Knowledge system construction and retrieval method of AI auxiliary compilation system for special construction scheme of dangerous and large project

Abstract

The invention provides a dangerous large project special construction scheme AI auxiliary compilation system knowledge system construction and retrieval method, which comprises the steps of S1, multi-dimensional knowledge base construction: constructing a multi-dimensional knowledge base for storing dangerous large project compilation rules, specification standards, enterprise standards and project knowledge; s2, establishing a multi-level knowledge graph, establishing the dangerous and large project construction multi-level knowledge graph, and enabling the multi-dimensional knowledge base to form a knowledge system; s3, intelligently analyzing the multi-type documents; s4, knowledge management system construction, wherein the construction steps comprise knowledge classification; importing knowledge; automatic label prediction; and S5, constructing a multi-level knowledge representation and retrieval model. According to the method, on the basis of fully researching the current situation of scheme compilation, the knowledge characteristics in the building field and the bottleneck of the current large language model, the targeted knowledge system construction and efficient retrieval method is realized, and the technical problems of knowledge system construction and efficient retrieval and utilization of the AI auxiliary compilation system for the special construction scheme of the dangerous and large project are solved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary integration of building construction technology and artificial intelligence technology, and in particular relates to a knowledge system construction and efficient retrieval method for an AI-assisted compilation system for special construction plans of critical and large-scale engineering projects. Background Technology

[0002] In the construction industry, specialized construction plans for high-risk and hazardous projects are essential for ensuring the management and safety of such projects. High-risk and hazardous projects refer to sub-projects with significant risks during construction that could easily lead to mass casualties or substantial economic losses. Since March 2018, regulations have been established for the management content and procedures of high-risk and hazardous projects, with the core being the management of specialized construction plans. As a guiding document for the management of high-risk and hazardous project construction, the entire plan encompasses multiple aspects of project management, including construction organization, technology, materials and equipment, safety precautions, and quality control, demanding a high level of experience and comprehensive abilities from the plan's drafter.

[0003] In the past two years, AI technology has rapidly penetrated and been applied to various fields and aspects of production and daily life, playing a significant role. Among these, document editing is considered one of the jobs most likely to benefit from enhanced productivity through artificial intelligence tools. Currently, AI-assisted editing of general documents such as job application reports and work summaries, which cater to the needs of the general public, has been developed to a relatively mature stage. However, in the construction industry, AI-assisted editing of core project management documents such as plans is still in the exploratory stage.

[0004] At present, using AI large-scale models to assist in the preparation of specialized construction plans is an extremely challenging but invaluable task. The reasons for this are as follows:

[0005] 1. Current Status of Plan Development

[0006] The construction plan is divided into nine chapters: project overview, basis for preparation, construction schedule, construction technology, construction assurance measures, construction management and personnel allocation and division of labor, acceptance requirements, emergency response measures, calculation sheets and related construction drawings. Each chapter is further divided into multiple sections and subsections. There is a strict logical correspondence between the chapters. For example, the sections "1.5 Hazard Identification and Classification," "5.2.1 Safety Assurance Measures," and "8.2 Emergency Events and Emergency Response Measures" should follow the logic of risk identification → prevention and control measures → emergency measures. The plan document is quite long, generally around 100,000 words, and includes various formats such as text, tables, and images. The plan is prepared by the project's construction technical personnel. Due to the complexity of high-risk construction projects, the preparation period for a single plan document is approximately 1-4 weeks.

[0007] 2. Characteristics of knowledge in the field of architecture

[0008] Generating professional documents using large-scale models requires high-quality, systematic knowledge as a prerequisite and foundation. Currently, professional knowledge in the construction field suffers from fragmented data, inconsistent standards, and difficulty in integrating tacit experience. For example, multiple standards exist for the same technical issue in national regulations, industry standards, and local regulations; the same construction process varies greatly due to differences in climate, region, geology, and surrounding environment; and tacit knowledge, often the core technology of a solution, is often buried in corporate management and expert experience but is difficult to extract and use in a structured way.

[0009] 3. Bottlenecks in large models

[0010] Large language models also face bottlenecks in generating long, professional documents. Firstly, their ability to handle long texts and logical connections is insufficient. The attention mechanism of the Transformer architecture weakens long-distance textual connections, and key logical chains, such as risk identification → prevention measures → emergency measures mentioned earlier, are prone to breakage beyond the model's context window. Secondly, text quality and compliance are inadequate. Due to insufficient modeling of tacit knowledge in vertical domains, lag in dynamic knowledge, and missing key knowledge, the illusion of a large model can easily arise. In high-compliance scenarios such as major engineering projects, this risk is often fatal.

[0011] Furthermore, the development of an AI-assisted system for compiling special construction plans for critical and major engineering projects requires the ability to achieve accurate and efficient knowledge retrieval and utilization. This knowledge retrieval and utilization necessitates the processing and integration of large amounts of data and knowledge from diverse sources and types to support the high-quality compilation of special construction plans for critical and major engineering projects. However, existing technologies have shortcomings in document parsing and knowledge retrieval, such as the inability to effectively handle multiple document types, the lack of an automated knowledge tagging system, and deficiencies in complex queries and professional content retrieval. Although some methods attempt to improve retrieval accuracy by increasing the complexity of retrieval algorithms, these methods often overlook the importance of document parsing and the crucial role of automatic knowledge tag construction, lacking intelligent tools, thus resulting in continued efficiency and accuracy issues in practical applications. Summary of the Invention

[0012] This invention proposes a knowledge system construction and retrieval method for an AI-assisted construction plan preparation system for critical and major engineering projects. Based on a thorough study of the current status of plan preparation, the characteristics of knowledge in the construction field, and the current bottlenecks of large language models, it realizes targeted knowledge system construction and efficient retrieval methods, solving the technical problems of knowledge system construction and efficient retrieval and utilization in the AI-assisted construction plan preparation system for critical and major engineering projects.

[0013] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0014] A knowledge system construction and retrieval method for an AI-assisted construction plan preparation system for critical and major engineering projects, including:

[0015] S1. Multidimensional Knowledge Base Construction: Construct a multidimensional knowledge base to store the compilation rules, specifications, standards, and project knowledge of critical and major engineering projects;

[0016] S2. Establishment of a multi-level knowledge graph: Establish a multi-level knowledge graph for the construction of critical and major projects, so that the multi-dimensional knowledge base can form a knowledge system;

[0017] S3, Intelligent Parsing of Multiple Document Types: Establishes a parsing strategy for multiple document types to extract document structure and knowledge from documents;

[0018] S4. Building a Knowledge Governance System: The construction steps include:

[0019] Knowledge classification divides the multidimensional knowledge base into general knowledge and project knowledge for importing corresponding knowledge documents. The multi-level knowledge graph provides semantic classification standards for knowledge classification.

[0020] Knowledge import involves parsing and extracting knowledge files using a multi-type document parsing strategy, then slicing them into knowledge fragments, converting them into semantic vectors, and storing them in the corresponding multi-dimensional knowledge base.

[0021] Automatic label prediction: Based on a large language model, by learning multi-dimensional knowledge base types, multi-level knowledge graph structures, and semantic features of knowledge fragments, labels for knowledge fragments at different levels are predicted. The labels include hierarchical labels and attribute labels.

[0022] S5. Construct a multi-level knowledge representation and retrieval model: Adopt a hybrid retrieval method, combine contextual semantics to retrieve structured and unstructured knowledge from a multi-dimensional knowledge base, and provide multi-hop association paths for hybrid retrieval through a multi-level knowledge graph.

[0023] Furthermore, the multidimensional knowledge base mentioned in step S1 includes a compilation basis base, a standard and specification base, an enterprise knowledge base, and a project knowledge base. The compilation basis base is used to store knowledge of the rules for compiling special construction plans for critical and major projects. The standard and specification base is used to store knowledge of the standard and specification that must be followed in the construction of critical and major projects. The enterprise knowledge base is used to store the enterprise's knowledge of the construction standards for critical and major projects. The project knowledge base is used to store the basic knowledge of the project to which the critical and major project is located.

[0024] Furthermore, the multi-level knowledge graph for critical and major projects mentioned in step S2 includes a general knowledge graph for the field of building engineering, a general chapter setting graph for critical and major projects, and a chapter setting graph for construction technology. It defines the hierarchical relationship of "unit project → sub-item project → construction technology" and establishes the edge relationship of "risk identification → prevention and control measures → emergency measures" as a logical chain.

[0025] Furthermore, the multi-type document parsing strategy in step S3 includes:

[0026] S301. Automatically identify file types based on file extensions of various document types, and select the corresponding parsing operation based on the file type;

[0027] S302. Perform layout analysis on the document, extract the document structure, and then parse and extract metadata from the knowledge including text, titles, tables, and images; when processing table information, automatically identify the table structure and convert it into an editable data format.

[0028] Furthermore, in step S302, text extraction is performed on scanned or image format documents using OCR technology based on large model fine-tuning, converting the text in the image into editable text; a layout border detection method based on large model fine-tuning is used to automatically detect table areas in the PDF and parse them using the PyPDF2 processing library to extract row and column knowledge.

[0029] Furthermore, the layout analysis described in step S302 supports user-defined deletion rules, including:

[0030] Custom settings allow you to delete special characters or irrelevant content. For irrelevant content, select a specific area in the file and use regular expression matching technology to convert the user-input special characters and specific area features into regular expressions. During layout analysis, match the document content according to the regular expressions, and block any content that matches the deletion rules.

[0031] Furthermore, in step S4, slicing uses a slicing strategy, which includes:

[0032] Perform structured operations, segmenting the data according to specified identifiers and combined with sentence identifiers extracted from the knowledge, to form knowledge fragments for subsequent retrieval and use.

[0033] Furthermore, the automatic label prediction in step S4 includes:

[0034] S401. Use the corpus in the special construction knowledge document of major and critical projects to conduct domain-adaptive training on the large language model, so that the large language model can understand the contextual semantics of professional terms.

[0035] This includes using masked language modeling to optimize domain semantic adaptation:

[0036]

[0037] Among them, L MLM To optimize domain semantic adaptation results using mask language modeling; i is the mask position, M is the set of mask positions, W \MFor context, θ represents the model parameters, and w represents the corpus. i From knowledge document; P(w i |w \M ,θ) indicates that in the context W \M The corpus w under the condition of model parameters θ i The prediction;

[0038] S402. Construct a multi-task learning architecture at the top level of the large language model to predict the hierarchical labels and multi-dimensional attribute labels of knowledge fragments, and verify the label logic through a rule engine.

[0039] Furthermore, in step S402, the hierarchical label prediction uses weighted cross-entropy loss to strengthen the logical consistency between parent and child labels:

[0040]

[0041] L multi This represents the weighted cross-entropy loss, where k represents the number of label levels, and the total number of label levels is K; α k For hierarchical weights, y k For the true label distribution, CE represents the model's predicted probability; CE is the cross-entropy loss function.

[0042] Furthermore, the hybrid retrieval method described in step S5 employs cosine similarity retrieval. Based on the contextual semantics of the knowledge document, it integrates data from different formats and sources to find the knowledge, document, or data table with the highest relevance to the content summary in the knowledge graph, handling complex cases including polysemous words, synonyms, and domain terms. On this basis, a modified retrieval enhancement generation technique is used to optimize the document retrieval process and enhance the relevance of the retrieval results.

[0043] Compared with the prior art, the beneficial effects of the present invention include:

[0044] (1) This invention proposes a knowledge base construction architecture that matches the AI-assisted compilation system for special construction plans of critical and major projects. It integrates the basis for plan compilation, domestic general technical specifications, core documents of construction enterprise technical management, core documents of project management, and implicit knowledge of expert experience, providing a basic corpus construction sample for AI-assisted compilation of professional documents in vertical fields.

[0045] (2) This invention establishes a knowledge graph with three levels in a progressive manner: the overall structure of the construction engineering field, the general chapter settings of the scheme, and the chapter settings of the "construction technology section". This enables the multi-dimensional knowledge base to form a knowledge system that matches the rules for compiling special construction schemes for major and critical projects in the construction industry, and provides a knowledge system construction idea for AI-assisted compilation of professional documents in vertical fields.

[0046] (3) The Tag Classification Algorithm (LCA) based on a large language model proposed in this invention automatically generates hierarchical and multi-dimensional tags for each knowledge point in the knowledge base, realizing intelligent classification and annotation of knowledge. Through an automated tagging system, the time-consuming, labor-intensive, and error-prone nature of traditional manual tagging is solved, significantly improving the structured management level of the knowledge base. Automatic tag prediction not only improves the usability of knowledge but also provides strong support for subsequent accurate retrieval. For example, tagging management of multiple dimensions such as construction technology, applicable scope, and standards and specifications enables the large language model to accurately find knowledge matching the chapters and paragraphs in the later stages of compiling special construction plans for critical and large-scale engineering projects, improving the efficiency and quality of plan compilation.

[0047] (4) This invention supports automatic parsing and information extraction of various document formats (such as PDF, DOCX, XLSX, PNG, JPG, etc.) through multi-type document intelligent parsing technology, solving the problem that current document parsing technology is limited to specific document formats. By combining OCR technology with a layout border detection method based on large language model fine-tuning, it can effectively handle scanned documents, images, and complex table structures. This technology not only improves the efficiency and accuracy of document parsing, but also extracts various unstructured data, greatly reducing the risk of manual operation and information omission. Especially in the field of bridge construction, document types are diverse and complex, and traditional parsing methods cannot meet the needs of accurate extraction. The intelligent document parsing solution provided by this invention greatly improves the breadth and accuracy of knowledge application.

[0048] (5) This invention proposes a full lifecycle knowledge governance system for bridge construction. This system uses a classification and hierarchical structure to assign tags to the knowledge base and knowledge, supporting personalized knowledge sources and providing customized knowledge services based on the needs of different projects, scheme types, and construction techniques. Simultaneously, it supports regular knowledge updates and version management, enabling timely updates to the knowledge base content based on new standards, policy guidance, or construction practices. This system ensures that the data in the knowledge base remains up-to-date, avoiding the problem of knowledge base obsolescence due to standard updates and project changes. This is particularly valuable in the field of bridge construction, where its development requires a high degree of specialization and is easily affected by policy and regulatory updates.

[0049] (6) The hybrid retrieval method proposed in this invention combines the advantages of structured and unstructured data retrieval, enabling efficient cross-format and cross-source information retrieval based on user-input queries. Through cosine similarity and deep semantic analysis, it accurately matches relevant documents and knowledge fragments, particularly excelling in handling complex queries and polysemous words. Combined with Corrective Retrieval Enhancement Generation (CRAG) technology, the retrieval results are further optimized by an evaluator, ensuring high relevance and usability of the generated results. This technology not only provides more accurate retrieval results but also flexibly supplements external information sources according to actual needs, ensuring reliability and accuracy when dealing with complex problems. This technology improves the comprehensiveness and accuracy of information retrieval, solving the problem that traditional single retrieval methods cannot handle complex and diverse needs.

[0050] (7) This invention makes up for the knowledge deficiency in the construction field by constructing a multi-dimensional knowledge base, establishes the logical relationship of long texts with a multi-level knowledge graph, and realizes the full life cycle governance from knowledge import to application by intelligent parsing, governance and hybrid retrieval of multi-type documents, laying the foundation for the development of AI-assisted compilation system for special construction plans of critical and major projects based on a multi-agent collaborative framework. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the multidimensional knowledge base construction process according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the multi-level knowledge graph construction process according to an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the overall knowledge graph in the field of construction engineering according to an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the general chapter settings for the critical care solution in this invention.

[0056] Figure 6 This is a schematic diagram showing the layout of the construction process technology section of this invention.

[0057] Figure 7 This is a schematic diagram of the document parsing process according to an embodiment of the present invention;

[0058] Figure 8 This is a schematic diagram of the labeling process according to an embodiment of the present invention;

[0059] Figure 9 This is a schematic diagram of the hybrid retrieval process in an embodiment of the present invention. Detailed Implementation

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0061] To make the objectives and features of the present invention more apparent and understandable, further explanation will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0062] This invention aims to compensate for the knowledge deficiencies in the construction field by constructing a multi-dimensional knowledge base, establishing logical relationships in long texts through a multi-level knowledge graph, and achieving full lifecycle governance from knowledge import to application through intelligent parsing, governance, and hybrid retrieval of multiple document types. This lays the foundation for the development of an AI-assisted compilation system for special construction plans of critical and major engineering projects based on a multi-agent collaborative framework.

[0063] Based on the above requirements, this embodiment analyzes and studies the current status of the preparation of special construction plans for critical and major engineering projects in the construction field, the characteristics of knowledge in the construction field, and the bottlenecks of large models, and proposes a knowledge system construction and efficient retrieval method for AI-assisted preparation system of special construction plans for critical and major engineering projects.

[0064] like Figure 1 As shown, it specifically includes:

[0065] 1. Multidimensional knowledge base construction: Compared with the knowledge required for manual scheme preparation, a knowledge base with five dimensions was established, including a preparation basis base, a standard and specification base, an enterprise knowledge base, a project knowledge base, and expert experience. This provides a basic expectation for the construction field in the preparation of special construction schemes for large models.

[0066] Similar to manually developing plans, the knowledge required for technical personnel to develop plans also applies to developing plans using large-scale models. Suppose a senior expert in the construction field needs to develop a special construction plan for a high-risk project. First, they need to master the rules for plan development, such as the "Regulations on Safety Management of High-Risk Sub-Projects" and the "Guidelines for the Development of Special Construction Plans for High-Risk Sub-Projects," among other documents. Second, they need to master the standards and specifications that must be followed in high-risk construction projects, such as the "General Specifications for Construction Scaffolding" (GB 55023-2022) and the "Safety Technical Standard for Socket-Type Disc-Type Steel Pipe Scaffolding in Building Construction" (JGJT231-2021), which must be based on the development plan for cast-in-place beams using a disc-lock full-span scaffolding system. Third, as a member of the company, they should master the company's standards for the construction of this project, such as process details, work instructions, project management manuals, plan templates, and handover templates. Finally, they should master the basic knowledge of the project, such as project bidding documents, preliminary planning reports, and construction organization designs.

[0067] Therefore, as Figure 2As shown, a large-scale model-based solution development program should be equipped with at least four dimensions of knowledge base: a development basis base, a standard and specification base, an enterprise knowledge base, and a project knowledge base. Furthermore, expert experience, as an implicit knowledge base, is primarily used for constructing the knowledge graph of solution chapters and for testing feedback and reinforcement learning after solution generation.

[0068] 2. Establishment of a multi-level knowledge graph:

[0069] Knowledge related to the specialized construction of critical and major engineering projects is scattered across diverse and heterogeneous data sources, including specifications, standards, corporate knowledge, and expert experience, lacking a unified structure. This embodiment integrates fragmented knowledge into a hierarchical semantic network through a multi-level knowledge graph, forming a reusable knowledge framework and avoiding the subjectivity and inefficiency of manual integration.

[0070] In response to the rules and requirements for scheme preparation in a series of documents such as the "Guidelines for the Preparation of Special Construction Schemes for High-Risk Sub-Projects", a three-level knowledge graph is established, consisting of an overall knowledge graph in the field of building engineering, a general chapter setting graph for schemes, and chapter settings for construction technology. This enables the multi-dimensional knowledge base to form a knowledge system that matches the preparation of special construction schemes for high-risk projects.

[0071] like Figure 3 As shown, in order to form a knowledge system from the multidimensional knowledge base and adapt to the rules of scheme preparation, the multidimensional knowledge base and expert experience are integrated to establish a three-level knowledge graph, namely the overall knowledge graph of the construction engineering field, the general chapter setting graph of the scheme, and the chapter setting graph of the "construction technology" section.

[0072] Taking bridge engineering as an example, the first step is to create a comprehensive knowledge map of bridge engineering. This map clearly defines unit projects, sub-projects, construction techniques, applicable scope, correspondences with templates, and correspondences with development guidelines, etc. Figure 4 As shown.

[0073] Secondly, there is a general chapter layout diagram for high-risk construction plans, mainly based on the Ministry of Housing and Urban-Rural Development's "Guidelines for the Preparation of Special Construction Plans for High-Risk Sub-Projects" and "List of Serious Defects in Special Construction Plans for High-Risk Sub-Projects (Trial Implementation)," enterprise management needs, and expert experience. This diagram clarifies the chapter names, logical relationships between chapters, chapter content summaries, and knowledge sources, such as... Figure 5 As shown.

[0074] Finally, there's the knowledge graph for the "Construction Technology" section of the scheme. Based on the knowledge graphs set up in the scheme chapters, the "Construction Technology" section varies significantly depending on the scheme type, so separate knowledge graphs are set up for each scheme type. Taking the cast-in-place beam scheme with beam-column support as an example, such as... Figure 6 As shown.

[0075] The three-tiered knowledge graph is progressively refined, defining a hierarchical relationship of "unit project → sub-item project → construction technology"; and establishing a logical chain of "risk identification → prevention and control measures → emergency measures". In the process of building this knowledge graph, not only was a multi-dimensional knowledge base systematized and logically connected, but it also summarized and refined tacit knowledge—expert experience—providing a deterministic framework for knowledge retrieval, utilization, and system development. Within the knowledge governance system, it plays a dual role: providing structured framework support and driving dynamic relationships.

[0076] 3. Intelligent parsing of multiple document types: Establish an automated parsing process, extract key information, eliminate irrelevant content, and transform it into structured knowledge through intelligent parsing of multiple document types in the knowledge system, facilitating subsequent governance and retrieval.

[0077] This step is as follows: Figure 7 As shown, it includes:

[0078] Building a multidimensional knowledge base involves various types of documents. To enable efficient retrieval and utilization of large models, an intelligent parsing strategy for multiple document types is established. This strategy automatically identifies the file type based on the file extension and performs parsing operations, supporting PDF, DOCX, XLSX, TXT, and MD format documents, as well as JPG and PNG format images.

[0079] This strategy performs layout analysis on documents, extracts document structure, and then parses and extracts metadata from text, titles, tables, and images, while automatically filtering out inefficient information such as headers and footers to ensure accurate parsing results. When processing table information, it automatically identifies the table structure and converts it into an editable data format.

[0080] Existing technologies have limitations in document parsing types. For documents of different formats, such as .pdf, docx, xlsx, txt, and md formats used in bridge construction, as well as images in jpg and png formats, each type has different layout formats. Existing technologies suffer from low robustness and insufficient compatibility in analyzing the layout of multiple document types, making it difficult to use a single layout analysis model to uniformly handle all document formats. Furthermore, when processing complex documents, existing layout analysis technologies cannot accurately extract key information, especially when the document content is extensive, the structure is complex, and the format is inconsistent, easily leading to information omissions or incorrect extraction. For example, when performing layout analysis on bridge construction documents containing numerous tables and images, it is difficult to accurately identify the table structure and the location of effective information in the images. In addition, existing technologies often employ fixed rules and processes, making it difficult for users to adjust and optimize them according to their own needs, and failing to adequately adapt to the specific document processing requirements of different projects in the bridge construction field.

[0081] Compared with existing layout analysis methods, the layout analysis method proposed in this embodiment offers significant performance improvements and supports various document formats. Whether it's text or image documents, layout analysis can be performed through a unified process, greatly enhancing the model's ability to handle different document types and meeting the diverse document needs in the bridge construction field. Furthermore, the layout analysis technology in this embodiment supports user-defined deletion rules. For example, users can define specific special characters or irrelevant content to be deleted, such as abbreviations of specific project codes or temporary special symbols used within a project in bridge construction documents. For irrelevant content, users can select specific areas in the file, and the system model automatically identifies the content characteristics of those areas, generating corresponding deletion rules. The algorithm uses regular expression matching technology; after the user completes the settings and submits, the system converts the user-input special characters and area features into regular expressions. During layout analysis, the system matches the document content based on these regular expressions, and once content matching the deletion rules is found, it is blocked, thereby improving document processing flexibility. Based on these improvements, the accuracy and parsing efficiency of the layout analysis method proposed in this embodiment are significantly improved.

[0082] To improve the efficiency and accuracy of document parsing, this embodiment uses OCR (Optical Character Recognition) technology based on large model fine-tuning to extract text from scanned or image-formatted documents, converting the text in the image into editable text; and uses a layout border detection method based on large model fine-tuning to automatically detect table areas in PDFs and parse them using the PyPDF2 processing library to extract row and column knowledge.

[0083] Existing OCR and bounding box detection technologies based on large models are mostly general-purpose and lack in-depth consideration of the complex needs of specific domains. This embodiment constructs a large-model fine-tuning OCR technology for bridge construction data. Utilizing a large amount of proprietary bridge construction data, including construction drawings, technical specifications, and design documents, the large model is fine-tuned and trained to improve the OCR technology's ability to deeply understand the unique symbols, complex annotations, and technical terms in bridge construction drawings, achieving a transformation from general recognition to domain-deep understanding.

[0084] Regarding layout border detection, this embodiment starts from the complex structure of tables in bridge construction documents and proposes a targeted optimization algorithm to enhance the border detection method's ability to deeply understand the logical relationships between table rows and columns. This enables it not only to accurately detect table borders but also to transform the knowledge within the tables into editable and searchable structured data, achieving intelligent parsing of table information. Specifically, the layout border detection method first collects bridge construction documents of various formats and types and performs image processing, using image enhancement technology to highlight table borders and lines to complete data preprocessing. Next, an attention mechanism and a multi-scale feature fusion module are introduced into a general large model. Subsequently, table data labeled with border coordinates and row and column division information are used to train the model, combining cross-entropy and IoU loss functions, and the stochastic gradient descent algorithm is used to optimize the training process. Throughout the entire document parsing process, OCR technology and layout border detection technology are two core capabilities, supporting efficient and accurate parsing of multiple types of documents.

[0085] 4. Construction of a knowledge governance system: such as Figure 8 As shown, through knowledge classification, import, slicing, and automatic tag prediction, structured knowledge is comprehensively managed throughout its entire lifecycle, establishing a systematic multidimensional knowledge base for accurate classification and storage, improving the usability of knowledge, and facilitating efficient retrieval and utilization.

[0086] This step builds a full lifecycle knowledge governance system based on a multi-dimensional knowledge base, a multi-level knowledge graph, and intelligent parsing strategies for various document types. This system includes knowledge classification, import, slicing, and automatic tag prediction.

[0087] Knowledge classification refers to the categorization of multidimensional knowledge into general knowledge and project knowledge based on knowledge attributes, to facilitate the use of AI-assisted scheme development systems. General knowledge includes the basis for development, standards and specifications, and enterprise knowledge, which are pre-installed in the system's basis for development, standard and specification library, and enterprise knowledge base, and are shared by various construction projects during scheme development. Project knowledge refers only to the project knowledge base. To avoid confusion, the project knowledge base is set up separately for each project and is imported by technical personnel before scheme development.

[0088] Multi-level knowledge graphs provide semantic classification standards for knowledge categorization. For example, the overall knowledge graph defines a hierarchical relationship of "unit project → sub-item project → construction technology" (such as "bridge engineering → pier construction → one-time completion"), which can guide the classification rules of general knowledge (such as dividing the standard library according to "technology → scope of application"). The chapter-based knowledge graph clarifies the logical chain of "risk identification → prevention and control measures → emergency measures," guiding the project knowledge base to be classified by chapter (such as "construction technology" and "safety assurance measures"), ensuring that the classification is consistent with the logic of scheme preparation.

[0089] By leveraging the semantic relationships within the graph, we can, to some extent, avoid the label confusion caused by insufficient human experience in traditional classification (such as "one-stop shop" being incorrectly classified as "materials management" instead of "construction technology"), thus improving classification accuracy.

[0090] Major engineering projects must follow a logical chain of "risk identification → prevention and control measures → emergency response," but large models may suffer from logical fragmentation due to token limitations. Multi-level knowledge graphs can provide external semantics for large models by explicitly encoding logical relationships, ensuring that the generated content complies with industry standards.

[0091] Knowledge import refers to the process of performing structured operations after document parsing, using different segmentation strategies such as segmentation by character or by title to create knowledge fragments that can be retrieved and used later. After segmentation, these fragments are automatically imported into the multidimensional knowledge base. During the knowledge import process, the BGE model is used to vectorize domain knowledge, preserving semantic relevance to the greatest extent possible.

[0092] Specifically, after importing the knowledge base files, the system automatically parses the file paragraphs, tables, and other content, converting them into text format. The converted text is then sliced; general knowledge bases are typically sliced ​​to a length of 1500 characters, while project knowledge bases are typically sliced ​​to a length of 800 characters. The sliced ​​text fragments are then converted into semantic vectors using word embedding and stored in a vector database for the system to access during solution generation.

[0093] Automatic tag prediction refers to the automatic tagging process during knowledge base construction. Based on a large language model, it performs automatic tag prediction by learning the semantic features of knowledge base types and knowledge content. It intelligently generates tags for different levels of knowledge bases and knowledge, achieving hierarchical tag construction and adding relevant attribute tags to each knowledge point. These tags include not only document type but also scope of application, structural type, construction technology, project stage, construction organization design, bidding documents, and tender documents. This enables efficient organization and governance of knowledge during AI-assisted solution development, facilitating accurate retrieval and retrieval by the large language model.

[0094] In the automatic label prediction scheme, the corpus in the special construction knowledge document of critical and major projects can be used to conduct domain-adaptive training of the large language model, so that the large language model can understand the contextual semantics of professional terms.

[0095] This includes using masked language modeling to optimize domain semantic adaptation:

[0096]

[0097] Among them, L MLMTo optimize domain semantic adaptation results using mask language modeling; i is the mask position, M is the set of mask positions, W \M For context, θ represents the model parameters, and w represents the corpus. i From knowledge document; P(w i |w \M ,θ) indicates that in the context W \M The corpus w under the condition of model parameters θ i The prediction;

[0098] Automatic label prediction also includes building a multi-task learning architecture at the top level of the large language model to predict hierarchical labels and multi-dimensional attribute labels for knowledge fragments, and verifying the label logic through a rule engine. Among them, hierarchical label prediction uses weighted cross-entropy loss to strengthen the consistency of parent-child label logic.

[0099]

[0100] L multi This represents the weighted cross-entropy loss, where k represents the number of label levels, and the total number of label levels is K; α k For hierarchical weights, y k For the true label distribution, CE represents the model's predicted probability; CE is the cross-entropy loss function.

[0101] Furthermore, project knowledge is dynamically updated as the project progresses (e.g., "construction phase from foundation construction → main structure construction"). The dynamic expansion capability of the knowledge graph (e.g., adding a node for "main structure construction → safety measures for high-altitude operations") can support real-time updates to the knowledge governance system, avoiding the problem of low retrieval accuracy caused by outdated static classification rules.

[0102] 5. Hybrid Knowledge Retrieval Method: The hybrid retrieval technology, which combines cosine similarity retrieval and modified retrieval augmentation (CRAG), enables the system to find the most relevant and matching content from a multi-dimensional knowledge base, achieving accurate retrieval and laying the foundation for the development of an AI-assisted compilation system for special construction plans of critical engineering projects based on a multi-agent collaborative framework.

[0103] A multidimensional knowledge base contains both structured and unstructured knowledge, which traditional retrieval methods struggle to integrate effectively. Therefore, this implementation, for example... Figure 9 As shown, a hybrid retrieval method is constructed, combining the advantages of structured and unstructured retrieval. By building a multi-level knowledge representation and retrieval model, the accuracy, comprehensiveness, and flexibility of knowledge acquisition are improved.

[0104] The hybrid retrieval method employs cosine similarity retrieval, which, based on the contextual semantics of knowledge documents, can integrate data from different formats and sources to find the knowledge, documents, or data tables with the highest relevance to the content summary in the knowledge graph. It handles complex situations such as polysemous words, synonyms, and domain terms, ensuring the comprehensiveness of the retrieval results in terms of type and information.

[0105] Structured retrieval directly employs cosine similarity retrieval, while unstructured retrieval uses semantic-based similarity retrieval, both evaluated by an evaluator. Building upon this, Corrective Retrieval Augmented Generation (CRAG) technology is used to optimize the retrieval process. By optimizing the document retrieval process and enhancing the relevance of retrieval results, the large model is ensured to be accurate and efficient in handling highly specialized and complex retrieval tasks.

[0106] Multi-level knowledge graphs provide multi-hop association paths for hybrid retrieval: In hybrid retrieval, the node weights of the graph can serve as weighting factors in similarity calculations, increasing the retrieval priority of high-value knowledge. In CRAG technology, the edge relationships of the graph (such as causal chains of "risk identification → prevention and control measures → emergency measures") can serve as logical verification criteria for retrieval results, reducing the likelihood of irrelevant retrieval results appearing.

[0107] By combining the above five steps, this embodiment can construct a knowledge system and efficient retrieval system for an AI-assisted construction plan compilation system for critical and major engineering projects. Based on this, further development of a multi-agent collaborative framework can realize an AI-assisted construction plan compilation system for critical and major engineering projects based on this framework. The "multi-agent collaboration" mentioned here is unrelated to the core technical solution of this embodiment and will not be elaborated further. This embodiment constructs a knowledge system that integrates a multi-dimensional knowledge base and a multi-level knowledge graph in the field of building construction, performs intelligent analysis and systematic governance of it, and combines it with hybrid retrieval technology to provide a foundational knowledge system and efficient retrieval for an AI-assisted construction plan compilation system for critical and major engineering projects based on multi-agent collaboration.

[0108] Specifically, this embodiment first compares the knowledge required for manually compiling plans and establishes a knowledge base in five dimensions: a compilation basis database, a standard database, an enterprise knowledge base, a project knowledge base, and expert experience. This provides basic corpus in the field of architectural engineering for the compilation of specialized construction plans using large models.

[0109] Secondly, in response to the rules and requirements for scheme preparation in a series of documents such as the Ministry of Housing and Urban-Rural Development's "Guidelines for the Preparation of Special Construction Schemes for Sub-projects with High Risk," a three-level knowledge graph is established, consisting of the overall framework for the construction engineering field, general scheme chapters, and "construction technology section" chapters. This enables the multi-dimensional knowledge base to form a knowledge system that matches the preparation of special construction schemes for high-risk projects.

[0110] Secondly, an automated parsing process is established. Through intelligent parsing of various document types in the knowledge system, key information is extracted, irrelevant content is eliminated, and the information is transformed into structured knowledge, which facilitates subsequent governance and retrieval.

[0111] Then, through knowledge classification, import, slicing, and automatic tag prediction, structured knowledge is comprehensively managed throughout its entire lifecycle, establishing a systematic multidimensional knowledge base for accurate classification and storage, improving the usability of knowledge, and facilitating efficient retrieval and utilization.

[0112] Finally, a hybrid retrieval technique combining cosine similarity retrieval and modified retrieval augmentation (CRAG) is used to enable the system to find the most relevant and matching content from a multi-dimensional knowledge base, achieving accurate retrieval and laying the foundation for the development of an AI-assisted compilation system for special construction plans of critical engineering projects based on a multi-agent collaborative framework.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge system construction and retrieval method for an AI-assisted construction plan preparation system for critical and major engineering projects, characterized in that, include: S1. Multidimensional Knowledge Base Construction: Construct a multidimensional knowledge base to store the compilation rules, specifications, standards, and project knowledge of critical and major engineering projects; S2. Establishment of a multi-level knowledge graph: Establish a multi-level knowledge graph for the construction of critical and major projects, so that the multi-dimensional knowledge base can form a knowledge system; S3, Intelligent Parsing of Multiple Document Types: Establishes a parsing strategy for multiple document types to extract document structure and knowledge from documents; S4. Building a Knowledge Governance System: The construction steps include: Knowledge classification divides the multidimensional knowledge base into general knowledge and project knowledge for importing corresponding knowledge documents. The multi-level knowledge graph provides semantic classification standards for knowledge classification. Knowledge import involves parsing and extracting knowledge files using a multi-type document parsing strategy, then slicing them into knowledge fragments, converting them into semantic vectors, and storing them in the corresponding multi-dimensional knowledge base. Automatic label prediction: Based on a large language model, by learning multi-dimensional knowledge base types, multi-level knowledge graph structures, and semantic features of knowledge fragments, labels for knowledge fragments at different levels are predicted. The labels include hierarchical labels and attribute labels. S5. Construct a multi-level knowledge representation and retrieval model: Adopt a hybrid retrieval method, combine contextual semantics to retrieve structured and unstructured knowledge from a multi-dimensional knowledge base, and provide multi-hop association paths for hybrid retrieval through a multi-level knowledge graph.

2. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects as described in claim 1, is characterized in that, The multidimensional knowledge base mentioned in step S1 includes a compilation basis base, a standard and specification base, an enterprise knowledge base, and a project knowledge base. The compilation basis base is used to store knowledge of the rules for compiling special construction plans for critical and major projects. The standard and specification base is used to store knowledge of the standard and specification that must be followed in the construction of critical and major projects. The enterprise knowledge base is used to store the enterprise's knowledge of the construction standards for critical and major projects. The project knowledge base is used to store the basic knowledge of the project to which the critical and major project is located.

3. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects as described in claim 1, is characterized in that, The multi-level knowledge graph for critical and major engineering projects mentioned in step S2 includes a general knowledge graph for the field of building engineering, a general chapter setting graph for critical and major engineering projects, and a chapter setting graph for construction technology. It defines the hierarchical relationship of "unit project → sub-item project → construction technology" and establishes the edge relationship of "risk identification → prevention and control measures → emergency measures" as a logical chain.

4. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects as described in claim 1, is characterized in that, The multi-type document parsing strategy in step S3 includes: S301. Automatically identify file types based on file extensions of various document types, and select the corresponding parsing operation based on the file type; S302. Perform layout analysis on the document, extract the document structure, and then parse and extract metadata from the knowledge including text, titles, tables, and images; when processing table information, automatically identify the table structure and convert it into an editable data format.

5. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects as described in claim 4, is characterized in that, In step S302, text extraction is performed on scanned or image format documents using OCR technology based on large model fine-tuning, converting the text in the image into editable text; a layout border detection method based on large model fine-tuning is used to automatically detect table areas in the PDF and parse them using the PyPDF2 processing library to extract row and column knowledge.

6. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects according to claim 4, is characterized in that, The layout analysis described in step S302 supports user-defined deletion rules, including: Custom settings allow you to delete special characters or irrelevant content. For irrelevant content, select a specific area in the file and use regular expression matching technology to convert the user-input special characters and specific area features into regular expressions. During layout analysis, match the document content according to the regular expressions, and block any content that matches the deletion rules.

7. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects according to claim 1, characterized in that, In step S4, slicing uses a slicing strategy, which includes: Perform structured operations, segmenting the data according to specified identifiers and combined with sentence identifiers extracted from the knowledge, to form knowledge fragments for subsequent retrieval and use.

8. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects according to claim 1, characterized in that, Automatic label prediction in step S4 includes: S401. Use the corpus in the special construction knowledge document of major and critical projects to conduct domain-adaptive training on the large language model, so that the large language model can understand the contextual semantics of professional terms. This includes using masked language modeling to optimize domain semantic adaptation: Among them, L MLM To optimize domain semantic adaptation results using mask language modeling; i is the mask position, M is the set of mask positions, W \M For context, θ represents the model parameters, and w represents the corpus. i From knowledge document; P(w i |w \M ,θ) indicates that in the context W \M The corpus w under the condition of model parameters θ i The prediction; S402. Construct a multi-task learning architecture at the top level of the large language model to predict the hierarchical labels and multi-dimensional attribute labels of knowledge fragments, and verify the label logic through a rule engine.

9. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects according to claim 1, characterized in that, In step S402, the hierarchical label prediction uses weighted cross-entropy loss to enhance the logical consistency between parent and child labels. L multi This represents the weighted cross-entropy loss, where k represents the number of label levels, and the total number of label levels is K; α k For hierarchical weights, y k For the true label distribution, CE represents the model's predicted probability; CE is the cross-entropy loss function.

10. The knowledge system construction and retrieval method of the AI-assisted compilation system for special construction plans of critical and major engineering projects according to claim 1, characterized in that, The hybrid retrieval method described in step S5 uses cosine similarity retrieval. Based on the contextual semantics of knowledge documents, it integrates data from different formats and sources to find the knowledge, documents, or data tables with the highest relevance to the content summary in the knowledge graph, handling complex cases including polysemous words, synonyms, and domain terms. On this basis, a modified retrieval enhancement generation technique is used to optimize the document retrieval process and enhance the relevance of the retrieval results.

Images