Nuclear power DCS knowledge base construction system and method

Abstract

The invention belongs to the technical field of nuclear power DCS knowledge base construction, and aims to solve the problems of knowledge dispersion, update lag, single expression, lack of intelligent application, difficulty in sharing and inheriting and the like in a traditional knowledge management mode of a nuclear power DCS. The invention provides a nuclear power DCS knowledge base construction system and method, the system comprises an application layer, a middle layer, a large model layer, a data layer and a computing power resource layer, the application layer comprises a knowledge base management module, an intelligent interaction module, a data service module, an office assistant module and an operation and maintenance assistant module; according to the method, a system architecture of a knowledge base is designed and comprises function division of a computing power resource layer, a data layer, a large model layer, a middle layer and an application layer. According to the method, an intelligent, systematized and efficient DCS knowledge base can be constructed, a powerful guarantee is provided for safe and stable operation of a nuclear power plant, and the efficiency, accuracy and practicability of nuclear power DCS knowledge management can be improved.

Description

Technical Field

[0001] The present application belongs to the technical field of nuclear power DCS knowledge base construction, and relates to a nuclear power DCS knowledge base construction system and method. Background Art

[0002] As the core of a nuclear power plant's safe and stable operation, the digital control system (DCS) is incredibly complex and crucial. With the continuous advancement of nuclear power technology, DCS systems are becoming increasingly large and complex, placing higher demands on the expertise and skills of operators and maintenance personnel. However, traditional DCS knowledge management methods have numerous drawbacks that make them difficult to meet the demands for efficient operation and maintenance in modern nuclear power plants. These drawbacks primarily include:

[0003] First, multi-source, heterogeneous data processing is difficult and lacks systematicity. DCS-related knowledge is scattered across various systems, and document formats include structured databases (such as SQL), unstructured documents (Word, PDF, images, etc.), and semi-structured data. This lacks unified management and integration, making it difficult to uniformly parse and correlate data using traditional methods.

[0004] Second, knowledge updates lag behind, and dynamic updating capabilities are insufficient. Traditional knowledge updates rely on manual maintenance, which is inefficient and unable to keep up with the rapid iteration and upgrade of DCS systems. The synchronization mechanism between the DCS health management data warehouse and the knowledge base is imperfect, making it difficult to reflect system changes in real time. This leads to outdated knowledge base content that cannot meet actual needs.

[0005] Third, knowledge extraction is inefficient. Existing technologies lack automated capabilities for semantic segmentation, table extraction, and image annotation of unstructured documents, relying instead on manual annotation, which is costly and error-prone.

[0006] Fourth, knowledge retrieval accuracy is poor. Traditional methods cannot effectively integrate the semantic information of multimodal data such as text, tables, and images, resulting in inaccurate retrieval results.

[0007] Fifth, knowledge is presented in a single format, making it difficult to understand and apply. Existing knowledge bases are mostly presented in static formats such as text and tables, which lack intuitiveness and interactivity, making it difficult for operations and maintenance personnel to quickly understand and apply knowledge.

[0008] Sixth, the lack of intelligent applications makes it difficult to assist in decision-making. Traditional knowledge bases lack intelligent application capabilities and are unable to provide accurate knowledge push and decision support based on specific scenarios and problems, making it difficult to effectively assist operations and maintenance personnel in fault diagnosis and resolution.

[0009] Seventh, knowledge sharing and inheritance are difficult. Traditional knowledge management methods make it difficult to effectively share and inherit knowledge, which can easily lead to knowledge silos and experience loss, hindering the long-term stable operation of nuclear power plants. Summary of the Invention

[0010] The purpose of this application is to provide a nuclear power DCS knowledge base construction system and method to solve the problems of knowledge dispersion, delayed update, single expression, lack of intelligent application and difficulty in sharing and inheritance in the traditional knowledge management method of nuclear power DCS system.

[0011] In order to achieve the above objectives, this application provides the following technical solutions:

[0012] In a first aspect, the present application provides a nuclear power DCS knowledge base construction system, comprising:

[0013] The application layer is used to provide specific functions and services to users;

[0014] The middle layer is used to connect the underlying technology and upper-layer applications;

[0015] The large model layer is used to integrate and utilize open source, general or domain-specific large models to achieve intelligent processing and application of knowledge;

[0016] Data layer, used for data collection, storage, management and preprocessing;

[0017] The computing resource layer is used to provide the computing resources required for system operation and support large-scale data processing and model training;

[0018] The application layer includes:

[0019] The knowledge base management module is used to classify and manage knowledge content of different types, businesses, and purposes;

[0020] The intelligent interaction module combines the data from the knowledge base to answer users' questions and trace the answers;

[0021] The data service module is used to query the structured data in the DCS data warehouse, call the structured data query results of the DCS data warehouse, and output the results, so as to quickly find the required information from the large amount of business and operation data of the power plant and quickly locate the relevant location;

[0022] The office assistant module is used to generate compliant texts based on different templates and user needs, quickly interpret user-uploaded documents, and extract key information to present to users;

[0023] The operation and maintenance assistant module is used to receive fault information, work order titles and completed work orders pushed by the business system, and after processing, push disposal suggestions, relevant historical work order results and experience feedback to the business system.

[0024] In some embodiments, the knowledge base management module includes:

[0025] Knowledge base operation unit, used to create, delete and edit knowledge base;

[0026] Knowledge base access unit, used to display the content list and classification structure of the corresponding knowledge base space;

[0027] The knowledge management unit is used to perform knowledge operations within a single knowledge base space, including uploading documents, adding text entries, and associating metadata.

[0028] In some embodiments, the knowledge management unit includes:

[0029] Data operation subunit, used to import and delete knowledge data;

[0030] The information management subunit is used to display the metadata of the data in the knowledge base, including the knowledge base name, file name, type, uploader, update time and parsing status;

[0031] The parsing processing subunit is used to extract paragraphs and vectorize the uploaded knowledge files, store the results in the vector database, update the parsing status in real time, and trigger re-parsing of data that failed to parse.

[0032] In some embodiments, the intelligent interaction module includes:

[0033] Intelligent retrieval unit, used to retrieve the corresponding content required by the user in the knowledge base according to the user's question, display the answer and trace the answer to the source;

[0034] The intelligent question-answering unit is used to retrieve the corresponding content required by the user in the knowledge base based on the user's question. The large model parses and summarizes the retrieved content and generates the answer corresponding to the question.

[0035] In some embodiments, the data service module includes:

[0036] The user question unit is used to receive the data query description input by the user and send the query request to the back-end processing module;

[0037] The data query unit is used to convert the user query description into SQL statements through the large model, retrieve the target data in the data warehouse and return the query results;

[0038] The data visualization icon generation unit is used to automatically generate visualization charts that match the query results based on the graphical requirements of the user's query needs, and return them synchronously with the data results;

[0039] The search history unit is used to store and display the user's historical query questions and corresponding search answers.

[0040] In some embodiments, the office assistant module includes:

[0041] The document generation unit is used to store various types of document templates and automatically generate documents that meet grammatical and semantic requirements by parsing the template structure and user input data through a large model;

[0042] The document interpretation unit is used to receive document files uploaded by users, analyze the document structure through a large model, extract key information, and output the interpretation results in structured text or visual form.

[0043] In some embodiments, the document generation unit includes:

[0044] The document generation requirement input subunit is used to receive the document generation requirement description input by the user and support the selection of document templates and document types;

[0045] The document generation processing subunit is used to call the large model to generate document content based on user needs and selected templates, and convert the results into a JSON file or text format that can be processed by the system;

[0046] The document generation result display subunit is used to render the generated document content online;

[0047] The generation record subunit is used to store and display the user's historical generation requests, generated documents and corresponding template information;

[0048] The template management subunit is used to manage document templates uploaded by users and provides template parsing, field storage and editing functions.

[0049] In some embodiments, the document interpretation unit includes:

[0050] The document interpretation requirement input subunit is used to receive the document parsing requirements input by the user and the uploaded document files;

[0051] The document interpretation subunit is used to parse the content of a single document and generate a summary or Q&A answers based on user needs, or to compare and analyze two documents and generate a summary of the differences;

[0052] The document interpretation record sub-unit is used to view the historical interpretation content.

[0053] In some embodiments, the operation and maintenance assistant module includes:

[0054] The status handling push unit is used to receive fault codes and descriptions pushed by the business system, process the fault information through a large model, retrieve the corresponding fault analysis results and handling suggestions from the knowledge base, sort them by relevance, and push multiple results to the business system, and provide document traceability information for the handling suggestions;

[0055] The work order generation unit is used to receive the work order title pushed by the business system, search the knowledge base for related historical work orders and maintenance plans through the big model, sort them by relevance, and push multiple results to the business system;

[0056] The experience feedback generation unit is used to extract content that meets the requirements of the experience feedback form based on the completed work order information selected by the user, generate structured feedback data and push it to the business system.

[0057] In some embodiments, the status handling push unit includes:

[0058] The status information receiving subunit is used by the nuclear power DCS knowledge base construction system to regularly synchronize equipment fault information from the DCS equipment health monitoring module and display it in a list format in the status disposal push module;

[0059] The treatment suggestion generation and viewing subunit is used to perform large-model driven fault analysis and treatment suggestion generation for one or more selected fault information, and record the treatment process and personnel information to optimize the model;

[0060] The status handling information management subunit is used to search based on fault description and processing status.

[0061] In some embodiments, the work order generating unit includes:

[0062] A new work order sub-unit is added to provide a status resolution push portal and a work order generation portal. The status resolution push portal can trigger work order generation based on alarm information and resolution suggestions, or a work order can be directly created through the work order generation portal.

[0063] The work order content generation sub-unit is used to fill in the work order title, equipment name and equipment code when adding a new work order. Based on the preset field requirements, the large model is called to push multiple maintenance plans and historical work order contents, and the content with the highest matching degree is selected for editing and saving;

[0064] The work order generation management subunit is used to store and manage saved work orders, and display the work order title, equipment code, equipment name, work order generation time and operating user in a list format.

[0065] In some embodiments, the experience feedback generating unit includes:

[0066] A new experience feedback sub-unit has been added to select work order information that has been completed on site and has a completion report;

[0067] The similar experience feedback retrieval sub-unit is used to parse the content of completed work orders through a large model, retrieve similar cases from the historical experience feedback library, and push multiple historical experience feedback orders according to relevance;

[0068] The experience feedback generation sub-unit is used to extract key information of completed work orders through the large model and generate structured experience feedback form content according to the preset template;

[0069] The experience feedback management subunit is used to store and manage saved experience feedback sheets, and display the feedback topics, involved equipment and operating users in a list format.

[0070] In a second aspect, the present application provides a method for constructing a nuclear power DCS knowledge base, comprising:

[0071] Step 1: Analyze the knowledge management requirements of nuclear power DCS, clarify the functional objectives and application scenarios of the knowledge base system, and design the system architecture of the knowledge base, including the functional division of computing resources layer, data layer, large model layer, middle layer, and application layer;

[0072] Step 2: Collect DCS-related data from the DCS health management data warehouse and pre-process the collected data;

[0073] Step 3: Use big model technology to extract key information from the data, convert the extracted knowledge into vector representation, associate and reason about the knowledge, and build logical relationships between the knowledge;

[0074] Step 4: Store structured knowledge into a vector database;

[0075] Step 5: Integrate open-source, general-purpose, or domain-specific large models, optimize them based on specific data from the nuclear power plant DCS domain, deploy large model services, and provide APIs for upper-layer applications to call.

[0076] Step 6: Based on the large model and knowledge base, realize multimodal knowledge expression through application functions;

[0077] Step 7: Conduct a comprehensive test on the knowledge base system, optimize system performance based on the test results, and deploy the nuclear power DCS knowledge base construction system into the nuclear power production environment or integrate it with the existing information business system.

[0078] Compared with the existing technology, the nuclear power DCS knowledge base construction system and method provided by this application have the following beneficial effects:

[0079] This application can build an intelligent, systematic and efficient DCS knowledge base, providing strong guarantee for the safe and stable operation of nuclear power plants.

[0080] This application is aimed at intelligent processing and knowledge integration of structured documents, unstructured documents and SQL databases.

[0081] This application significantly improves the efficiency, accuracy and practicality of nuclear power DCS knowledge management through a systematic, intelligent and dynamic knowledge base construction method, solves the problems of knowledge dispersion, delayed updates, single expression and lack of intelligent applications in traditional technologies, and provides strong support for the safe, stable and efficient operation of nuclear power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solution of this application, the following is a brief introduction to the drawings required for the technical description.

[0083] Figure 1 A flowchart of the knowledge base business process provided in the embodiment of this application;

[0084] Figure 2 This is a diagram of the DCS expert knowledge base system architecture provided in the embodiment of this application;

[0085] Figure 3 Schematic diagram of the intelligent question-answering and retrieval business process provided in the embodiment of this application;

[0086] Figure 4 Schematic diagram of the data service business process provided in this embodiment of the application;

[0087] Figure 5 A schematic diagram of the SQL flow for generating a large model provided in an embodiment of the present application;

[0088] Figure 6 A schematic diagram of the document generation process provided in this application embodiment;

[0089] Figure 7 A schematic diagram of the document interpretation business process provided in the embodiment of this application;

[0090] Figure 8 Schematic diagram of the operation and maintenance assistant business process provided in this embodiment of the application;

[0091] Figure 9 A data connection flow chart provided for an embodiment of the present application;

[0092] Figure 10 This is a flowchart of the method for constructing a nuclear power DCS knowledge base provided in an embodiment of the present application. DETAILED DESCRIPTION

[0093] The following is further explained in detail through specific implementation methods.

[0094] like Figures 1 to 9 As shown, the embodiment of the present application provides a nuclear power DCS knowledge base construction system, including:

[0095] The application layer is used to provide specific functions and services to users to meet the actual application needs of the nuclear power DCS knowledge base;

[0096] The middle layer connects underlying technologies and upper-layer applications, enabling intelligent processing, efficient retrieval, and precise application of knowledge through functional enhancement and system optimization.

[0097] The large model layer is used as the core intelligent engine of the system. It is mainly responsible for integrating and utilizing open source, general or domain large models to realize intelligent processing and application of knowledge.

[0098] The data layer is responsible for data collection, storage, management and preprocessing;

[0099] The computing resource layer is used to provide the computing resources required for system operation and support large-scale data processing and model training.

[0100] The middle layer is the core hub connecting the underlying technologies (computing resources, data, and large models) with upper-level applications (the application layer). This layer primarily includes components such as knowledge engineering, prompt word engineering, embedding models, vector databases, hybrid search engines, layout models, and third-party tools, enabling intelligent knowledge processing, efficient retrieval, and precise application.

[0101] The big model layer is the core intelligent engine of the system, mainly responsible for integrating and utilizing open source, general or domain big models to realize intelligent processing and application of knowledge.

[0102] The data layer is responsible for data collection, storage, management, and preprocessing, providing high-quality data support for the upper layer and serving as the basis for knowledge base construction and model training.

[0103] The computing resource layer provides underlying computing support for the entire system, ensuring the efficiency and stability of data processing and model operation.

[0104] Furthermore, the application layer includes:

[0105] The knowledge base management module is used to classify and manage knowledge content of different types, businesses, and purposes;

[0106] The intelligent interaction module combines the data from the knowledge base to answer users' questions and trace the answers;

[0107] The data service module is used by users to query the structured data in the DCS data warehouse. Users enter query conditions in natural language. The large model converts the natural language into corresponding SQL query statements, calls the structured data in the DCS data warehouse to query the data results, and outputs the results. This allows users to quickly find the required information from the large amount of business and operation data of the power plant and quickly locate the relevant location;

[0108] The office assistant module is used to generate compliant texts based on different templates and user needs, quickly interpret user-uploaded documents, and extract key information to present to users;

[0109] The operation and maintenance assistant module is used to receive fault information, work order titles and completed work orders pushed by the business system, and after processing, push disposal suggestions, relevant historical work order results and experience feedback to the business system.

[0110] In one embodiment, the knowledge base management module includes:

[0111] The knowledge base operation unit is used to create, delete and edit knowledge bases. The creation operation requires entering the knowledge base name, category and description information, and the deletion and editing operations are triggered by user interface buttons.

[0112] The knowledge base access unit enters the corresponding knowledge base space by clicking the knowledge base icon and displays its content list and classification structure.

[0113] The knowledge management unit supports knowledge addition, deletion, modification and query operations within a single knowledge base space, including uploading documents, adding text entries and associating metadata.

[0114] In one embodiment, the knowledge management unit includes:

[0115] The data operation subunit is used to import and delete knowledge data. The import supports local file upload and menu data synchronization of the DCS health management data warehouse. The deletion operation supports single or batch deletion of knowledge data and synchronous removal of related content in the vector database.

[0116] The information management subunit is used to display the metadata of the data in the knowledge base, including the knowledge base name, file name, type, uploader, update time and parsing status.

[0117] The parsing processing subunit is used to extract paragraphs and vectorize the uploaded knowledge files, store the results in the vector database, and update the parsing status (parsing, parsing completed, parsing failed) in real time, and supports triggering re-parsing of data that failed to parse.

[0118] Importing knowledge data: Select local files and upload them to the knowledge base. You can also choose to synchronize files (data) from one or more menus in the DCS health management data warehouse. The knowledge base can automatically synchronize new data with the data warehouse it connects to on a regular basis, for example, once a month.

[0119] Among them, knowledge data information: for data stored in the knowledge base, the knowledge base name, file name, type, uploader, update time, and parsing status will be displayed.

[0120] Knowledge data parsing: Extract knowledge segments from the original knowledge file, vectorize them, and store them in a vector database. The parsing status information shows whether each piece of knowledge has been successfully parsed. The system returns three statuses based on the data parsing process: parsing in progress, parsing completed, and parsing failed. Data that failed parsing can be re-parsed.

[0121] Among them, knowledge data deletion: For invalid data, you can delete or delete in batches, delete the corresponding data, and at the same time, the related data in the vector library will also be deleted.

[0122] In one embodiment, knowledge base permissions are managed through the knowledge base management module. Knowledge base user permissions are centrally managed by the platform user center. Ordinary users cannot operate the knowledge base, but knowledge base administrators can operate the knowledge base. Depending on the type or specialty of the knowledge base, multiple knowledge base administrators are set up, and each administrator can only operate the knowledge base related to him or her.

[0123] In one embodiment, the intelligent interaction module includes:

[0124] The intelligent retrieval unit searches the knowledge base for the corresponding content required by the user based on the user's question, displays the answer, and traces the answer to its source;

[0125] The intelligent question-answering unit retrieves the relevant content from the knowledge base based on the user's question. The large model parses and summarizes the retrieved content to generate the corresponding answer. The large model can be an existing open source, general-purpose large model or a fine-tuned / trained domain-specific large model.

[0126] Among them, the answer tracing positioning is: the displayed answer provides a tracing function. If it comes from an unstructured document, the position of the answer in the document is displayed; if it comes from a structured text, the entire text information is displayed.

[0127] After the user asks a question, the large model is first used to identify the intent and extract keywords. The intent is divided into general information and specific information. General information is usually some relatively broad or holistic questions raised by users, covering a relatively broad topic or field, and usually without clear specific details. Such questions usually focus on general concepts, summary information or a wide range of content. Specific information is a specific question with clear details raised by the user. Such questions usually require accurate answers to be extracted from existing information. The focus of the question is clear and often points to a specific thing, event or data. General information uses machine learning methods such as the nearest neighbor algorithm (KNN) to make an approximate comparison between the question and all vectors in the vector database, so as to identify paragraphs with a high degree of similarity as the recall result. Specific information uses BM25 (an algorithm widely used in the field of information retrieval) and keyword matching (keywords are generated through label configuration). This method can not only retrieve the content of all documents and ensure the comprehensiveness of the recall, but also return the original text content in the recalled vector, which is convenient for tracing the results generated by the large model and improving the interpretability and transparency of the system. The business process of the intelligent interaction module is as follows: Figure 3 shown.

[0128] In one embodiment, the difference between the intelligent retrieval unit and the intelligent question-answering unit in the intelligent interaction module is whether the content of answering the user is parsed and summarized based on the big model, and the intelligent question-answering unit is parsed and summarized based on the big model.

[0129] Take the intelligent question-answering unit as an example to illustrate its functions and usage:

[0130] (1) User question box: Enter the question description in the user question box, send it, and wait for the answer to be returned.

[0131] (2) Knowledge base call: All knowledge bases are used by default. If the user clearly knows the knowledge classification of the question when asking, he or she can directly select the knowledge base or select multiple knowledge bases at the same time. The system will answer according to the content of the knowledge base selected by the user to improve the efficiency and accuracy of retrieval questions and answers.

[0132] (3) Answer Source: The system displays the answer generated by the large model and the document source in a dialog box. Users can view the original document content based on the file name in the "Source" position below the answer. If the search result comes from an unstructured document, the location of the result in the document is displayed; if the search result comes from a structured text, the entire text information is displayed. For unstructured documents, the full document can be viewed and downloaded online.

[0133] (4) Search history: In the history box, you can view the questions asked in the past and the corresponding answers.

[0134] (5) Add the functions of copy, regenerate, approve, and disapprove under the answer box. The copy function corresponds to copying the generated text, and the regenerate function corresponds to re-answering the user's question. When the user likes or disapproves the question and answer, the system will store the approved question and answer and the disapproved question and answer in the database respectively. This function is mainly used for subsequent improvement and optimization of the large model.

[0135] In one embodiment, the data service business process includes user query, data demand / graphic demand, user intent completion, natural language to SQL, data result / graphic result, display / download, such as Figure 4 shown.

[0136] like Figure 5 The SQL generation process of the large model is shown. The technology of generating SQL from the large model includes three parts: recalling table names and field names, prompts and large model generation, and error correction of the large model generation content.

[0137] When recalling table and field names, the nuclear power DCS knowledge base construction system uses a variety of recall methods, including regular expressions and semantic vectors, to improve retrieval accuracy and efficiency. Based on the AST structure of user questions, a large model is used to extract the table names, field names, and other constraints in the question. Table and field names are first matched at the word level with the table and field names in the database using a combination of dynamic programming, Levenstein distance, and regular matching. For any unmatched data, semantic vectors are used to perform a semantic match, thereby ensuring the accuracy of the table and field name recall data.

[0138] After obtaining the recalled table and field names, this information is combined with the user's instructions to form prompts, which are then fed into the large model, which has been pre-trained using SQL statements. To improve the accuracy of the generated SQL statements, an agent is employed to cause the large model to repeatedly reflect on the output SQL statements. Specifically, a threshold is set. When the similarity between the SQL statement generated by the large model and the expected result falls below this threshold, the large model is prompted to reconsider until the generated SQL statement meets the requirements. This approach gradually optimizes the performance of the large model, enabling it to generate more accurate SQL statements.

[0139] While large models excel at generating SQL statements, they can still introduce errors and hallucinations. To address this, we use table name and field name comparison to correct errors in SQL statements generated by large models. Specifically, we compare the generated SQL statements with the actual table and field names in the database to check for inconsistencies. If an inconsistency is found, it can be identified as a potential error or hallucination, which can then be corrected. This approach effectively improves the executable nature of SQL statements and avoids data query failures or errors caused by incorrect SQL statements.

[0140] In one embodiment, the data service module includes:

[0141] The user question unit is used to receive the data query description entered by the user and send the query request to the back-end processing module. The user can enter the data description to be queried in the user question box and wait for the query result to be returned after sending it.

[0142] The data query unit converts user query descriptions into SQL statements using the big model, retrieves the target data from the data warehouse, and returns the query results. Based on the query description entered by the user, the big model automatically converts it into SQL query statements, retrieves the target data from the data warehouse, and displays the search results to the user.

[0143] The data visualization icon generation unit automatically generates visualizations that match the query results based on the graphical requirements of the user's query, and returns them synchronously with the data results. Data can be displayed graphically based on the data characteristics required. The user needs to specify the graphical generation requirements when asking the question, and the system will return the results and visualizations based on the user's needs.

[0144] The search history unit is used to store and display the user's historical query questions and corresponding search answers. In the history box, you can view the historical query questions and corresponding search answers.

[0145] In one embodiment, the office assistant module includes:

[0146] The document generation unit stores various document templates and uses a large model to analyze the template structure and user input data to automatically generate documents that meet grammatical and semantic requirements. The system can store different types of document templates, and the large model analyzes the templates and automatically generates various types of documents according to user requirements. When generating documents, the large model understands the subject and content based on user requirements and input data, and generates text that meets grammatical and semantic requirements.

[0147] The document interpretation unit receives uploaded documents, analyzes their structure using a large model, extracts key information, and outputs the interpretation results in structured text or visual form. The large model quickly interprets uploaded documents (supporting Word, PDF, and other file uploads), analyzes their structure and content, extracts key information, and presents it to the user in an easy-to-understand format.

[0148] like Figure 6 As shown, the large model generates the corresponding document content based on user requirements and templates in the template library. The JSON file required for document generation is pushed to the system. If a new template document needs to be generated, the large model first parses the new template document uploaded by the user, generates structured content, and then receives any adjustments to the structured content from the system and stores them in the template library.

[0149] In one embodiment, the document generation unit includes:

[0150] The document generation requirement input subunit receives user-entered document generation requirement descriptions and supports the selection of document templates and document types. A document generation requirement input box is provided where users can enter the document generation requirement description, select the document template and document type to be generated, and then wait for the macro model to be generated.

[0151] The document generation subunit is responsible for invoking the large model to generate document content based on user requirements and the selected template, and converting the results into a JSON file or text format that the system can process. After the large model generates the corresponding document content, it pushes the JSON file required for document generation to the system, which then presents the results to the user.

[0152] The document generation result display subunit is used to render the generated document content online, allowing users to preview, download, or copy the text. Users can preview and download the generated document online; if the user does not select a template, the result will be returned in text form, and the user can copy the text;

[0153] The generation record subunit is used to store and display the user's historical generation requests, generated documents and corresponding template information.

[0154] The Template Management subunit manages document templates uploaded by users, providing template parsing, field storage, and editing capabilities. Users can upload, edit, delete, and view templates, view the content of uploaded templates, and provide document parsing and field storage capabilities. Parsed fields can be displayed, and users can manually add and delete them. Parsed fields can also be saved, facilitating content generation based on fields in large models.

[0155] like Figure 7As shown in the figure, the system provides a parsing algorithm for temporary documents waiting to be interpreted. The large model receives the temporary document content uploaded by the user (supports Word and PDF files) and the interpretation requirements, generates the corresponding answer, and pushes the result to the business system.

[0156] In one embodiment, the document interpretation unit includes:

[0157] The Document Interpretation Requirement Input subunit is used to receive user-entered document parsing requirements and uploaded document files. A Document Interpretation Requirement Input box is provided where users can enter their document parsing requirements, select the document to upload, and then wait for the large model to generate the results.

[0158] The document interpretation subunit is used to parse the content of a single document and generate a summary or answers to questions based on user needs, or to compare and analyze two documents and generate a summary of the differences. When a user uploads a single document, the document content is parsed and a summary or answers to the document questions are returned to the user. When a user chooses to upload two documents, the document is parsed and compared. The large model summarizes the differences between the two documents and returns the summary to the user.

[0159] The document interpretation record sub-unit is used to view historical interpretation content. It stores and displays the user's historical interpretation requests and corresponding results.

[0160] In one embodiment, the operation and maintenance assistant module includes:

[0161] The status handling push unit receives fault codes and descriptions from the business system, processes the fault information using the large model, retrieves the corresponding fault analysis results and handling suggestions from the knowledge base, sorts them by relevance, and pushes multiple results to the business system. It also provides document traceability information for the handling suggestions. The large model receives fault codes and descriptions from the business system, processes the fault information, retrieves the corresponding fault analysis results and handling suggestions from the knowledge base, and provides multiple results by relevance, which are then pushed to the business system. It also provides the source of the handling suggestions, enabling document traceability.

[0162] The work order generation unit receives work order titles pushed by the business system, searches the knowledge base for related historical work orders and troubleshooting solutions using the big model, sorts them by relevance, and pushes multiple results to the business system. The big model receives work order titles pushed by the business system, searches the knowledge base for related historical work orders and troubleshooting solutions, and pushes multiple results by relevance to the business system.

[0163] The experience feedback generation unit is used to extract content that meets the experience feedback requirements based on the completed work order information selected by the user, generate structured feedback data, and push it to the business system. Based on the completed on-site work order information selected by the user, the corresponding content is extracted according to the experience feedback requirements and pushed to the business system.

[0164] In one embodiment, the status handling push unit includes:

[0165] The status information receiving subunit is used by the nuclear power DCS knowledge base construction system to regularly synchronize the equipment fault information of the DCS equipment health monitoring module, including alarm time, equipment code, fault code and fault description, and display it in a list form in the status disposal push module.

[0166] The disposal suggestion generation and viewing subunit is used to perform large-model driven fault analysis and generate disposal suggestions for one or more selected fault information. It supports users to interactively view, select or manually edit the suggestion content, and records the processing process and personnel information to optimize the model.

[0167] The status handling information management subunit is used to provide retrieval functions, which can be retrieved according to fault description and processing status, as well as deletion and batch deletion functions.

[0168] In one embodiment, the handling suggestion generation and viewing subunit analyzes the fault information selected by the user, retrieves and returns suggestions, supports viewing, selection and editing, stores question and answer pairs and records the processing personnel, specifically including: selecting one or more fault information for fault analysis, the large model retrieves the corresponding fault analysis results and handling suggestions in the knowledge base and sorts and returns them to the system; viewing on the handling suggestions, one or more "fault analysis" and "handling suggestions" returned by the large model are displayed in a form, and the user can choose to determine one of them as the handling suggestion for the alarm. If the generated answers are not correct, the user is supported to manually edit the "fault analysis" and "handling suggestion"; the system stores the question and answer pairs finally selected or modified by the user, including the alarm log information and fault analysis handling suggestions in the table, to facilitate subsequent model improvement and optimization. At the same time, the system automatically records the processing personnel of the status information based on the user information.

[0169] In one embodiment, the work order generating unit includes:

[0170] A new work order sub-unit has been added, providing both a status action push portal and a work order generation portal. The status action push portal allows users to trigger work order generation based on alarm information and action suggestions, while the work order generation portal allows users to directly create work orders. There are two new work order portals: the status action push portal allows users to generate a work order based on an alarm information and corresponding action suggestions; the work order generation portal allows users to directly generate a work order.

[0171] The work order content generation subunit is used to enter the work order title, equipment name, and equipment code when adding a new work order. Based on the preset field requirements, the large model is called to push multiple maintenance plans and historical work order content. The content with the highest degree of matching is selected for editing and saving. When adding a new work order, the work order title, equipment name, and equipment code are entered in the new window. Based on the preset generation field requirements of the nuclear power DCS knowledge base construction system, the large model pushes multiple maintenance plans and corresponding content of past work orders. The business system selects the one with the highest degree of matching for display, and the user can edit and save the generated content.

[0172] The Work Order Generation Management subunit stores and manages saved work orders, displaying the work order title, device code, device name, work order generation time, and operating user in a list format. Saved work orders can be viewed in the Work Order Generation subunit, displaying a list of work order information, including the work order title, device code, device name, work order generation time, and operating user. Work order details can be viewed, edited, and deleted, and a search function is provided by entering the work order title to search.

[0173] In one embodiment, the experience feedback generating unit includes:

[0174] A new experience feedback subunit has been added to select work order information that has been completed on site and has a completion report. The completion information of the work order is automatically obtained from the DCS health management data warehouse. When adding experience, select work order information that has been completed on site and has a completion report. The completion information is obtained from the DCS health management data warehouse.

[0175] The Similar Experience Feedback Retrieval sub-unit uses the large model to analyze the content of completed work orders, retrieve similar cases from the historical experience feedback library, and push multiple historical experience feedback sheets based on relevance. The large model retrieves historical experience feedback information based on the description of the completed work order and pushes and displays multiple historical experience feedback sheets based on relevance. If similar experience feedback already exists, the user can choose not to generate it. If the user chooses to generate it, the user will be prompted to generate the experience feedback sheet.

[0176] The Experience Feedback Generation subunit extracts key information from completed work orders through the large model and generates structured experience feedback forms based on pre-set templates. The large model extracts key information from completed work orders and summarizes and generates the required experience feedback forms according to the nuclear power DCS knowledge base construction system. Users can edit and save the generated content.

[0177] The Experience Feedback Management subunit stores and manages saved experience feedback sheets, displaying the feedback topics, involved equipment, and operating users in a list format. Saved experience feedback can be viewed in the Experience Feedback module within the application layer of the nuclear power DCS knowledge base construction system, displaying the experience feedback information in a list format, including the topic, involved equipment, and operating users. Experience feedback details can be viewed, edited, and deleted, and a search function is provided for searching by entering the experience feedback topic.

[0178] In addition, the DCS data connection used in the embodiment of the present application includes:

[0179] Data reading interface: Based on the data interface of the database in the data warehouse, develop a corresponding data reading interface to process the structured data and unstructured data related to the data warehouse business respectively; at the same time, connect to the existing business system to realize the reading and parsing of documents uploaded by the business system.

[0180] Data parsing and storage: Read the existing data warehouse API, parse the unstructured data and structured data received and store them in the warehouse.

[0181] Data docking process is as follows Figure 9 As shown in the figure, the nuclear power DCS knowledge base construction system first issues a data synchronization request and sends the request through the data service interface. Next, it pulls the required data from the data warehouse and returns it to the knowledge base. Subsequently, the nuclear power DCS knowledge base construction system calls the model parsing interface, and the large model parses the file data and stores the parsing results in the vector library. Finally, the system returns the operation status and responds to the parsing operation, completing the entire data docking process. This process achieves efficient data synchronization, intelligent parsing, and storage, providing reliable data support for the construction and application of the knowledge base.

[0182] like Figure 10 As shown, the embodiment of the present application provides a method for constructing a nuclear power DCS knowledge base, including:

[0183] Step 1: Requirements Analysis and System Design. Analyze the nuclear power DCS knowledge management requirements and clarify the functional objectives and application scenarios of the knowledge base system. Design the knowledge base system architecture, including the functional divisions of the computing resource layer, data layer, large model layer, middle layer, and application layer.

[0184] Step 2: Data Collection and Preprocessing. Collect nuclear power plant DCS-related data from the DCS health management data warehouse. Perform preprocessing operations such as cleaning, deduplication, and format standardization on the collected data.

[0185] Step 3: Knowledge extraction and structured representation: Using large model technology, key information is extracted from the data, the extracted knowledge is converted into vector representation, and the knowledge is associated and reasoned to build logical relationships between the knowledge.

[0186] Step 4: Knowledge Storage and Management. Store structured knowledge in a vector database to support efficient retrieval and updating. Design knowledge base management functions to support knowledge entry, classification, and version control. Implement knowledge base permission management to ensure data security and compliance.

[0187] Step 5: Large Model Integration and Fine-tuning. Integrate open-source, general-purpose, or domain-specific large models to provide capabilities such as natural language processing and knowledge reasoning. Optimize the large model based on specific data from the nuclear power plant DCS domain to improve its domain performance. Deploy the large model service and provide an API for upper-layer applications to call.

[0188] Step 6: Knowledge retrieval and application development. Develop application functionality based on the large model and knowledge base. Implement multimodal knowledge representation (e.g., text, charts, etc.) to enhance user experience.

[0189] Step 7: System testing, optimization, and deployment. Conduct comprehensive testing of the knowledge base system, optimize system performance based on the test results, deploy the nuclear power DCS knowledge base system into the nuclear power production environment or integrate it with existing information business systems to ensure compatibility with existing systems, and provide system operation and maintenance support.

[0190] The knowledge base in the nuclear power DCS knowledge base construction system is the data source and basis for intelligent question and answer, operation and maintenance assistant and other functions. The knowledge base mainly extracts knowledge paragraphs from the original knowledge files and stores them in the vector database. These knowledge mainly include text, pictures, tables and other modalities. The extracted data is vectorized by the embedding model and stored in the vector database. In the process of storing the file in the vector database, the label configuration will be carried out according to the full text summary and vectorized. Finally, these vectors are stored in the Milvus vector database according to the hierarchical rules of the text to facilitate subsequent efficient retrieval and data analysis. The knowledge base business process is as follows Figure 1 shown.

[0191] Multimodal knowledge extraction methods for knowledge bases include:

[0192] (1) Word-level parsing and summary generation

[0193] S11: Document structured segmentation.

[0194] Use the LayoutLM model to parse Word / PFD documents, identify title levels, paragraphs, tables, and image locations, and generate a tree-like directory structure.

[0195] S12: Semantic segmentation and summary generation.

[0196] The fine-tuned BART model is called for each paragraph to generate a summary, and the parent node summary covers the child node semantics.

[0197] S13: Multimodal association.

[0198] Convert table content to Markdown format, add OCR-extracted annotation text to images, establish bidirectional links with adjacent paragraphs, and store them in JSON metadata.

[0199] (2) Table annotation association

[0200] S21: Table structure analysis.

[0201] The Table Transformer model is used to detect the table area, identify the row and column structure, and extract the cell content.

[0202] S22: Semantic annotation binding.

[0203] If the table has no title or annotation, call the GPT-4 model to generate a descriptive title based on the context and associate it with the nearest paragraph.

[0204] S23: Vectorized storage.

[0205] The table content and annotation text are input into the BGE model together to generate a joint vector, which is then stored in the "table-text" association index of the Milvus database.

[0206] (3) SQL field mapping

[0207] S31: Dynamic field extraction.

[0208] Extract table names, field names, and data types from the SQL database schema and build a field-semantic dictionary (e.g., `pump_pressure → "pump pressure value"`).

[0209] S32: Large model driven mapping optimization.

[0210] Use the CodeLlama model to analyze field annotations and query logs and automatically add synonyms (for example, `pressure` is mapped to `pressure` and `pressure`).

[0211] S33: Vector index construction.

[0212] Field names, annotations, and synonyms are jointly encoded into 256-dimensional vectors, sharing the same vector space with unstructured data vectors, supporting cross-modal retrieval.

[0213] Furthermore, to ensure the effectiveness of the knowledge base, a dynamic synchronization mechanism for structured and unstructured data is established.

[0214] (1) Real-time synchronization trigger design

[0215] S41: Structured data monitoring.

[0216] Set up a Binlog listener in a SQL database (such as MySQL), capture INSERT / UPDATE / DELETE operations, and push change events to a message queue (Kafka).

[0217] S42: Unstructured file monitoring.

[0218] Use the inotify tool to monitor the knowledge base file directory, detect new / modified Word / PDF files, and trigger incremental parsing tasks.

[0219] (2) Incremental parsing and version control

[0220] S51: Change data classification processing.

[0221] Structured data: Update the corresponding field vector in the vector database based on the Binlog event type.

[0222] Unstructured data: Only the modified sections in the document (identified by the Git Diff algorithm) are parsed and a delta vector is generated.

[0223] S52: Version snapshot storage.

[0224] After each update, a SHA-256 hash snapshot is generated for the affected data and stored in the version library (such as Git LFS) along with the timestamp, supporting rollback by timeline.

[0225] (3) Consistency check

[0226] S61: Timed proofreading task.

[0227] The proofreading script is started at dawn every day to compare the number of records and hash values of key fields in the SQL database and the vector library, and mark inconsistent items.

[0228] S62: Exception handling.

[0229] Trigger full analysis of inconsistent data and notify administrators through the alarm system to ensure data integrity.

[0230] Furthermore, in order to facilitate the intelligent application of the knowledge base, a hybrid retrieval system is developed that integrates semantic vectors, keywords and SQL queries.

[0231] (1) Retrieval request classification routing

[0232] Intent Recognition:

[0233] After the user enters a query, the BERT-INTENT model is called to determine the intent type:

[0234] Precise query (such as device code) → trigger the SQL engine;

[0235] Semantic search (e.g., “pump pressure anomaly analysis”) → vector similarity search;

[0236] Hybrid queries (such as "Pump pressure value of unit X in 2023") → Parallel execution of SQL and vector search.

[0237] (2) Multi-engine parallel execution

[0238] S71: SQL query generation.

[0239] Use the DIN-SQL model to convert natural language into SQL, prioritizing structured data (such as `SELECT pressure FROM pump_log WHERE date='2023-01-01'`).

[0240] S72: Semantic vector retrieval.

[0241] After the query statement is vectorized by the BGE model, an approximate nearest neighbor search (ANN) is performed in Milvus to recall the top 50 relevant paragraphs.

[0242] S73: Keyword enhancement.

[0243] Apply the BM25 algorithm to the recall results for secondary sorting, and increase the priority of documents containing precise terms (such as `GST018 valve`).

[0244] The above description is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be covered by the scope of protection of the present application.

Claims

1. A nuclear power DCS knowledge base construction system, characterized in that: include: The application layer is used to provide specific functions and services to users; The middle layer is used to connect the underlying technology and upper-layer applications; The large model layer is used to integrate and utilize open source, general or domain-specific large models to achieve intelligent processing and application of knowledge; Data layer, used for data collection, storage, management and preprocessing; The computing resource layer is used to provide the computing resources required for system operation and support large-scale data processing and model training; The application layer includes: The knowledge base management module is used to classify and manage knowledge content of different types, businesses, and purposes; The intelligent interaction module combines the data from the knowledge base to answer users' questions and trace the answers; The data service module is used to query the structured data in the DCS data warehouse, call the structured data query results of the DCS data warehouse, and output the results, so as to quickly find the required information from the large amount of business and operation data of the power plant and quickly locate the relevant location; The office assistant module is used to generate compliant texts based on different templates and user needs, quickly interpret user-uploaded documents, and extract key information to present to users; The operation and maintenance assistant module is used to receive fault information, work order titles and completed work orders pushed by the business system, and after processing, push disposal suggestions, relevant historical work order results and experience feedback to the business system.

2. The nuclear power DCS knowledge base construction system according to claim 1, characterized in that: The knowledge base management module includes: Knowledge base operation unit, used to create, delete and edit knowledge base; Knowledge base access unit, used to display the content list and classification structure of the corresponding knowledge base space; The knowledge management unit is used to perform knowledge operations within a single knowledge base space, including uploading documents, adding text entries, and associating metadata.

3. The nuclear power DCS knowledge base construction system according to claim 2, characterized in that: The knowledge management unit includes: Data operation subunit, used to import and delete knowledge data; The information management subunit is used to display the metadata of the data in the knowledge base, including the knowledge base name, file name, type, uploader, update time and parsing status; The parsing processing subunit is used to extract paragraphs and vectorize the uploaded knowledge files, store the results in the vector database, update the parsing status in real time, and trigger re-parsing of data that failed to parse.

4. The nuclear power DCS knowledge base construction system according to claim 1, characterized in that: The intelligent interaction module includes: Intelligent retrieval unit, used to retrieve the corresponding content required by the user in the knowledge base according to the user's question, display the answer and trace the answer to the source; The intelligent question-answering unit is used to retrieve the corresponding content required by the user in the knowledge base based on the user's question. The large model parses and summarizes the retrieved content and generates the answer corresponding to the question.

5. The nuclear power DCS knowledge base construction system according to claim 1, characterized in that: The data service modules include: The user question unit is used to receive the data query description input by the user and send the query request to the back-end processing module; The data query unit is used to convert the user query description into SQL statements through the large model, retrieve the target data in the data warehouse and return the query results; The data visualization icon generation unit is used to automatically generate visualization charts that match the query results based on the graphical requirements of the user's query needs, and return them synchronously with the data results; The search history unit is used to store and display the user's historical query questions and corresponding search answers.

6. The nuclear power DCS knowledge base construction system according to claim 1, characterized in that: The office assistant module includes: The document generation unit is used to store various types of document templates and automatically generate documents that meet grammatical and semantic requirements by parsing the template structure and user input data through a large model; The document interpretation unit is used to receive document files uploaded by users, analyze the document structure through a large model, extract key information, and output the interpretation results in structured text or visual form.

7. The nuclear power DCS knowledge base construction system according to claim 1, characterized in that: The operation and maintenance assistant module includes: The status handling push unit is used to receive fault codes and descriptions pushed by the business system, process the fault information through a large model, retrieve the corresponding fault analysis results and handling suggestions from the knowledge base, sort them by relevance, and push multiple results to the business system, and provide document traceability information for the handling suggestions; The work order generation unit is used to receive the work order title pushed by the business system, search the knowledge base for related historical work orders and maintenance plans through the big model, sort them by relevance, and push multiple results to the business system; The experience feedback generation unit is used to extract content that meets the requirements of the experience feedback form based on the completed work order information selected by the user, generate structured feedback data and push it to the business system.

8. The nuclear power DCS knowledge base construction system according to claim 7, characterized in that: The status handling push unit includes: The status information receiving subunit is used by the nuclear power DCS knowledge base construction system to regularly synchronize equipment fault information from the DCS equipment health monitoring module and display it in a list format in the status disposal push module; The treatment suggestion generation and viewing subunit is used to perform large-model driven fault analysis and treatment suggestion generation for one or more selected fault information, and record the treatment process and personnel information to optimize the model; The status handling information management subunit is used to search based on fault description and processing status.

9. The nuclear power DCS knowledge base construction system according to claim 7, characterized in that: The experience feedback generation unit includes: A new experience feedback sub-unit has been added to select work order information that has been completed on site and has a completion report; The similar experience feedback retrieval sub-unit is used to parse the content of completed work orders through a large model, retrieve similar cases from the historical experience feedback library, and push multiple historical experience feedback orders according to relevance; The experience feedback generation sub-unit is used to extract key information of completed work orders through the large model and generate structured experience feedback form content according to the preset template; The experience feedback management subunit is used to store and manage saved experience feedback sheets, and display the feedback topics, involved equipment and operating users in a list format.

10. A method for constructing a nuclear power DCS knowledge base, characterized in that: include: Step 1: Analyze the knowledge management requirements of nuclear power DCS, clarify the functional objectives and application scenarios of the knowledge base system, and design the system architecture of the knowledge base, including the functional division of computing resources layer, data layer, large model layer, middle layer, and application layer; Step 2: Collect DCS-related data from the DCS health management data warehouse and pre-process the collected data; Step 3: Use big model technology to extract key information from the data, convert the extracted knowledge into vector representation, associate and reason about the knowledge, and build logical relationships between the knowledge; Step 4: Store structured knowledge into a vector database; Step 5: Integrate open-source, general-purpose, or domain-specific large models, optimize them based on specific data from the nuclear power plant DCS domain, deploy large model services, and provide APIs for upper-layer applications to call. Step 6: Based on the large model and knowledge base, realize multimodal knowledge expression through application functions; Step 7: Conduct a comprehensive test on the knowledge base system, optimize system performance based on the test results, and deploy the nuclear power DCS knowledge base construction system into the nuclear power production environment or integrate it with the existing information business system.

Images