis algorithm further processes the recognized text content, analyzes the layout structure of the document, and determines the boundary positions of chapter titles, text paragraphs, tables and formulas, etc. different regions, providing accurate region division for subsequent text structure analysis and information extraction. For layout analysis, target detection algorithms in deep learning such as Faster R-CNN or YOLO can be used to locate and classify different regions in the document, thereby realizing automated layout structure analysis. In addition, natural language processing technology can also be combined to perform semantic analysis and understanding on the extracted text content, further improving the intelligent level of document preprocessing.

[0082] S13. According to the chapter title, the extracted text content is hierarchically divided, the aligned and classified text content is aligned and classified according to the paragraph semantics, the table region is analyzed and converted into key-value pair format, and the redundant information is removed, generating standardized structured text data;

[0083] Specifically, for step S13, after the extraction and layout analysis of the text content are completed, the text content needs to be structured. According to the chapter title, the text is divided into hierarchical sections, which can organize the document content into hierarchical information units, facilitating subsequent knowledge extraction and semantic understanding. Aligning and classifying the text paragraphs according to semantics can help integrate related text information and improve the coherence and readability of the information. For table regions, row and column analysis and conversion to key-value pair format can convert table data into structured data format for computer processing and querying. At the same time, during the entire processing process, redundant information such as repeated text and irrelevant annotations is removed, which can reduce data volume, improve data quality and processing efficiency, and finally generate standardized structured text data, providing high-quality input for subsequent knowledge graph construction.

[0084] The embodiment improves the accuracy and efficiency of document preprocessing by specifying the preprocessing method of multi-source heterogeneous power grid design specification documents, and converts power grid design specification documents of different formats and complex sources into standardized structured text data, providing a high-quality, unified format data basis for subsequent knowledge extraction and knowledge graph construction.

[0085] Further, in some embodiments, step S2 "extracting power grid domain knowledge triples including confidence, original text location and source metadata from structured text data based on domain-specific prompt engineering driven multi-modal large language model" can specifically include:

[0086] S21. Constructing a power grid domain-specific prompt template, the prompt template including a power grid professional term dictionary, a knowledge triple mandatory output format definition and a thinking chain reasoning guide sentence, the power grid professional term dictionary containing polysemous term interpretation, and the thinking chain reasoning guide sentence used to indicate the multi-modal large language model to analyze the logical association of specification clauses step by step;

[0087] Specifically, for step S21, constructing the power grid field-specific prompt template is the key to driving the MLLM to accurately extract knowledge. The prompt template includes a power grid professional terminology dictionary, a knowledge triple mandatory output format definition, and a thinking chain reasoning guide sentence. The power grid professional terminology dictionary covers polysemous terms and their interpretations in the power grid field, helping the MLLM understand the specific meanings of terms in different contexts. The knowledge triple mandatory output format definition ensures that the knowledge triples output by the MLLM have a uniform structure, facilitating subsequent processing. The thinking chain reasoning guide sentence guides the MLLM to analyze step by-step, identify core entities and logical relationships in the standard clauses. When constructing the prompt template, expert knowledge and historical data in the power grid field can be fully utilized. By analyzing a large number of power grid design specification documents, common professional terms and their polysemous usage are extracted and organized into a semantic dictionary. At the same time, multiple thinking chain reasoning guide sentence templates are designed, providing corresponding guide sentences for different standard clause types (such as safety distance requirements, equipment parameter standards, etc.), helping the MLLM more accurately identify and analyze key information in the clauses.

[0088] S22. The structured text data is split by chapter into text segments, each text segment is spliced with the prompt template and input into the multi-modal large language model, driving the multi-modal large language model to identify core entities and logical relationships in the standard clauses through thinking chain reasoning, generating preliminary knowledge triples, the core entities include the subject and the object;

[0089] Specifically, for step S22, the structured text data obtained by preprocessing is split by chapter to form multiple text segments. Each text segment is spliced with the constructed prompt template and provided as input to the MLLM. This step enables the MLLM to focus on the content in each text segment under the guidance of the prompt template, accurately identifying and extracting knowledge triples therein. The text segment splitting strategy can be optimized according to the characteristics of the power grid design specification documents. For example, according to the hierarchical structure of the documents, the content can be split into finer granularity segments, such as by clause, sub-clause, to enable the MLLM to more accurately process each independent specification content. In addition, natural language processing techniques can be used to pre-analyze the text segments, identifying key sentences and key terms therein, providing auxiliary information for the knowledge extraction of the MLLM.

[0090] S23. Associate the source document identifier, chapter number and page number information corresponding to the preliminary knowledge triples to complete the source metadata, bind the confidence level output by the multi-modal large language model with the preliminary knowledge triples, source metadata and original text location, form power grid field knowledge triples;

[0091] Specifically, for step S23, after the MLLM outputs the preliminary knowledge triples, it is necessary to associate these triples with their location information in the original document, including the unique identification of the source document, chapter number, and page number information, etc. In this way, complete source metadata is supplemented for each knowledge triple, ensuring the traceability and accuracy of the knowledge. Finally, the confidence level output by the MLLM is bound with the preliminary knowledge triples, source metadata, and original text location information to form a complete power grid domain knowledge triple. The confidence level reflects the credibility of the model's extraction results, while the original text location records the specific location of the knowledge in the original document. This step provides the necessary information for subsequent knowledge evaluation and conflict detection.

[0092] The present embodiment improves the accuracy and consistency of knowledge extraction in power grid design specification documents by constructing power grid domain-specific prompt templates; splitting structured text data into text segments and inputting them one by one into the MLLM after splicing with prompt templates, which can more accurately identify and extract knowledge triples while ensuring traceability of the knowledge; binding the confidence level output by the MLLM with the knowledge triples and their source metadata, providing a reliable basis for subsequent knowledge quality evaluation and conflict detection.

[0093] Further, in some embodiments, step S3 "quality evaluation and conflict detection of power grid domain knowledge triples, automatic resolution of conflicting knowledge according to preset priority rules, to obtain a candidate knowledge set" can specifically include:

[0094] S31. Set a preset confidence threshold, and remove the power grid domain knowledge triples with a confidence level lower than the preset confidence threshold, and keep the high-confidence knowledge triples;

[0095] Specifically, for step S31, a preset confidence threshold is set for screening knowledge triples. Knowledge triples below this threshold are removed due to low credibility, and high-confidence knowledge triples above the threshold are kept to ensure the reliability of the knowledge for subsequent processing. The confidence threshold should be reasonably set according to the characteristics of the power grid domain knowledge and actual needs. The threshold range can be determined by analyzing historical data, and the sample data can be evaluated and optimized by manual annotation.

[0096] S32. Compare the high-confidence knowledge triples through a knowledge matching algorithm, and when the subject and relationship of two high-confidence knowledge triples are completely consistent and the object is different, it is determined that there is a knowledge conflict between the two high-confidence knowledge triples;

[0097] Specifically, for step S32, a knowledge matching algorithm is used to compare the subject and the relationship of the high-confidence knowledge triples. If the subject and the relationship are the same but the object is different, it is determined that there is a knowledge conflict, and the contradiction between different specifications is discovered and processed in a timely manner. A hybrid method based on rules and machine learning can be used to improve accuracy. At the same time, combining professional knowledge and experience in the power grid field, the algorithm is continuously optimized to improve the conflict detection efficiency.

[0098] S33. Invoke the preset priority rule to process the high-confidence knowledge triples with knowledge conflicts, mark the knowledge triples that meet the preset priority rule as candidate effective knowledge, mark the remaining knowledge triples with knowledge conflicts as expert-confirmed knowledge, merge the candidate effective knowledge and the expert-confirmed knowledge, and obtain a candidate knowledge set;

[0099] Specifically, for step S33, the preset priority rule is triggered to automatically process the conflicting knowledge. According to the rule, such as national standards being superior to line standards and new versions being superior to old versions, the knowledge triples that meet the conditions are marked as candidate effective knowledge, and the rest are marked as expert-confirmed knowledge. The priority rule should be formulated according to the specifications and standards of the power grid industry to ensure that the processing result meets the actual requirements. The formulation of the rule needs to invite experts in the power grid field to participate, and it should be updated regularly to adapt to changes in specifications.

[0100] The embodiment sets a confidence threshold and uses a knowledge matching algorithm to effectively screen and identify low-quality data and conflicting information in the knowledge triples in the power grid field. In combination with the preset priority rule, the conflicting knowledge is automatically processed to generate a reliable candidate knowledge set, improving the quality and credibility of the knowledge graph.

[0101] Further, in some embodiments, step S4 "auditing and confirming the candidate knowledge set through a human-machine collaborative verification mechanism to construct a power grid design specification knowledge graph" can specifically include:

[0102] S41. Based on the active learning mechanism, analyze the historical verification records of experts in the field, identify knowledge fields with an extraction error rate of the multi-modal large language model higher than a preset error threshold, and preferentially push the candidate knowledge in the knowledge field to an expert verification interface;

[0103] Specifically, for step S41, the historical verification records of domain experts are analyzed by an active learning mechanism to count the extraction error rates of the multi-modal large language model in different knowledge domains. If the error rate of a certain knowledge domain exceeds the preset threshold, the candidate knowledge in that domain will be prioritized for verification by experts to ensure that the knowledge in these error-prone domains receives special attention and correction. The core of the active learning mechanism lies in the accurate assessment of the credibility of the model extraction results through effective analysis of the expert verification history. To further improve the accuracy of intelligent verification, deep learning techniques can be introduced into the active learning mechanism, such as using recurrent neural networks (RNN) or long short-term memory networks (LSTM) to model the verification history. At the same time, an error rate dynamic adjustment model can be established to update the error rate in real time based on expert verification results, optimizing the priority of knowledge pushing.

[0104] S42. Clustering the knowledge triples in the candidate knowledge set with the same source and consistent relationship type, generating batch verification tasks, and providing operation entry in the expert verification interface;

[0105] Specifically, for step S42, the knowledge triples in the candidate knowledge set are clustered and analyzed according to their source and relationship type, and the knowledge triples with the same source and relationship type are clustered together to form batch verification tasks. In the expert verification interface, special operation entries are provided for these batch verification tasks to facilitate the batch processing of these tasks by experts. The key to clustering lies in designing an effective similarity measurement method. In addition to traditional text-based clustering methods, semantic-based clustering techniques can be introduced to deeply explore the semantic similarity of knowledge triples. Graph neural network (GNN)-based clustering methods can also be introduced to construct semantic graphs of knowledge triples. Meanwhile, combined with knowledge graph technology, knowledge triples can be embedded into knowledge graphs to utilize the structural information of the graph to assist clustering. In the knowledge graph, each knowledge triple can be represented as a node, and the edges represent the association between them. By analyzing the connection strength and path length of the graph structure features, the similarity of knowledge triples can be measured.

[0106] S43. Obtaining the operation data of the domain experts confirming or adjusting the candidate effective knowledge through the expert verification interface;

[0107] Specifically, for step S43, the domain experts carefully review the candidate valid knowledge on the expert verification interface, and verify and correct the accuracy of these knowledge through confirmation or adjustment operations. The system will obtain the operation data of the experts, including the confirmed knowledge, adjusted knowledge, and the experts' modification opinions, etc. To ensure accurate recording and efficient use of operation data, the system captures the experts' operation behaviors in the interface in real time, such as clicking, inputting text, etc., and converts them into structured operation data. A rule-based and machine learning combined method is used for semantic analysis of the operation data. The rule-based method identifies the experts' modification intentions based on pre-defined patterns, while the machine learning model automatically learns the experts' modification patterns and preferences by analyzing a large amount of operation history data.

[0108] S44. Based on the operation data, write the knowledge triples that pass the audit into the power grid design specification knowledge graph, and record the effective time, expiration time and version number of each knowledge triple;

[0109] Specifically, for step S44, according to the operation data of the experts, the knowledge triples that pass the audit are officially written into the power grid design specification knowledge graph. At the same time, the system records important information such as the effective time, expiration time and version number of each knowledge triple, in order to facilitate subsequent knowledge management and traceability. At the same time, combined with the time series database, the system stores the time series data such as the effective and expiration time of the knowledge triples, in order to facilitate knowledge analysis and mining in the time dimension. Blockchain technology can also be introduced to record the writing operation of the knowledge triples in a non-tamperable way, improving the security and credibility of the knowledge graph. Each writing operation is recorded as a blockchain transaction, containing information such as the timestamp of the writing and the hash value of the operation data.

[0110] The embodiment effectively improves the accuracy and authority of the knowledge graph through the active learning mechanism and clustering processing, prioritizes the knowledge of high error rate fields to ensure the key review of easy-to-mistake knowledge, generates batch verification tasks and provides operation entrances to greatly improve the efficiency of expert verification.

[0111] Further, in some embodiments, step S5 "in response to the drawing review task, identifying the to-be-reviewed elements in the to-be-reviewed power grid drawing through a multi-modal large language model, and generating a structured query statement" can specifically include:

[0112] S51. In response to the drawing review task, input the image data of the to-be-reviewed power grid drawing into the multi-modal large language model, and identify the to-be-reviewed elements in the to-be-reviewed power grid drawing through the multi-modal large language model; the to-be-reviewed elements include power grid equipment, equipment parameters and spatial relationships;

[0113] Specifically, for step S51, when receiving the grid drawing review task, input the image data of the grid drawing to be reviewed into the multi-modal large language model (MLLM). Utilizing the powerful visual understanding and language generation capabilities of MLLM, analyze and identify the drawing, and find out the elements to be reviewed, such as grid equipment, equipment parameters and spatial relationships, etc. Further, advanced computer vision techniques such as object detection algorithms (Faster R-CNN, YOLO, etc.) and semantic segmentation algorithms in deep learning can be used to more accurately locate and identify the equipment, parameters and spatial relationships in the drawing. At the same time, combined with the professional knowledge and experience in the field of power grid, the MLLM is trained and optimized to better understand and identify the complex information in the grid drawing.

[0114] S52. Automatically generate a structured query statement conforming to the knowledge graph query syntax of the grid design specification according to the elements to be reviewed by the multi-modal large language model, and the structured query statement includes the subject to be checked and the relationship to be checked;

[0115] Specifically, for step S52, according to the identified elements to be reviewed, MLLM will automatically generate a structured query statement conforming to the knowledge graph query syntax of the grid design specification. This query statement contains the subject to be checked (such as device name, parameter name, etc.) and the relationship to be checked (such as the connection relationship between devices, the compliance relationship of parameters, etc.), which is used for subsequent knowledge graph query operations. When generating the structured query statement, natural language processing techniques can be further combined to optimize and check the query statement, ensuring its grammatical correctness and semantic accuracy. At the same time, the pattern and structure information of the knowledge graph can also be used to intelligently complete and recommend the generated query statement, improving the efficiency and effectiveness of the query.

[0116] This embodiment utilizes the visual understanding and language generation capabilities of the multi-modal large language model, which can efficiently and accurately identify key elements in the drawing and generate query statements conforming to the knowledge graph query syntax, not only improving the efficiency and accuracy of grid drawing review, but also laying a solid foundation for subsequent compliance judgment and review report generation.

[0117] Further, in some embodiments, step S6 "query the grid design specification knowledge graph based on the structured query statement to obtain a query data package, and the query data package includes compliance knowledge and its corresponding metadata", specifically can include:

[0118] S61. After receiving the structured query statement in the grid design specification knowledge graph, match the compliance knowledge currently in the effective state;

[0119] Specifically, for step S61, during the grid drawing paper review process, the system generates a structured query statement based on the elements in the to-be-reviewed drawing. This query statement is used to retrieve compliance knowledge related to the to-be-reviewed elements and their corresponding metadata from the grid design specification knowledge graph. The metadata includes important information such as specification name, standard number, and specific clause number. The generation of the query statement can combine natural language processing techniques to convert the elements in the drawing into statements that conform to the knowledge graph query syntax. The system can use the knowledge graph's index and retrieval functions to quickly locate the relevant compliance knowledge based on the query statement and extract the corresponding metadata. After the grid design specification knowledge graph receives the structured query statement, it searches for compliance knowledge that is currently in effect in the grid design specification knowledge graph. Since the knowledge in the knowledge graph has a time dimension and state marker, it can directly identify which knowledge is currently valid and which may be outdated. During the construction of the knowledge graph, a time-series database can be used to store the knowledge in the knowledge graph to facilitate the management and query of the time dimension information of the knowledge.

[0120] S62. Extract the metadata corresponding to the effective state of the compliance knowledge, package the compliance knowledge and the corresponding metadata to form a query data package; the metadata includes specification name, standard number, and specific clause number;

[0121] Specifically, for step S62, for the matched compliance knowledge in the effective state, the system extracts its corresponding metadata. The metadata provides detailed background information of the compliance knowledge, such as specification name, standard number, and specific clause number. Efficient query statements can be written using the query language of the graph database, such as Cypher (for Neo4j database), to extract the metadata. The extracted compliance knowledge and corresponding metadata are integrated to form a query data package, which contains all the compliance knowledge and metadata related to the to-be-reviewed drawing, providing a comprehensive information foundation for subsequent compliance judgment.

[0122] This embodiment refines the process of querying the grid design specification knowledge graph based on structured query statements and generating query data packages, improving the accuracy and efficiency of the query, and quickly locating the currently effective compliance knowledge and extracting relevant metadata, providing comprehensive and accurate information support for subsequent compliance judgment.

[0123] Further, in some embodiments, step S7 "compliance judgment of the to-be-reviewed grid drawing based on the query data package, generating a review report with original specification clause references" can specifically include:

[0124] S71. Extract the requirement values of compliance knowledge from the query data packet, compare the actual values of the equipment parameters in the grid drawing under review with the requirement values, and generate the corresponding compliance determination results;

[0125] Specifically, for step S71, the query data packet contains the compliance knowledge retrieved from the grid design specification knowledge graph and its corresponding metadata. The system first extracts the specific requirement values in the compliance knowledge from the query data packet, which are the key basis for judging whether the grid drawing under review conforms to the specification. The requirement values can be specific numerical ranges (such as the ground safety distance of equipment), technical parameters, equipment models, etc. When extracting the requirement values, the system can use data parsing techniques to quickly locate and extract the required information based on the structured storage characteristics of the grid design specification knowledge graph. For example, using XML or JSON data format parsing methods, the corresponding field values in the query data packet can be directly obtained.

[0126] The actual values of the equipment parameters identified from the grid drawing under review are compared with the requirement values of the compliance knowledge extracted from the query data packet. Through this comparison, the system can determine whether the actual values meet the requirement values in the specification, thereby determining whether the drawing has compliance issues. The comparison process can use numerical comparison algorithms, string matching techniques, etc., and select appropriate comparison methods according to different data types. For numerical data, size comparison, range judgment, etc. can be performed; for text data, exact matching or fuzzy matching can be performed.

[0127] S72. Integrate the compliance determination results, equipment parameter actual values, requirement values, and corresponding metadata to generate an examination report; the examination report includes conclusions, problem descriptions, drawing values, requirement values, and original specification clause reference information;

[0128] Specifically, for step S72, the system integrates the compliance determination results (such as "compliance" or "non-compliance"), the actual values of the equipment parameters in the grid drawing under review, the requirement values of the compliance knowledge, and the corresponding metadata (such as specification name, standard number, specific clause number, etc.). Through this integration, the system can generate a complete examination report that details the review process and results, providing clear and accurate review evidence for the reviewer.

[0129] This embodiment refines the process of compliance judgment based on the query data packet and generates an examination report, realizes the automation process from data extraction, comparison to report generation, can accurately extract the requirement values of the compliance knowledge from the query data packet, and compare them with the actual values of the equipment parameters in the drawing under review, generate a detailed examination report, effectively improve the efficiency and accuracy of the grid drawing review, and also ensure the traceability and explainability of the review results.

[0130] Further, in some embodiments, the power grid design specification knowledge graph adopts a three-level hierarchical architecture including an ontology layer, a data layer, and a service layer. The ontology layer is used to define the core concepts of the power grid field and their associated relationships, and to construct the knowledge system of the power grid field. The data layer is used to store knowledge triples and version information using a time series database for multi-dimensional retrieval. The service layer is used to provide knowledge query and incremental update services through standardized interfaces, and to realize integration with the power grid drawing review system.

[0131] Specifically, for the ontology layer, the ontology layer is the basis of the power grid design specification knowledge graph, and is responsible for defining the core concepts of the power grid field (such as "transformer", "circuit breaker", "busbar", etc.) and their associated relationships (such as "connected to", "protected", "controlled", etc.). By constructing these core concepts and their relationships, a complete knowledge system of the power grid field is formed, providing a semantic framework for the construction of the knowledge graph.

[0132] For the data layer, it is responsible for storing specific data in the knowledge graph, including knowledge triples and version information. Knowledge triples represent the relationships between entities (such as "a certain type of circuit breaker - rated voltage - 126kV"), while version information records the effective time, expiration time, and version number of the knowledge, etc. The use of a time series database can effectively manage these time-dimensioned data, supporting multi-dimensional retrieval (such as by time, by version, by entity, etc.).

[0133] The service layer provides interface support for the use and maintenance of the knowledge graph. Through standardized interfaces (such as RESTful API, GraphQL, etc.), other systems (such as the power grid drawing review system) can easily query the data in the knowledge graph and obtain the required compliance knowledge and metadata. At the same time, the service layer also supports incremental updates of the knowledge graph, ensuring that the knowledge graph can reflect the latest power grid design specifications in a timely manner.

[0134] This embodiment provides a solid semantic foundation for the knowledge graph through the ontology layer, ensures accurate storage and fast retrieval of knowledge through the data layer, and realizes seamless integration of the knowledge graph and the review system through the service layer. Not only does this improve the construction and maintenance efficiency of the knowledge graph, but also enhances its practicality and reliability in intelligent review of power grid drawings, effectively promoting the intelligentization process of power grid engineering design review.

[0135] In order to facilitate understanding of the power grid drawing intelligent review method based on the knowledge graph provided in this embodiment, the following will be described in conjunction with specific scenarios.

[0136] Scenario setting:

[0137] Existing knowledge: In the system's knowledge graph, there is already a piece of knowledge from the national standard (GB): (Subject: "220kV live equipment", Relation: "Minimum safety distance to ground", Object: "3.0 meters", Source: "GB50061-2010 Article X").

[0138] New regulation: A certain power grid company releases a new internal enterprise standard (PDF format) to improve safety margin, which states: "To strengthen safety management, all 220kV live equipment within our jurisdiction should have a minimum safety distance to ground of not less than 3.5 meters."

[0139] Pending drawing: A design drawing of a new substation, in which a certain 220kV equipment is marked with a distance to ground of 3.2 meters.

[0140] First stage: Intelligent updating of knowledge graph.

[0141] Knowledge extraction: The administrator inputs the new company standard PDF file into the knowledge graph construction pipeline of the invention. The MLLM-driven knowledge extraction module successfully extracts a new knowledge triple from it through its domain-specific prompt engineering: (Subject: "220kV live equipment", Relation: "Minimum safety distance to ground", Object: "3.5 meters", Source: "Q / GDW12345-2025 Article Y").

[0142] Conflict detection and resolution: When this new knowledge is about to be integrated into the knowledge graph, the knowledge quality assessment and conflict detection module is activated. It finds that there is already a piece of knowledge about the same topic (220kV equipment safety distance to ground) but with different content (3.0 meters vs 3.5 meters) in the library. The conflict resolution algorithm is immediately activated, which compares the sources of the two pieces of knowledge, according to the preset priority rule "enterprise standard requirements are stricter than national standards", and the version timestamp (newly released enterprise standard is superior to the national standard of 2010), automatically makes a decision: marks "3.5 meters" from the enterprise standard as "currently effective (Active)", and marks "3.0 meters" from the national standard as "superseded (Superseded)" but retains it for historical tracing.

[0143] Human-computer collaborative verification: The system pushes this conflict and its automatic processing suggestion to the intelligent human-computer collaborative verification module. The domain expert sees the sources of the two rules and the conflict processing suggestion clearly listed on the interface, confirms that there is no error, clicks "approve", and the new knowledge state is officially written into the knowledge graph.

[0144] Second stage: Traceable intelligent drawing review.

[0145] Scene recognition and query: When the main patent system reviews the drawing, it identifies "220kV live equipment" and its distance to ground "3.2 meters". The intelligent query generation and result tracing module is activated, and the MLLM generates a query accordingly: "Query the effective rule of'minimum safety distance to ground' of '220kV live equipment'".

[0146] Knowledge query and tracing: The knowledge graph executes the query, and since the conflict has been resolved, it directly returns the rule marked as "currently valid" and its complete source information: "Knowledge: '≥3.5 meters', Source: 'Q / GDW12345-2025 Article Y'".

[0147] Compliance judgment and report generation: The MLLM compares the "3.2 meters" identified from the drawing with the authoritative basis returned by the knowledge graph "≥3.5 meters" and determines that it is not compliant. The system finally generates a high-credibility defect report with the following content:

[0148] Conclusion: Serious non-compliance.

[0149] Problem description: The safety distance of 220kV live equipment to ground is insufficient.

[0150] Drawing value: 3.2 meters.

[0151] Required value: ≥3.5 meters.

[0152] Specification basis: Q / GDW12345-2025 Article Y.

[0153] In summary, the knowledge graph-based intelligent review method for power grid drawings provided in this embodiment first preprocesses the multi-source heterogeneous power grid design specification documents to generate standardized structured text, solving the basic problem of different formats and different sources of specifications being difficult to unify, and providing regular data support for subsequent knowledge extraction. Then, based on the field-specific prompt engineering, the multi-modal large language model extracts power grid knowledge triples containing confidence, original text location and source metadata, which not only realizes the structured transformation of specification knowledge, but also retains the association between knowledge and original specifications through metadata. After building the knowledge graph through quality evaluation, conflict resolution and human-machine collaborative verification, the priority rule is used to process conflicting knowledge and combined with expert review to ensure the accuracy and authority of the knowledge in the graph. Finally, in response to the review task, the multi-modal large language model identifies the drawing elements to generate query statements, queries the graph to obtain compliance knowledge, and completes the compliance judgment and generates a report with original specification references, realizing the closed loop from specification knowledge structured construction to intelligent review of drawings.

[0154] To facilitate better implementation of the knowledge graph-based power grid drawing intelligent review method of the embodiments of the present application, the embodiments of the present application further provide a knowledge graph-based power grid drawing intelligent review system based on the above knowledge graph-based power grid drawing intelligent review method. The meanings of the terms are the same as in the above knowledge graph-based power grid drawing intelligent review method, and the specific implementation details can be referred to the description in the method embodiment.

[0155] Please refer to Figure 3 , Figure 3 The structure schematic diagram of the knowledge graph-based power grid drawing intelligent review system provided by the embodiments of the present application is shown in the figure. The knowledge graph-based power grid drawing intelligent review system can specifically include:

[0156] The document preprocessing module 201 is configured to perform preprocessing operations on the multi-source heterogeneous power grid design specification document to generate standardized structured text data.

[0157] Specifically, for the document preprocessing module 201, the module is the entrance of the knowledge pipeline, responsible for receiving and parsing original specification documents in multiple formats such as PDF, Word, and HTML. It integrates advanced OCR and layout analysis technology to extract pure text content and perform chapter segmentation, paragraph alignment, table parsing, and other preprocessing to provide clean and formatted text data for subsequent knowledge extraction.

[0158] The knowledge extraction module 202 is configured to drive a multi-modal large language model based on domain-specific prompt engineering to extract power grid domain knowledge triples including confidence, original text location, and source metadata from the structured text data.

[0159] Specifically, for the knowledge extraction module 202, the module is the engine of automatic knowledge extraction. Instead of using a general MLLM, it is customized through domain-specific prompt engineering (Prompt Engineering). The prompt template is designed to guide the MLLM to perform "thinking chain" reasoning to handle the polysemy of power grid terminology and force it to output knowledge triples in the format containing "[subject | relationship | object | confidence | original text location]". Among them, the "original text location" metadata accurately establishes the mapping relationship from knowledge to the original source of the document, which is the basis for realizing source tracing.

[0160] The knowledge quality evaluation and conflict detection module 203 is configured to perform quality evaluation and conflict detection on the power grid domain knowledge triples, automatically resolve the conflicting knowledge according to the preset priority rules, and obtain a candidate knowledge set.

[0161] Specifically, for the knowledge quality evaluation and conflict detection module 203, the extracted knowledge triples must be processed by this module before being stored in the database. It first performs preliminary screening based on the "confidence" output by the MLLM to filter out low-quality extraction results. More importantly, it contains a knowledge conflict detection algorithm. When a new knowledge has the same subject and relationship as the existing knowledge in the database, but the object is different, the algorithm is triggered. It compares the "source" attributes of the two pieces of knowledge and automatically marks the authoritative knowledge to be adopted and the outdated knowledge to be archived according to the preset priority rules of "national standard > line standard > enterprise standard" and the time sequence rules of "new version > old version".

[0162] The intelligent human-machine collaborative verification module 204 is used for auditing and confirming the candidate knowledge set through a human-machine collaborative verification mechanism to build the power grid design specification knowledge graph.

[0163] Specifically, for the intelligent human-machine collaborative verification module 204, the module provides a high-efficiency auditing interface for domain experts. It not only confirms simply, but also integrates intelligent mechanisms, such as an active learning mechanism, which can analyze the historical correction patterns of experts and actively push the knowledge in the uncertain model or the field frequently corrected by experts to the experts; and a batch verification tool, which can cluster knowledge points with similar structures and the same sources to enable experts to confirm in batches with one key, greatly improving the verification efficiency.

[0164] The knowledge graph of the embodiment supports multi-version time sequence management of knowledge at the storage level, that is, the effective time, invalidation time and version information of each knowledge point are recorded completely, ensuring that the review of historical projects can still call the effective specifications at that time. At the system architecture level, to ensure scalability and maintainability, the system adopts micro-service design, decouples the core functions such as knowledge extraction, storage and query into independent services, and integrates with other systems (such as the main patent review system) through standardized API interfaces

[0165] The scene understanding and query generation module 205 is used for identifying the elements to be reviewed in the power grid drawing to be reviewed through a multi-modal large language model in response to the drawing review task, and generating a structured query statement.

[0166] The compliance judgment and report generation module 206 is used for querying the power grid design specification knowledge graph based on the structured query statement to obtain a query data package, the query data package including compliance knowledge and corresponding metadata; judging the compliance of the power grid drawing to be reviewed based on the query data package, and generating a review report with original specification clause references;

[0167] Specifically, for the scene understanding and query generation module 205 and the compliance judgment and report generation module 206, in the drawing review application end, when the compliance verification needs to be performed, the module is activated. The MLLM in the module will understand the visual scene on the drawing and automatically construct a structured query for the knowledge graph. The result returned by the query is a data package containing the knowledge content itself and all its metadata (especially "original text location" and "version information"). Finally, the defect report generated by the system will explicitly cite these metadata as its "normative basis", realizing the presentation of the complete evidence chain.

[0168] In specific embodiments, as shown in FIG. 6, the embodiments provided by the present application provide an architecture diagram of an MLLM-driven human-machine collaborative knowledge graph construction pipeline, which is a complete architecture for transforming a semi-automated pipeline into a structured knowledge graph. The pipeline is divided into six stages: stage 1 (file input) is responsible for receiving original documents in multiple formats. Stage 2 (preprocessing) is responsible for text extraction and structured parsing. Stage 3 (intelligent extraction) is the core link, and the extraction module driven by the MLLM automatically extracts knowledge triples and their sources. Stage 4 (quality control) is responsible for confidence evaluation and automatic resolution of version / priority conflicts of the extracted knowledge. Stage 5 (expert verification) is an intelligent human-machine collaborative link, and the domain experts perform the final review of the knowledge processed by the machine on an interface with active recommendation and batch processing functions. Only the knowledge confirmed by the experts can be formally written into the final "power grid design specification knowledge graph" with core features such as multi-version management and hierarchical ontology in stage 6 (construction completion). Figure 4 In specific embodiments, as shown in FIG. 6, the embodiments provided by the present application provide an architecture diagram of an MLLM-driven human-machine collaborative knowledge graph construction pipeline, which is a complete architecture for transforming a semi-automated pipeline into a structured knowledge graph. The pipeline is divided into six stages: stage 1 (file input) is responsible for receiving original documents in multiple formats. Stage 2 (preprocessing) is responsible for text extraction and structured parsing. Stage 3 (intelligent extraction) is the core link, and the extraction module driven by the MLLM automatically extracts knowledge triples and their sources. Stage 4 (quality control) is responsible for confidence evaluation and automatic resolution of version / priority conflicts of the extracted knowledge. Stage 5 (expert verification) is an intelligent human-machine collaborative link, and the domain experts perform the final review of the knowledge processed by the machine on an interface with active recommendation and batch processing functions. Only the knowledge confirmed by the experts can be formally written into the final "power grid design specification knowledge graph" with core features such as multi-version management and hierarchical ontology in stage 6 (construction completion).

[0169] Figure 5 In specific embodiments, as shown in FIG. 6, the embodiments provided by the present application provide an architecture diagram of an MLLM-driven human-machine collaborative knowledge graph construction pipeline, which is a complete architecture for transforming a semi-automated pipeline into a structured knowledge graph. The pipeline is divided into six stages: stage 1 (file input) is responsible for receiving original documents in multiple formats. Stage 2 (preprocessing) is responsible for text extraction and structured parsing. Stage 3 (intelligent extraction) is the core link, and the extraction module driven by the MLLM automatically extracts knowledge triples and their sources. Stage 4 (quality control) is responsible for confidence evaluation and automatic resolution of version / priority conflicts of the extracted knowledge. Stage 5 (expert verification) is an intelligent human-machine collaborative link, and the domain experts perform the final review of the knowledge processed by the machine on an interface with active recommendation and batch processing functions. Only the knowledge confirmed by the experts can be formally written into the final "power grid design specification knowledge graph" with core features such as multi-version management and hierarchical ontology in stage 6 (construction completion).

[0170] ​The specific limitations of the power grid drawing intelligent review system based on the knowledge graph can refer to the limitations of the power grid drawing intelligent review method based on the knowledge graph in the above, and will not be repeated here. Each module in the above power grid drawing intelligent review system based on the knowledge graph can be realized by software, hardware and combinations thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls and executes the operations corresponding to each of the above modules.

[0171] The power grid drawing intelligent review system based on the knowledge graph provided in the embodiment replaces a large amount of manual sorting and comparison work through automatic processing, improves the review efficiency and accuracy, and enhances the reliability and interpretability of the review results due to the authority of the knowledge graph and the explicit reference of the original specification in the defect report, thereby promoting the intelligent upgrading of the power grid drawing review.

[0172] In addition, the embodiment of the present application also provides an electronic device, as shown in the Figure 6 The electronic device structure involved in the embodiment of the present application is shown in the

[0173] The electronic device can include a processor 301 with one or more processing cores, a memory 302 with one or more computer readable storage media, a power supply 303, and an input unit 304, and the like. Those skilled in the art can understand that the electronic device structure shown in the Figure 6 The electronic device structure shown in the

[0174] The processor 301 is the control center of the electronic device, which connects various parts of the entire electronic device through various interfaces and lines, executes the software programs and / or modules stored in the memory 302 and the data stored in the memory 302, and processes various functions and data of the electronic device, thereby overall monitoring the electronic device. Optionally, the processor 301 can include one or more processing cores; preferably, the processor 301 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 301.

[0175] The memory 302 can be used to store software programs and modules, and the processor 301 executes various function applications and the power grid drawing intelligent review method based on the knowledge graph by running the software programs and modules stored in the memory 302. The memory 302 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 302 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 302 can also include a memory controller to provide access for the processor 301 to the memory 302.

[0176] The electronic device also includes a power supply 303 for powering various components. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 303 can also include one or more direct current or alternating current power supplies, a recharging system, a power supply failure detection circuit, a power supply converter or inverter, a power supply state indicator, and the like.

[0177] The electronic device can also include an input unit 304, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0178] Although not shown, the electronic device can also include a display unit, etc., which will not be described here. Specifically, in the present embodiment, the processor 301 in the electronic device will load the executable file corresponding to the process of one or more application programs into the memory 302 according to the following instructions, and run the application programs stored in the memory 302 by the processor 301, so as to realize various functions, as follows:

[0179] The pre-processing operation is performed on the multi-source heterogeneous power grid design specification document to generate standardized structured text data; the domain-specific prompt drives the multi-modal large language model to extract power grid domain knowledge triples including confidence, original text location and source metadata from the structured text data; the power grid domain knowledge triples are subjected to quality evaluation and conflict detection, and the conflicting knowledge is automatically resolved according to the preset priority rules to obtain a candidate knowledge set; the candidate knowledge set is audited and confirmed through a man-machine cooperative verification mechanism to construct a power grid design specification knowledge graph; in response to a drawing review task, the multi-modal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed, and a structured query statement is generated; the power grid design specification knowledge graph is queried based on the structured query statement to obtain a query data package, which includes compliance knowledge and corresponding metadata; the power grid drawing to be reviewed is subjected to compliance judgment based on the query data package, and a review report with original specification clause references is generated.

[0180] The specific implementation of each operation can refer to the foregoing embodiments, which will not be repeated here.

[0181] The embodiments of the present application replace a large amount of manual sorting and comparison work through automatic processing, improve the review efficiency and accuracy, and enhance the reliability and explainability of the review results due to the authority of the knowledge graph and the explicit reference of the original specification in the defect report, thereby promoting the intelligent upgrading of the power grid drawing review.

[0182] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or controlled by instructions related to hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0183] To this end, the embodiments of the present application provide a storage medium, which stores a plurality of instructions capable of being loaded by a processor to execute the steps in any one of the power grid drawing intelligent review methods based on a knowledge graph provided by the embodiments of the present application. For example, the instructions can execute the following steps:

[0184] The pre-processing operation is performed on the multi-source heterogeneous power grid design specification document to generate standardized structured text data; the domain-specific prompt is used to drive the multi-modal large language model to extract power grid domain knowledge triples including confidence, original text location and source metadata from the structured text data; the quality of the power grid domain knowledge triples is evaluated and conflict detection is performed, and the conflicting knowledge is automatically eliminated according to the preset priority rules to obtain a candidate knowledge set; the candidate knowledge set is audited and confirmed through a man-machine cooperative verification mechanism to construct a power grid design specification knowledge graph; in response to a drawing review task, the multi-modal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed to generate a structured query statement; the power grid design specification knowledge graph is queried based on the structured query statement to obtain a query data package, and the query data package includes compliance knowledge and corresponding metadata; the compliance of the power grid drawing to be reviewed is judged based on the query data package to generate a review report with original specification clause references.

[0185] The specific implementation of each operation can refer to the foregoing embodiments, which will not be repeated here.

[0186] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0187] Due to the instructions stored in the storage medium, the steps of any one of the power grid drawing intelligent review methods based on the knowledge graph provided by the embodiments of the present application can be performed, and thus the beneficial effects of any one of the power grid drawing intelligent review methods based on the knowledge graph provided by the embodiments of the present application can be achieved. Details are described in the foregoing embodiments, which will not be repeated here.

[0188] The power grid drawing intelligent review method and system based on the knowledge graph provided by the embodiments of the present application are described in detail above, and specific examples are applied to describe the principles and implementation modes of the present application. The above embodiment descriptions are only used to help understand the method and its core idea of the present application; meanwhile, for those skilled in the art, the specific implementation modes and application ranges will be changed according to the idea of the present application, and the above description should not be understood as a limitation of the present application.

Claims

1. A knowledge graph-based intelligent review method for power grid drawings, characterized in that, include: Perform preprocessing operations on the design specification documents of multi-source heterogeneous power grids to generate standardized structured text data; Based on a domain-specific prompting engineering-driven multimodal large language model, this method extracts power grid domain knowledge triples, including confidence level, original text location, and source metadata, from the structured text data. The process includes: constructing a power grid domain-specific prompting template, which includes a power grid terminology dictionary, a mandatory output format definition for knowledge triples, and thought chain reasoning guidance statements. The power grid terminology dictionary contains definitions of polysemous terms, and the thought chain reasoning guidance statements instruct the multimodal large language model to parse the logical relationships of the specification clauses step by step. The structured text data is then split into text fragments by chapter, and each text fragment is concatenated with the prompting template and input into the multimodal large language model. This drives the multimodal large language model to identify core entities and logical relationships in the specification clauses through thought chain reasoning, generating preliminary knowledge triples. The core entities include subjects and objects. The source document identifier, chapter number, and page number information corresponding to the preliminary knowledge triples are associated to complete the source metadata. The confidence level output by the multimodal large language model is bound to the preliminary knowledge triples, source metadata, and original text location to form the power grid domain knowledge triples. The knowledge triples in the power grid domain are subjected to quality assessment and conflict detection. Conflicting knowledge is automatically resolved according to preset priority rules to obtain a set of candidate knowledge. The candidate knowledge set is reviewed and confirmed through a human-machine collaborative verification mechanism to construct a power grid design specification knowledge graph. This includes: analyzing the historical verification records of domain experts based on an active learning mechanism to identify knowledge domains where the multimodal large language model extraction error rate is higher than a preset error threshold, and prioritizing the push of candidate knowledge within these domains to the expert verification interface; clustering knowledge triples with the same source and consistent relationship type in the candidate knowledge set to generate batch verification tasks, and providing an operation entry point on the expert verification interface; obtaining operation data of domain experts confirming or adjusting candidate valid knowledge through the expert verification interface; and writing the approved knowledge triples into the power grid design specification knowledge graph based on the operation data, and recording the effective time, expiration time, and version number of each knowledge triple. In response to the drawing review task, the multimodal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed, and a structured query statement is generated. The power grid design specification knowledge graph is queried based on the structured query statement to obtain a query data package, which includes compliance knowledge and its corresponding metadata. Based on the query data package, the compliance of the power grid drawings to be reviewed is assessed, and a review report with references to the original regulatory clauses is generated.

2. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, The preprocessing operation performed on the multi-source heterogeneous power grid design specification document to generate standardized structured text data includes: Obtain design specification documents for multi-source heterogeneous power grids in different formats, and establish a mapping relationship between the unique identifier of each specification document and the document format type; The text content of the scanned standard document is extracted using optical character recognition technology, and the boundary information of chapter titles, body paragraphs, table areas and formula areas in the standard document is identified and determined by layout analysis algorithm. The extracted text content is hierarchically segmented according to the chapter titles, and the segmented text content is aligned and categorized according to paragraph semantics. The table area is parsed row and column and converted into key-value pair format. Redundant information is removed to generate standardized structured text data.

3. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, The process involves quality assessment and conflict detection of the knowledge triples in the power grid domain, automatically resolving conflicting knowledge according to preset priority rules to obtain a candidate knowledge set, including: Set a preset confidence threshold, remove knowledge triples in the power grid domain with a confidence level lower than the preset confidence threshold, and retain knowledge triples with high confidence. The high-confidence knowledge triples are compared using a knowledge matching algorithm. When the subjects and relations of two high-confidence knowledge triples are completely identical but the objects are different, it is determined that there is a knowledge conflict between the two high-confidence knowledge triples. A preset priority rule is invoked to process high-confidence knowledge triples with knowledge conflicts. Knowledge triples that meet the preset priority rule are marked as candidate valid knowledge, and the remaining knowledge triples with knowledge conflicts are marked as knowledge to be confirmed by experts. The candidate valid knowledge and the knowledge to be confirmed by experts are merged to obtain the candidate knowledge set.

4. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, In response to the drawing review task, the multimodal large language model is used to identify the elements to be reviewed in the power grid drawing to be reviewed, and a structured query statement is generated, including: In response to the drawing review task, the image data of the power grid drawing to be reviewed is input into a multimodal large language model, and the multimodal large language model is used to identify the elements to be reviewed in the power grid drawing; the elements to be reviewed include power grid equipment, equipment parameters, and spatial relationships; The multimodal large language model automatically generates a structured query statement that conforms to the query syntax of the power grid design specification knowledge graph based on the elements to be reviewed. The structured query statement includes the subject to be verified and the relationship to be verified.

5. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, The process of querying the power grid design specification knowledge graph based on the structured query statement yields a query data package, which includes compliance knowledge and its corresponding metadata, including: After the power grid design specification knowledge graph receives the structured query statement, it matches the compliance knowledge that is currently in effect. Extract the metadata corresponding to the compliance knowledge of the effective status, and package the compliance knowledge and the corresponding metadata to form the query data package; the metadata includes the specification name, standard number and specific clause number.

6. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, The process of performing a compliance assessment on the power grid drawings to be reviewed based on the query data packet, and generating a review report with references to the original regulatory clauses, includes: The required values ​​of compliance knowledge are extracted from the query data package, and the actual values ​​of equipment parameters in the power grid drawings to be reviewed are compared with the required values ​​to generate corresponding compliance judgment results. The compliance determination results, the actual values ​​of the equipment parameters, the required values, and the corresponding metadata are integrated to generate the review report; the review report includes conclusions, problem descriptions, drawing values, required values, and reference information to original specification clauses.

7. The intelligent review method for power grid drawings based on knowledge graphs according to claim 1, characterized in that, The power grid design specification knowledge graph adopts a three-tiered architecture including an ontology layer, a data layer, and a service layer. The ontology layer is used to define core concepts in the power grid field and their relationships, and to construct a knowledge system for the power grid field. The data layer is used to store knowledge triples and version information in a time-series database for multi-dimensional retrieval. The service layer is used to provide knowledge query and incremental update services through standardized interfaces, and to achieve integration and docking with the power grid drawing review system.

8. A knowledge graph-based intelligent review system for power grid drawings, characterized in that, include: The document preprocessing module is used to perform preprocessing operations on multi-source heterogeneous power grid design specification documents to generate standardized structured text data; The knowledge extraction module is used to extract power grid domain knowledge triples, including confidence level, original text location, and source metadata, from the structured text data based on a domain-specific prompting engineering-driven multimodal large language model. This includes: constructing a power grid domain-specific prompt template, which includes a power grid terminology dictionary, a forced output format definition for knowledge triples, and thought chain reasoning guidance statements. The power grid terminology dictionary contains definitions of polysemous terms, and the thought chain reasoning guidance statements instruct the multimodal large language model to parse the logical connections of specification clauses step by step; and processing the structured text... The data is split into text fragments by chapter. Each text fragment is then concatenated with the prompt template and input into a multimodal large language model. This model drives the multimodal large language model to identify the core entities and logical relationships in the normative clauses through thought chain reasoning, generating preliminary knowledge triples. The core entities include subjects and objects. The source document identifier, chapter number, and page number information corresponding to the preliminary knowledge triples are associated to complete the source metadata. The confidence level output by the multimodal large language model is bound to the preliminary knowledge triples, the source metadata, and the original text location to form the power grid domain knowledge triples. The knowledge quality assessment and conflict detection module is used to assess the quality and detect conflicts in the knowledge triplet of the power grid domain, and automatically resolve conflicting knowledge according to preset priority rules to obtain a candidate knowledge set. The intelligent human-machine collaborative verification module is used to review and confirm the candidate knowledge set through a human-machine collaborative verification mechanism, and to construct a power grid design specification knowledge graph. This includes: analyzing historical verification records of domain experts based on an active learning mechanism; identifying knowledge domains where the multimodal large language model extraction error rate exceeds a preset error threshold; prioritizing the submission of candidate knowledge within these domains to the expert verification interface; clustering knowledge triples from the same source and with consistent relationship types within the candidate knowledge set to generate batch verification tasks, and providing an operation entry point on the expert verification interface; acquiring operation data from domain experts confirming or adjusting candidate valid knowledge through the expert verification interface; and, based on this operation data, writing the approved knowledge triples into the power grid design specification knowledge graph, and recording the effective date, expiration date, and version number of each knowledge triple. The scenario understanding and query generation module is used to respond to the drawing review task by identifying the elements to be reviewed in the power grid drawing to be reviewed through the multimodal big language model and generating structured query statements. The compliance assessment and report generation module is used to query the power grid design specification knowledge graph based on the structured query statement to obtain a query data package, which includes compliance knowledge and its corresponding metadata; and to perform compliance assessment on the power grid drawings to be reviewed based on the query data package, and generate a review report with references to the original specification clauses.

Images