Power engineering construction file supervision and auditing method and system based on large language model

Abstract

The application discloses a power engineering construction file supervision and auditing method and system based on a large language model, and relates to the technical field of power engineering informatization.The application analyzes a to-be-audited construction file through a multi-modal document analysis tool, extracts content elements such as tables, formulas, title paragraphs, pictures and handwritten annotations in the document, outputs structured content and extracts structured auditing information of the construction file, performs auditing from four dimensions of compliance, integrity, matching degree and risk and logic, controls output through a large language model and reasoning parameters for each auditing dimension, collects auditing results of each dimension after analyzing the output, generates an examination report in a structured form of 'problem point-basis-reform suggestion' and visually displays and pushes the examination report to a business system, and stores auditing operation records in an audit log.The application realizes automatic multi-dimensional auditing and structured examination report generation of power engineering construction files, and improves auditing efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of information technology in power engineering, and in particular to a method and system for supervising and reviewing construction documents for power engineering based on a large language model. Background Technology

[0002] As a crucial link in ensuring the quality and safety of power grid construction, one of the core tasks of power engineering supervision is to professionally review various construction documents (including construction organization designs, specialized construction plans, and construction schedules) submitted by construction units to ensure their compliance, completeness, and feasibility. With the accelerated development of my country's new power system, the scale of power grid projects continues to expand, and their technical complexity is constantly increasing, leading to a sharp rise in the workload of construction document review. However, current construction document review still heavily relies on expert experience and manual interpretation. Reviewers must compare and analyze each item against national and industry standards and internal company standards, making the review process time-consuming and labor-intensive. The quality of the review is easily affected by subjective factors such as the professional level and work status of the personnel, making it difficult to meet the practical demands of "fast, accurate, and rigorous" supervision and review work in the context of high-quality development of power engineering.

[0003] In recent years, with the rapid development of artificial intelligence technology, some research has begun to explore the application of technologies such as deep learning and natural language processing to intelligent document review. For example, Chinese patent application CN121301632A proposes an intelligent review system for engineering industry documents based on a large language model, and Chinese patent application CN120598724A proposes an intelligent review system for building construction plans based on a large model. Both of them utilize the semantic understanding capabilities of language models to automatically review construction documents and generate review reports.

[0004] However, the aforementioned existing technologies still have the following shortcomings: First, they all treat construction documents as plain text and cannot parse the complex layout content such as nested tables, handwritten annotations, and formulas widely present in power engineering construction documents, resulting in incomplete extraction of review information; Second, they all use a single semantic matching or rule matching method to compare the normative clauses, resulting in a single recall method and limited matching accuracy; Third, the review dimensions are limited to general dimensions such as completeness, consistency, accuracy, and compliance, and none of them propose a review method for the matching degree between construction content and technical measures in construction documents, nor can they output a structured review report combining "problem point-basis-rectification suggestion"; Fourth, none of them are specifically designed for the technical specification system unique to power engineering (such as power grid construction standards, power safety regulations, etc.), and they are all independent operating tools, lacking a closed-loop management mechanism for review with the supervision business system.

[0005] In summary, existing technologies require an automated construction document supervision and review solution that can comprehensively analyze multimodal content such as nested tables, handwritten annotations, and formulas in power engineering construction documents, improve the accuracy of matching standard provisions through multi-path recall and fusion sorting mechanisms, and conduct comprehensive review from four dimensions: compliance, completeness, matching degree, risk, and logic. Ultimately, it should form an automated review and review closed-loop management system with the power supervision business system. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention discloses a method and system for supervising and reviewing construction documents for power engineering based on a large language model. This method solves the technical problems of the lack of automated review methods for construction documents in power engineering, the inability to perform multi-dimensional intelligent review of construction documents for compliance and completeness and output structured review reports, and the inability of the review process to form a closed-loop management with the supervision business system.

[0007] To achieve the above objectives, the present invention provides the following solution: On the one hand, a method for supervising and reviewing construction documents for power engineering based on a large language model is provided, and the specific steps include the following: The process involves: acquiring construction documents to be reviewed; performing multimodal document parsing on the documents to extract various content elements and converting the parsed content into a structured format; extracting structured review information from the parsed structured content; reviewing the construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic. Each review dimension uses a prompt word template containing system role definitions and output format requirements to call a large language model, and controls the model output according to set inference parameters. The structured results output by the model are then parsed to generate review results for each dimension; the review results from each dimension are collected and a review report is generated in a structured format of "problem point - basis - rectification suggestion"; the review report is visualized, pushed to the business system, and all operation records of the review process are stored in the audit log.

[0008] Preferably, the multimodal document parsing includes: parsing tables in the document and identifying the hierarchical relationship of nested tables; identifying formulas and converting them into machine-readable format; identifying titles and paragraphs and establishing a hierarchical document structure; extracting text from images using OCR; extracting text from handwritten annotations and establishing a connection between the text and the annotation's position; after parsing, intelligently sorting content elements based on the multimodal document parsing model and outputting a standardized Markdown format file; the table parsing uses a graph neural network-based table parsing model for recursive identification of row and column structures and nesting levels; the text extraction of handwritten annotations uses a model architecture based on convolutional recurrent neural networks and connection-time classification, and the connection between annotations and the text position uses an attention matching method based on position coordinates.

[0009] Preferably, the structured review information includes at least text paragraph information, table information, standard basis information, and handwritten annotation information; the handwritten annotation information is used to compare consistency with the main construction content in the matching degree review, to assist in identifying risk warning information in the risk and logic review, and to add supplementary information to the problem points when the review report is generated.

[0010] Preferably, the compliance review specifically involves: accessing a dedicated knowledge base for power engineering supervision, semantically matching the referenced standard clauses in the construction documents with the standard clauses in the knowledge base, and determining whether the content of the construction documents conforms to the corresponding standard requirements. The semantic matching employs a multi-path recall method combining vector retrieval and keyword retrieval, and the matching result is determined after fusing and sorting the vector retrieval results and keyword retrieval results. The vector retrieval uses a semantic embedding model to encode the standard clause text into dense vectors, and performs similarity retrieval in an approximate nearest neighbor index structure. The keyword retrieval uses a retrieval algorithm based on term frequency-inverse document frequency. The fusion sorting uses a reciprocal fusion sorting or weighted fusion sorting algorithm, and the similarity score of the vector retrieval and the relevance score of the keyword retrieval are normalized and fused, and then sorted from high to low scores. The knowledge base stores at least national and power industry engineering construction standards, power grid enterprise technical specifications, safety operating procedures, and enterprise internal management systems.

[0011] Preferably, the integrity audit specifically involves: matching the corresponding predefined audit checklist according to the document type, and checking whether each key module is included in the construction document one by one; the matching degree audit specifically involves: extracting the construction content information and technical measure information from the construction document, using a large language model for semantic understanding, and determining whether each technical measure covers the corresponding requirements of each stage of the construction content.

[0012] Preferably, the risk and logic review specifically involves: using a large language model to perform risk semantic analysis on the construction document text to identify potential safety risks, verifying the logical consistency between different chapters of the construction document, and verifying the accuracy of key calculations in the construction document.

[0013] Preferably, the inference parameters include temperature parameters, kernel sampling parameters, and maximum output length parameters. The prompt word templates are configured according to different review tasks in four dimensions: compliance review, integrity review, matching degree review, and risk and logic review. The structured results output by the model inference are processed by a parser to extract fields. If the output format does not meet expectations, a retry mechanism is triggered.

[0014] On the other hand, a power engineering construction document supervision and review system based on a large language model is provided, including an application layer, a core function layer, and a deployment layer. The application layer includes a graphical configuration interface; the core function layer includes a document parsing module, a multi-dimensional review module, a review report generation module, a knowledge base management module, and a system interaction module; the deployment layer includes a training and promotion integrated machine and a business system interface. The document parsing module is used to perform multi-modal parsing of construction document documents and output structured content; the knowledge base management module is used to store a dedicated knowledge base for power engineering supervision; the multi-dimensional review module is used to review construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic. The review process involves each review dimension calling upon a large language model deployed on the training and inference integrated machine using prompt word templates that include system role definitions and output format requirements. The model output is controlled according to set inference parameters, and the structured output is parsed to generate review results for each dimension. The review report generation module generates review reports in a structured format of "problem point - basis - rectification suggestion." The system interaction module enables AI-driven business system data interaction and provides security control over data modification operations. The training and inference integrated machine provides training and inference computing power support for the large language model. The business system interface communicates with the integrated supervision management platform and the Supervision Treasure APP.

[0015] Preferably, the document parsing module parses the mixed text, image, and table layout format based on the multimodal document parsing big model, intelligently converts the content into a reading order, and provides specialized parsing capabilities for the unique table nesting and handwritten annotation format of construction documents.

[0016] Preferably, the system interaction module includes a security authentication unit and an operation traceability unit. The security authentication unit is used to implement hierarchical access control, and the operation traceability unit is used to log the entire data modification process, and the logs are tamper-proof.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention realizes automated review of construction documents for power engineering projects. By constructing a dedicated knowledge base for power engineering supervision and combining it with the semantic understanding capabilities of a large language model, it achieves automated compliance judgment of construction documents for power engineering projects, filling the gap in dedicated automated review solutions in the field of power engineering supervision.

[0018] 2. This invention enables multi-dimensional intelligent review of construction documents and outputs structured review reports. It comprehensively reviews construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic, and generates review reports in a structured form of "problem point - basis - rectification suggestion", making the review results clear and operable.

[0019] 3. This invention achieves deep integration and closed-loop management of the audit process with the power supervision business system. Through the business system interface, it connects with the integrated supervision management platform and the Supervision Treasure APP to realize the automatic acquisition of audit tasks, automatic push of results, and automatic archiving of operation logs. Through hierarchical access control and tamper-proof audit logs, a complete closed-loop audit management is formed, which meets the compliance requirements of the power supervision industry for traceability and auditability of the audit process. Attached Figure Description

[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention. Various modifications and equivalent substitutions made by those skilled in the art without departing from the spirit and scope of the present invention should be considered within the scope of protection of the present invention.

[0021] Figure 1 This is a diagram illustrating the overall architecture of the power engineering construction document supervision and review system based on a large language model, as described in this invention.

[0022] Figure 2 This is an overall flowchart of the power engineering construction document supervision and review method based on a large language model, as described in this invention.

[0023] Figure 3 This is a flowchart of the multimodal document parsing process of the present invention.

[0024] Figure 4 This is a flowchart of the four-dimensional review process for this invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited thereto.

[0026] Example 1 Please see Figures 1 to 4This embodiment provides a method for supervising and reviewing construction documents for power engineering based on a large language model. The method is applied to a supervision AI large model review platform to automatically review various construction documents submitted by power engineering construction units.

[0027] Figure 2 This embodiment illustrates the steps of the power engineering construction document supervision and review method based on a large language model. (See flowchart below.) Figure 2 As shown, the method includes the following steps S1-S6.

[0028] S1. Obtain the construction documents to be reviewed, which includes the following steps: S11. Obtain the construction documents to be reviewed through the business system interface. The business system interface is connected to the integrated supervision management platform and / or the Supervision Treasure APP. Users select the construction documents to be reviewed through a graphical configuration interface (review dashboard). The types of construction documents include construction organization design, special construction plan, construction schedule, and other construction documents.

[0029] S12. Preprocess the acquired construction documents. The preprocessing includes identifying the document type, extracting document metadata (including document name, document type, submission time, submitting unit, and compiler), and matching the corresponding review rule template according to the document type.

[0030] Step S1 enables the automatic acquisition and preprocessing of construction documents to be reviewed, providing standardized input data for subsequent document parsing and review processes, and ensuring the accurate initiation of review tasks and the correct identification of document types.

[0031] S2. Perform multimodal document parsing on construction documents. Figure 3 This embodiment illustrates a multimodal document parsing flowchart for the power engineering construction document supervision and review method S2 based on a large language model. (See attached flowchart.) Figure 3 As shown, the multimodal document parsing process includes the following steps S21-S25.

[0032] S21. Call the document content parsing tool to obtain the construction document to be parsed. The document content parsing tool supports parsing various file formats, including pdf, doc, docx, ppt, mhtml, pptx (single file not exceeding 100MB), xlsx, xls, md, txt, csv, html (single file not exceeding 20MB), and jpg, png, jpeg image files (single file not exceeding 50MB).

[0033] S22. Identify and classify document types. The document content parsing tool automatically identifies document types based on file extensions and header information, and selects the corresponding parsing strategy.

[0034] S23. Extract various content elements from the document, specifically: 1) Table Parsing: Identifies the table structure in the document, including row, column, cell boundaries, and merging relationships. For nested tables, the hierarchical relationship is preserved layer by layer during parsing. A nested table refers to a complex table structure where cells within a table contain sub-tables. During parsing, the row and column structure of the outer table is identified first, followed by the structure of the inner sub-tables layer by layer, ultimately generating a complete tree-like description of the table structure. The table structure identification employs a graph neural network-based table parsing model, treating cells as nodes in the graph and adjacency relationships as edges. A message passing mechanism is used to identify row and column structures and merging relationships. Nested parsing uses a recursive detection strategy, checking layer by layer whether a table cell contains a sub-table region and recursively calling the structure identification process.

[0035] 2) Formula recognition: Recognizes mathematical formulas and logical expressions in documents and converts them into machine-readable formats.

[0036] 3) Title and paragraph recognition: Identify the title hierarchy and body paragraphs of a document to establish a hierarchical structure of the document.

[0037] 4) Image and Handwritten Annotation Recognition: Image content is extracted from the document, and text information within the images is recognized using OCR technology. For handwritten annotations, a handwriting recognition model is used to extract the text, and the handwritten content is associated with the corresponding text content at the annotation location. The handwritten text recognition employs a model architecture based on Convolutional Recurrent Neural Network (CRNN) and Connectionist Temporal Classification (CTC). The association between handwritten content and the text uses a positional coordinate-based attention matching method. By calculating the association probability between the annotation location and each text content block location, the annotation content is associated with the text content block with the highest association probability. If the maximum association probability is lower than a preset threshold, the annotation is marked as an independent annotation.

[0038] S24. Intelligently sort the extracted content elements according to the reading order.

[0039] Based on a large multimodal document parsing model, the position and order information of each content element is generated in a coarse-grained manner. The semantic awareness of the context is given by the content generation method, and the sorting results are corrected to ensure that the reading order of complex layout formats such as nested tables and mixed text and images conforms to the original logical structure of the document.

[0040] The multimodal document parsing model employs a multimodal neural network architecture that integrates text, layout, and image information. It includes: a text encoding layer for encoding the text sequence obtained from OCR recognition; a layout encoding layer for encoding the position coordinates and size of each text block on the page; an image encoding layer for extracting visual features from the document page image; and a cross-modal attention fusion layer that fuses the outputs of the three encoding layers to output the type label and position coordinates of each content element. The model is trained on a labeled dataset of power engineering construction documents. The labeled data was completed by power engineering supervision experts, and the labeled content includes table areas and nested levels, formula areas, title levels, image areas, handwritten annotation areas, and the relationship between annotations and the main text.

[0041] S25. Output a standardized Markdown file. Convert all sorted content elements to Markdown format, preserving the formatting and hierarchy of tables, formulas, image references, headings, paragraphs, headers, footers, and other content elements.

[0042] Step S2 enables comprehensive parsing and structured output of multimodal content such as tables, formulas, headings, images, and handwritten annotations in power engineering construction documents. This solves the problem that existing technologies cannot parse complex layouts such as nested tables and handwritten annotations, providing a complete and accurate structured data foundation for subsequent review.

[0043] S3. Extract structured review information from construction documents, specifically including the following steps: S31. Semantic segmentation of Markdown content. Based on the document hierarchy (heading hierarchy) established in step S2, the document content is divided into several semantic blocks. Each semantic block contains a heading and its corresponding body text; for table content, the table heading and table data are treated as a single semantic block; for pages containing handwritten annotations, the body text of the annotation area and the annotation content together constitute a single semantic block. Each semantic block is then categorized by information type. Based on preset information type classification rules, combined with the heading hierarchy and content characteristics within the semantic block, the information type of each semantic block is determined.

[0044] S32. Based on the classification results in S31, extract the structured audit information for each type, including at least: 1) Text paragraph information: Extract text paragraphs from the main text that contain key content such as construction methods, technical measures, safety measures, quality assurance measures, and schedule plans, and mark their positions in the original text; 2) Table information: Extract table data such as resource plan, schedule plan, and personnel allocation table from the construction documents, and retain the row and column structure and numerical information of the tables; 3) Standards and specifications information: Extract the clause numbers and contents of national and industry standards and enterprise technical standards referenced in the construction documents; 4) Handwritten annotation information: Extract the text content of handwritten annotations and their corresponding location information from the construction document.

[0045] The extracted handwritten annotation information is used in subsequent reviews: (1) In the matching review, the modification opinions of construction methods or technical measures involved in the annotation are compared with the construction content in the main text; (2) In the risk identification of risk and logic review, the risk warning information in the annotation content is identified as an auxiliary judgment basis; (3) When the review report is generated, if there is a handwritten annotation in the main text corresponding to the problem point, the annotation content is attached to the problem description.

[0046] Step S3 enables the precise extraction of structured information required for review, such as text paragraphs, tables, standard references, and handwritten annotations, from the parsed structured content. The information is then categorized and labeled, providing data support for the four-dimensional review process. Furthermore, the association between handwritten annotations and the main text content is established, ensuring the effective use of handwritten annotation information in subsequent review stages.

[0047] S4. Review the construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic. Figure 4 This embodiment illustrates a four-dimensional review flowchart for the power engineering construction document supervision and review method S4 based on a large language model. (See attached flowchart.) Figure 4 As shown, the four-dimensional audit process includes the following steps S41-S44.

[0048] The large language model used in this invention is a general-purpose large language model based on the Transformer decoder architecture, with a parameter scale of no less than 70 bytes, deployed on a training and inference integrated machine. The following parameters are configured during inference: a temperature parameter (temperature) controls output randomness, a kernel sampling parameter (top_p) controls the size of the candidate word set, and a maximum output length (max_tokens) controls the maximum number of tokens generated in a single iteration. Each review dimension uses a dedicated prompt word template for model inference, which includes system role definitions, specific instruction formats, and output format requirements. The structured results output by the model inference are processed by a parser to extract fields; if the output format does not meet expectations, a retry mechanism is automatically triggered.

[0049] S41. Compliance Audit. Extract the regulatory basis information and text paragraph information from the structured audit information obtained in step S3, and call the power engineering supervision-specific knowledge base. The knowledge base stores at least national and power industry engineering construction standards, power grid enterprise technical specifications, safety operating procedures, and enterprise internal management systems. The regulatory clauses in the knowledge base are stored in vectorized form and have a vector index.

[0050] Preferably, the vector retrieval employs a semantic embedding model based on the Transformer architecture to encode the specification clause text into dense vectors, the vector index uses a graph-based approximate nearest neighbor index structure, and the similarity metric uses cosine similarity or inner product similarity. The keyword retrieval uses a retrieval algorithm based on term frequency-inverse document frequency (such as the BM25 algorithm). The fusion ranking uses a reciprocal fusion ranking or weighted fusion ranking algorithm to normalize and fuse the semantic similarity score of the vector retrieval with the text relevance score of the keyword retrieval, and determines the final matching result according to the fusion score from high to low.

[0051] The process involves semantically matching the referenced standard clauses in the construction documents with standard clauses in the knowledge base. The content of the referenced standard clauses in the construction documents is encoded into query vectors. An approximate nearest neighbor search is then performed on the vector index of the knowledge base to recall semantically similar standard clauses. Simultaneously, the standard clause numbers referenced in the construction documents are extracted and matched precisely with the clause numbers in the knowledge base. The vector search results and precise matching results are then merged and sorted to obtain the standard clauses that match the referenced clauses in the construction documents.

[0052] Determine whether the content of the construction documents complies with the corresponding specifications. Perform a semantic comparison between the matched specification clauses and the specific measures in the construction documents to determine whether the technical measures, safety measures, and construction methods proposed in the construction documents meet the requirements of the specification clauses. If the measures in the construction documents meet the specification requirements, mark the audit item as "compliant"; if not, mark it as "non-compliant" and record the relevant specification clauses.

[0053] S42. Integrity Review. Extract text paragraph information from the structured review information obtained in step S3 and match it to the corresponding integrity review checklist based on the document type. The integrity review checklist is predefined according to the requirements of the power engineering supervision specifications. The construction method of the review checklist is as follows: The initial checklist is compiled by the power engineering supervision expert team based on the "Construction Engineering Supervision Specification" (GB / T 50319), the "Power Construction Engineering Supervision Specification" (DL / T 5434), and the company's internal supervision work standards for each document type (construction organization design, special construction plan, construction schedule, etc.). Each checklist item includes an item number, item name, item description, importance level (high / medium / low), and corresponding specification basis. The system provides a review checklist management interface, supporting users with administrator privileges to add, edit, delete, and manage versions of the checklist. Each modification operation records the operator, modification time, and version number. The document type matching mechanism works as follows: After identifying the document type during the preprocessing stage, the system queries the pre-configured document type-audit checklist mapping table using the document type name as the key, obtains the corresponding audit checklist ID, and loads the checklist. If the document type is not configured in the mapping table, the system defaults to using the "general audit checklist" for integrity checks. For example, the integrity audit checklist for construction organization design includes key modules such as project overview, construction deployment, construction schedule, construction preparation and resource allocation plan, main construction schemes, construction site layout, quality assurance measures, safety management measures, civilized construction and environmental protection measures, and seasonal construction measures. The integrity audit checklist for specialized construction schemes includes key modules such as the basis for preparation, project overview, construction plan, construction technology, quality assurance measures, safety assurance measures, emergency plans, calculation sheets, and related drawings.

[0054] Check each construction document to ensure it contains all the key modules from the audit checklist. Use semantic matching to determine if the construction document text covers the thematic content corresponding to each key module in the audit checklist. For included modules, mark them as "complete"; for missing modules, mark them as "missing" and record the module's name and importance level.

[0055] S43. Matching Verification. Extract text paragraphs and table information from the structured verification information obtained in step S3, and simultaneously extract construction content information described in the construction documents. The construction content information includes at least one or more of the following: construction procedures, construction methods, construction process parameters, construction period, and resource input. Also extract information on technical measures described in the construction documents. The technical measures information includes at least one or more of the following: safety measures, quality measures, environmental protection measures, civilized construction measures, and seasonal construction measures.

[0056] The consistency of construction content information and technical measures information is compared. A large language model is used to perform semantic understanding of the construction content information and technical measures information to determine whether each technical measure covers the corresponding requirements of each stage of the construction content. If the technical measures fully cover the requirements of each stage of the construction content, the audit item is marked as "matched"; if there is a lack of corresponding measures in a construction stage or the measures are inconsistent with the construction stage, it is marked as "mismatched" and the specific details of the mismatch are recorded.

[0057] S44. Risk and Logic Review. Extract text paragraphs and table information from the structured review information obtained in S3, and then conduct a risk and logic review, specifically including: Risk Identification: A large language model is used to perform risk semantic analysis on the construction document text to identify potential safety and quality risks. These risks include one or more common risk types in power engineering construction, such as risks associated with working at heights, deep foundation pit collapses, temporary power supply, hoisting and lifting operations, fire, electric shock, falling objects, mechanical injuries, and poisoning from harmful gases.

[0058] Logical consistency check: This checks whether there are logical contradictions between different parts of the construction document text. These contradictions include mismatches between the construction schedule and resource allocation plan, inconsistencies in the description of construction methods, discrepancies between safety measures and construction process requirements, and mismatches between emergency plans and risk identification results.

[0059] Calculation accuracy verification: Verify the accuracy of key calculations involved in the construction documents. These key calculations include construction load calculations, hoisting stress calculations, temporary structure stability verifications, and electrical parameter calculations. Verification is performed by utilizing the numerical reasoning capabilities of a calculation engine or a large language model.

[0060] If no risks, vulnerabilities, logical contradictions, or calculation errors are identified in the construction documents, mark the audit item as "passed"; if the above problems exist, mark it as "failed" and record the problem type and specific details.

[0061] Step S4 involves a comprehensive review of the construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic. By combining the power engineering supervision-specific knowledge base and the semantic understanding capabilities of the large language model, a comprehensive judgment is made on the compliance of the construction documents with specifications, the completeness of the content, the matching degree of measures, and the consistency of safety risks and logic, which effectively improves the comprehensiveness and accuracy of the review.

[0062] S5. Generate a structured review report. Compile the review results from steps S41 to S44 for each dimension, and generate a review report in a structured format of "Issue Point - Basis - Rectification Suggestions". For each dimension's review items marked as "Non-compliant", "Missing", "Mismatch", or "Failed", generate corresponding issue point entries. Each issue point entry includes: 1) Problem Description: Describe the specific problems existing in the construction documents in concise language; 2) Source of the issue: Indicate whether the issue originated from a compliance review, completeness review, matching degree review, or risk and logic review; 3) Problem Location: Mark the chapter, page number, or paragraph of the construction document where the problem occurs; 4) Based on the clauses: For compliance issues, list the name of the relevant standard and clause number; for completeness issues, list the names of the key modules in the audit checklist; for matching degree and risk logic issues, list the relevant enterprise standards or industry practices. 5) Suggestions for improvement: Provide specific suggestions for improvement and reference examples for this issue.

[0063] Generate a complete review report document, specifically including the review report document output in Word, PDF or HTML format, including report title, basic review information (including construction document name, review time, reviewer), issue summary table and detailed description of each issue.

[0064] Step S5 compiles the audit results from each dimension into a structured report in the form of "problem point - basis - rectification suggestion". This makes the audit results clear and operable, providing clear modification guidelines for construction document preparers and effectively solving the problem that existing technologies cannot output structured review reports.

[0065] S6. Output and archiving of audit results, specifically including visually displaying the review report through a graphical configuration interface (audit dashboard); pushing the audit results and review report to the business system (integrated supervision management platform and / or Supervision Treasure APP) for construction document preparation personnel to view and modify; and storing all operation records of the audit process in the audit log. The audit log shall at least include the operation initiator (account / name), operation time, name of the audited construction document, audit result, and report version, and the log shall be tamper-proof to meet the audit traceability requirements.

[0066] Step S6 enables the visualization of audit results, data exchange with business systems, and tamper-proof storage of the entire audit process operation record. This ensures a complete closed-loop management of audit task initiation, result push, report archiving, and audit traceability, solving the technical problem in existing technologies where the audit process cannot form a closed-loop management with the supervision business system.

[0067] In summary, the power engineering construction document supervision and review method based on a large language model provided in Example 1, through the coordinated operation of steps S1 to S6, achieves fully automated construction document supervision and review throughout the entire process, from document acquisition, multimodal parsing, structured information extraction, four-dimensional intelligent review, structured review report generation to review result output and closed-loop management.

[0068] Example 2 Please see Figure 1 This embodiment provides a power engineering construction document supervision and review system based on a large language model. The system supports private deployment and is deployed in the internal network of power engineering supervision enterprises. It achieves data interoperability with the enterprise's existing integrated supervision management platform and Supervision Treasure APP through business system interfaces, and is used to automatically review various construction documents submitted by power engineering construction units.

[0069] like Figure 1 As shown, the system includes: application layer 1, core function layer 2, and deployment layer 3.

[0070] The application layer 1 includes a graphical configuration interface 11.

[0071] The core functional layer 2 includes a document parsing module 21, a multi-dimensional review module 22, a review report generation module 23, a knowledge base management module 24, and a system interaction module 25.

[0072] The deployment layer 3 includes a training and promotion integrated machine 31 and a business system interface 32.

[0073] The training and inference integrated machine 31 of the deployment layer 3 is communicatively connected to each module of the core functional layer 2, and is used to provide training and inference computing power support for large language models. The business system interface 32 of the deployment layer 3 is connected to the integrated supervision management platform and the Supervision Treasure APP respectively, and is used to obtain the construction documents to be reviewed and push the review results back to the business system; The document parsing module 21 of the core functional layer 2 is communicatively connected to the graphical configuration interface 11. It is used to receive construction document uploaded by the user, perform multimodal parsing on the construction document, and output the parsed content in Markdown format. The knowledge base management module 24 of the core functional layer 2 is communicatively connected to the multi-dimensional review module 22 and is used to provide the multi-dimensional review module 22 with semantic matching query services for the normative clauses; The multi-dimensional audit module 22 of the core functional layer 2 is communicatively connected to the document parsing module 21 and the knowledge base management module 24, respectively. It is used to obtain the structured content of the construction documents from the document parsing module 21 and the standard clauses from the knowledge base management module 24, and to conduct compliance audit, completeness audit, matching degree audit and risk and logic audit on the construction document documents. The review report generation module 23 of the core functional layer 2 is communicatively connected to the multi-dimensional review module 22, and is used to collect the review results of each dimension and generate a review report in a structured form of "problem point-basis-rectification suggestion". The system interaction module 25 of the core functional layer 2 is communicatively connected to the graphical configuration interface 11, and is used to realize AI command-driven business system data interaction and to perform security control on data modification operations. The graphical configuration interface 11 of the application layer 1 is communicatively connected to the document parsing module 21, the review report generation module 23, and the system interaction module 25 of the core function layer 2, and is used to allow users to upload construction documents to be reviewed, view review results, export review reports, and view audit logs.

[0074] The document parsing module 21 supports parsing various file formats, including pdf, doc, docx, ppt, mhtml, pptx, xlsx, xls, md, txt, csv, html, jpg, png, and jpeg. Based on a multimodal document parsing model, the module generates the position and order of content elements with coarse granularity, supplemented by semantic awareness of contextualized content generation, parsing mixed text, image, and table layouts and intelligently converting the content into a reading order. The module provides specialized parsing capabilities for the unique table nesting and handwritten annotation formats in construction documents: for table nesting, it uses a nested table hierarchy recognition model to parse the structure of outer and inner sub-tables layer by layer, generating a complete table tree structure description; the nested table hierarchy recognition model adopts a graph attention network architecture, using cells as nodes and adjacency relationships as edges to identify row and column structures, and uses a recursive detection strategy to achieve layer-by-layer parsing of nested levels. The model training data comes from manually annotated nested table samples in power engineering construction documents, totaling 20,000 table image annotations. For handwritten annotations, text is extracted using a handwritten character recognition model, and a positional association is established between the handwritten content and the text content at the annotation location. The handwritten character recognition model adopts a convolutional recurrent neural network and connection-temporal classification (CRNN+CTC) architecture. The training data includes the publicly available handwritten dataset IAM and 50,000 self-collected handwritten annotation samples from power engineering projects. The association between annotations and text is achieved using a positional alignment method based on an attention mechanism.

[0075] The multi-dimensional audit module 22 is used to conduct compliance audits, completeness audits, matching degree audits, and risk and logic audits on construction documents. The compliance audit determines whether the content of the construction documents conforms to the corresponding regulatory requirements by semantically matching the normative clauses referenced in the construction documents with normative clauses in the knowledge base. This semantic matching employs a multi-path recall method combining vector retrieval and keyword retrieval. The completeness audit checks whether each key module is included in the construction documents by matching them to a predefined audit checklist based on the document type. The matching degree audit extracts construction content information and technical measure information from the construction documents, uses a large language model for semantic understanding, and determines whether the technical measures fully cover the needs of each stage of the construction content. The risk and logic audit uses a large language model to perform risk semantic analysis on the construction document text to identify potential safety risks, verify the logical consistency between different chapters of the construction documents, and verify the accuracy of key calculations in the construction documents.

[0076] The knowledge base management module 24 is used to store and manage a dedicated knowledge base for power engineering supervision. The knowledge base stores at least national and power industry engineering construction standards, power grid enterprise technical specifications, safety operating procedures, and internal management systems. The knowledge base management module 24 supports adding, adding / removing, viewing details, and managing documents in the knowledge base. It supports uploading files from local storage and / or FTP, managing document tags, automatically generating question-and-answer pairs based on documents, and automatically extracting and summarizing document abstracts. The knowledge base management module 24 supports storing normative clauses in the knowledge base in vector form and establishing a vector index. Specifically, during knowledge base construction, the knowledge base is divided into blocks based on normative clauses. Each clause text is encoded into a dense vector using a semantic embedding model, and an index is built on the vectors using an approximate nearest neighbor index structure to achieve semantic similarity retrieval of normative clauses. When a new normative document is imported, it is automatically parsed, divided into blocks, and vectorized, incrementally updated to the existing index, and semantic matching queries of normative clauses are achieved through vector retrieval.

[0077] The review report generation module 23 is used to aggregate the review results from various dimensions and generate a review report in a structured format of "problem point - basis - rectification suggestion". Each problem point includes at least the fields of problem description, problem source, problem location, basis clause, and rectification suggestion. The review report is output in Word, PDF, or HTML format.

[0078] The system interaction module 25 includes a security authentication unit and an operation traceability unit. The security authentication unit implements hierarchical access control, classifying operation permissions by role and restricting unauthorized personnel from accessing and / or modifying core data. The operation traceability unit logs the entire data modification process. The logs include at least the operation initiator (account / name), operation time, content before modification, content after modification, and operation type (add / edit / delete). The logs support filtering and querying by time, operator, and data type, and are tamper-proof, meeting audit traceability requirements.

[0079] The training and inference integrated machine 31 is a domestically produced inference integrated machine, using domestically produced hardware and software platforms. It has the capability to manage, monitor, and schedule large models throughout their entire lifecycle, including large model management, large model data management, large model training, large model inference, large model training and inference computing resources, and other functions. The parameter scale of the large language model is ≥70B. The configuration requirements of the training and inference integrated machine 31 are as follows: ≥2 CPUs, each CPU having ≥64 cores; ≥1536GB of memory; ≥8 computing power modules, with a single module computing power of ≥560TOPS@INT8 and 280TFlops@FP16, a single card having ≥96GB of video memory, and a single card power consumption of no more than 400W; ≥2×480GB of storage + 6×3.84TB of storage; and a network card with at least two gigabit Ethernet ports and two 25GE optical ports.

[0080] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for supervising and reviewing construction documents for power engineering based on a large language model, characterized in that, Includes the following steps: Obtain the construction documents pending review; The construction document is parsed using a multimodal method to extract various content elements and convert the parsed content into a structured format for output. Extract structured review information from the parsed structured content of the construction documents; The construction documents are reviewed from four dimensions: compliance, completeness, matching degree, and risk and logic. Each review dimension calls the large language model through prompt word templates that include system role definitions and output format requirements, and controls the model output according to the set inference parameters. The structured results of the model output are parsed and the review results of each dimension are generated respectively. The audit results from various dimensions are aggregated and a review report is generated in a structured format of "problem point - basis - rectification suggestion". The review report is then visualized, pushed to the business system, and all operation records of the audit process are stored in the audit log.

2. The method according to claim 1, characterized in that, The multimodal document parsing includes: parsing tables in the document and identifying the hierarchical relationship of nested tables; recognizing formulas and converting them into machine-readable format; recognizing titles and paragraphs and establishing a hierarchical document structure; extracting text from images using OCR; extracting text from handwritten annotations and establishing a connection between the text and the annotations; and after parsing, intelligently sorting the content elements based on the multimodal document parsing model and outputting a standardized Markdown format file.

3. The method according to claim 1, characterized in that, The structured review information includes at least text paragraph information, table information, standard basis information, and handwritten annotation information; the handwritten annotation information is used to compare consistency with the main construction content in the matching degree review, to help identify risk warning information in the risk and logic review, and to add supplementary information to the problem points when the review report is generated.

4. The method according to claim 1, characterized in that, The compliance review specifically involves: accessing a dedicated knowledge base for power engineering supervision, semantically matching the referenced standard clauses in the construction documents with the standard clauses in the knowledge base, and determining whether the content of the construction documents complies with the corresponding standard requirements. The semantic matching employs a multi-path recall method combining vector retrieval and keyword retrieval, and the matching result is determined after fusing and sorting the vector retrieval results and keyword retrieval results. The vector retrieval uses a semantic embedding model to encode the standard clause text into dense vectors, and performs similarity retrieval in an approximate nearest neighbor index structure. The keyword retrieval uses a retrieval algorithm based on term frequency-inverse document frequency. The fusion sorting uses a reciprocal fusion sorting or weighted fusion sorting algorithm, and the similarity score of the vector retrieval and the relevance score of the keyword retrieval are normalized and fused, and then sorted from high to low scores. The knowledge base stores at least national and power industry engineering construction standards, power grid enterprise technical specifications, safety operating procedures, and internal management systems of enterprises.

5. The method according to claim 1, characterized in that, The integrity audit specifically involves: matching the predefined audit checklist according to the document type, and checking whether each key module is included in the construction document one by one; the matching degree audit specifically involves: extracting the construction content information and technical measure information from the construction document, using a large language model for semantic understanding, and determining whether each technical measure covers the corresponding requirements of each stage of the construction content.

6. The method according to claim 1, characterized in that, The risk and logic review specifically involves: using a large language model to perform risk semantic analysis on the construction document text to identify potential safety risks, verifying the logical consistency between different chapters of the construction document, and verifying the accuracy of key calculations in the construction document.

7. The method according to claim 1, characterized in that, The inference parameters include temperature parameters, kernel sampling parameters, and maximum output length parameters. The prompt word templates are configured according to different review tasks in four dimensions: compliance review, integrity review, matching degree review, and risk and logic review. The structured results output by the model inference are processed by a parser to extract fields. If the output format does not meet expectations, a retry mechanism is triggered.

8. A power engineering construction document supervision and review system based on a large language model, characterized in that, include: The application layer includes a graphical configuration interface; The core functional layer includes a document parsing module, a multi-dimensional review module, a review report generation module, a knowledge base management module, and a system interaction module. The deployment layer includes a training and promotion integrated machine and a business system interface; The document parsing module is used to perform multimodal parsing on construction documents and output structured content; The knowledge base management module is used to store a dedicated knowledge base for power engineering supervision. The multi-dimensional audit module is used to audit construction documents from four dimensions: compliance, completeness, matching degree, and risk and logic. Each audit dimension calls the large language model deployed on the training and push integrated machine through prompt word templates containing system role definitions and output format requirements, and controls the model output according to the set inference parameters. After parsing the structured output results, the audit results of each dimension are generated. The review report generation module is used to generate review reports in a structured format of "problem point - basis - rectification suggestion"; The system interaction module is used to realize data interaction of business systems driven by AI commands and to perform security control on data modification operations; The training and inference integrated machine is used to provide training and inference computing power support for large language models; The business system interface is used to communicate with the integrated supervision management platform and the Supervision Treasure APP.

9. The system according to claim 7, characterized in that, The document parsing module is based on a multimodal document parsing model to parse the layout of mixed text, images, and tables, intelligently converting the content into a reading order, and providing specialized parsing capabilities for the unique table nesting and handwritten annotation formats of construction documents.

10. The system according to claim 7, characterized in that, The system interaction module includes a security authentication unit and an operation traceability unit. The security authentication unit is used to implement hierarchical control of permissions, and the operation traceability unit is used to log the entire process of data modification, and the logs are tamper-proof.

Images