Nuclear power company group plant management requirement validity judgment method based on event characteristics

Abstract

The invention belongs to the technical field of nuclear power operation, and particularly relates to a nuclear power company group plant management requirement validity judgment method based on event characteristics. Comprising the following steps: step 1, constructing a structured management requirement database; 2, event report information is extracted and stored; 3, screening a management requirement validity judgment range; 4, effective question answering and calculation of management requirements are carried out; and 5, outputting a management requirement validity judgment result. The method has the beneficial effects that a large model tool is applied to nuclear power group factory experience feedback work, the implementation effectiveness of management requirements in a nuclear power plant is judged by using a mode of constructing a prompt project, and the processing capability of informatization on mass data and the reasoning capability of an intelligent model are fully applied; and a measurable judgment standard is given, so that the effectiveness judgment of the management requirements is more intelligent, efficient and accurate.

Description

Technical Field

[0001] This invention belongs to the field of nuclear power operation technology, specifically relating to a method for judging the effectiveness of management requirements for nuclear power company clusters based on event characteristics. Background Technology

[0002] To more effectively conduct experience feedback across multiple nuclear power plants, nuclear power operators have established an A / B category management mechanism. This mechanism involves in-depth analysis of the causes of events occurring at individual power plants, extracting lessons learned, and proposing targeted management requirements. Each power plant then conducts its own analysis, evaluation, and investigation according to these requirements, and formulates and implements corrective actions based on the results to prevent similar events from recurring. Therefore, identifying whether newly occurring events at nuclear power plants, after undergoing analysis, evaluation, and investigation according to management requirements, share similar causes and failure points with related nuclear power A / B category events, and whether they address the problems that management requirements aim to solve, is a means of assessing the effectiveness of these management requirements.

[0003] In practice, when assessing the effectiveness of a management requirement, plant experience feedback managers primarily rely on staff members' familiarity with numerous nuclear power plant incidents and their personal abilities and experience. This approach has significant limitations. First, the staff member needs to be extremely familiar with nuclear power plant incidents and their characteristics; second, they need to be well-versed in the problems the management requirement aims to address; and third, they need strong logical reasoning skills to determine whether the problems addressed by the management requirement exist in newly occurring incidents. Therefore, the current assessment method is highly dependent on staff ability and work experience, resulting in low efficiency and a high risk of omissions and incorrect judgments.

[0004] Therefore, it is necessary to study a more clearly defined and intelligent method for judging the effectiveness of management requirements. This method should rely on information technology to automatically monitor and judge the effectiveness of management requirements, thereby improving the convenience and accuracy of judging the effectiveness of management requirements as the number of management requirements and events continues to increase, and providing assistance to managers who provide experience feedback to multiple factories. Summary of the Invention

[0005] The purpose of this invention is to provide a method for judging the effectiveness of management requirements for nuclear power plant clusters based on event characteristics. By determining whether there are management requirements that need to be addressed in an event, and combining this with pre-set rules, the method assesses the effectiveness of the management requirements implemented at the relevant nuclear power plants. This method is applicable to monitoring nuclear power plant events using information and intelligent technologies, automatically judging the effectiveness of management requirements, providing prompts and warnings, effectively reducing workload and significantly decreasing reliance on personnel skills and experience.

[0006] The technical solution of this invention is as follows: A method for judging the validity of management requirements of a nuclear power company's group of plants based on event characteristics, including a design method for a group of plant management requirement validity judgment agent, a list of management requirement validity judgment questions, and conversion rules for validity question answers and validity judgments, including the following steps:

[0007] Step 1: Build a structured management requirements database;

[0008] By manually sorting and extracting information from structured fields, information on management requirements is extracted, and a structured management requirements database is constructed.

[0009] Step 2: Extraction and Storage of Event Report Information

[0010] Through natural language processing, unstructured nuclear power incident report texts are transformed into machine-readable, understandable, and computable structured event feature information, providing standardized data input for subsequent validity assessments;

[0011] Step 3: Screening the Scope of Management Requirement Validity Judgment

[0012] Based on pre-defined business rules, a subset of management requirements that are associated with the "current event" and require validity judgment are selected from the structured management requirements database.

[0013] Step 4: Management Requirements Validity Questions and Calculations

[0014] By utilizing large language models, pre-screened management requirements are accurately correlated with events;

[0015] Step 5: Output the results of the management requirement validity assessment

[0016] By setting dynamic thresholds and multi-level early warning mechanisms, quantitative calculation results are transformed into intuitive management and drive intervention actions.

[0017] Step 1 includes:

[0018] 1) Data Acquisition and Preprocessing

[0019] The system obtains the original text information of management requirements from the source system, and performs data cleaning and standardization. Data collection is carried out through the standardized RESTful API provided by the source system to obtain management requirement data for a specified year in batches in a programmatic manner. If the above method is not feasible, PDF or Excel reports are exported, and text and table information are extracted using a document parsing library. After data collection, data cleaning and standardization are performed, including checking and handling missing values, removing irrelevant symbols and garbled characters from the text, unifying the date format, and establishing a standard dictionary to map synonyms and variant words in the original data to a unified standard.

[0020] 2) Machine-assisted structured information extraction

[0021] Standardized field information is extracted from management requirement text to achieve information structuring. Natural language processing technology and rule engine are used to automatically identify and extract the release time and applicable objects of management requirements, and extract and solidify other core attributes of management requirements, including: management requirement number, requirement type, corresponding event title, release date, management requirement description, and responsible unit.

[0022] 3) Data pattern definition and manual sorting

[0023] Define the data structure schema for management requirements, and then conduct manual review. First, manually remove interfering data. Domain experts review the original set of management requirements and remove obsolete, replaced, or irrelevant management requirement items to ensure the accuracy and timeliness of the data source. Second, manually review the validity judgment questions. For each management requirement, domain experts analyze its core purpose and the problem it is expected to solve, and transform it into multiple specific and measurable "validity judgment questions".

[0024] 4) Database integration and storage

[0025] Information that has been manually sorted and extracted by machines is integrated and structured into a database of management requirements according to a predefined data pattern.

[0026] Step 2 includes:

[0027] 1) Incident Report Preprocessing

[0028] Preprocessing of various monitoring event reports obtained from the experience feedback database provides high-quality input text for subsequent large-scale model analysis;

[0029] 2) Structured extraction of key information;

[0030] Extract entities and content that are crucial to the validity of management requirements from the cleaned text and store them in a structured event library;

[0031] 3) Text Vectorization and Indexing

[0032] Convert the full text or core paragraphs of the event report into numerical vectors and build an efficient vector index for subsequent rapid semantic similarity matching and pre-screening with management requirements.

[0033] The preprocessing process in step 2 includes:

[0034] (1) Format parsing: Use document parsing tools to extract the raw text stream from the document;

[0035] (2) Text cleaning: Remove irrelevant characters, including special symbols, garbled characters, excessive spaces and newlines, convert full-width characters to half-width characters, unify the case of English letters, and standardize the date and number formats;

[0036] (3) Key chapter segmentation: Based on the standard structure of nuclear power incident reports, key chapters in the report are identified and segmented through keyword matching or rule engine, including “Event Overview”, “Cause Analysis”, “Corrective Actions”, and “Lessons Learned”.

[0037] The key information extraction process in step 2 is as follows:

[0038] (1) Use the named entity recognition model or predefined regular expression rules to extract key entities and core content, including: power plant name, unit number, system / equipment, event occurrence time, event summary, basic information, description, consequences, failure / fault point determination, and analysis conclusions;

[0039] (2) The extracted information is filled into the corresponding fields according to the predefined database schema to form structured records.

[0040] The text vectorization and indexing process in step 2 is as follows:

[0041] (1) Document slicing

[0042] The slices are divided according to the chapters identified in the preprocessing, and each slice retains the metadata information of the chapter to which it belongs;

[0043] (2) Use a pre-trained language model to convert each text slice into a high-dimensional floating-point vector;

[0044] (3) Store the vectors of all event report text slices, along with their corresponding metadata, in a dedicated vector database.

[0045] The business rules in step 3 include the following three dimensions:

[0046] 1) Time range rule: The event must occur later than the effective time required by management;

[0047] 2) Coverage of power plant: The power plant where the incident occurred must be located within the scope of the applicable management requirements;

[0048] 3) Power plant completion status rule: The corrective action taken by the power plant in response to the management requirement must be marked as "completed".

[0049] Step 4 consists of the following three parts:

[0050] 1) Based on RAG context enhancement and retrieval, for each management requirement question to be judged, the most relevant evidence fragments are accurately located from massive event information to provide decision-making basis for large models;

[0051] 2) Large-scale intelligent question answering and reasoning guides large-scale models to perform logical analysis based on the provided context and make clear judgments for specific questions;

[0052] 3) Rule-based conversion from answer to score: A set of mapping rules is predefined to convert qualitative answers from large models into quantitative scores.

[0053] Step 5 includes the following two steps:

[0054] 1) Setting and adjusting the judgment threshold;

[0055] Based on historical data or expert experience, an initial judgment threshold T is set for the effectiveness score E. When E≥T, the management requirement is judged to be "insufficiently effective". A management interface is provided, allowing administrators to fine-tune the threshold T for the whole or specific categories according to changes in the overall safety performance of the group plant, the category of management requirements, or the distribution of judgment results over a period of time, in order to optimize the sensitivity and accuracy of the early warning.

[0056] 2) The generation and display of early warning information will be further improved by setting multi-level thresholds to achieve graded early warning.

[0057] The beneficial effects of this invention are as follows: By constructing a structured management requirements database, staff only need to sort out the problems that the management requirements aim to solve, forming a set of management requirements validity judgment questions. Intelligent technology is then used to automatically determine whether these problems exist in new events. Based on the weights and scoring rules set for different questions, a quantitative score for the validity judgment of the management requirements is calculated, ultimately providing a validity judgment conclusion. Compared to the previous method that relied entirely on manual judgment, applying large-scale model tools to the experience feedback work of nuclear power plants and using a constructive engineering approach to judge the effectiveness of management requirements implemented in nuclear power plants fully utilizes the information technology's ability to process large amounts of data and the reasoning ability of intelligent models, providing measurable judgment standards, making the judgment of management requirements validity more intelligent, efficient, and accurate. Attached Figure Description

[0058] Figure 1 The flowchart of a method for judging the validity of management requirements of nuclear power company clusters based on event characteristics provided by the present invention is shown. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] This invention relates to a method for judging the effectiveness of management requirements for nuclear power plant clusters based on a large-scale model. The core concept of the method lies in: constructing a structured management requirement library and utilizing the powerful reasoning capabilities of a large-scale language model to perform automated and intelligent correlation analysis and logical judgment between newly occurring events and historical management requirements, and finally outputting a validity conclusion in a quantitative manner, thereby achieving continuous monitoring and early warning of the effectiveness of management requirements.

[0061] A method for determining the validity of management requirements for nuclear power plant clusters based on event characteristics includes a design method for a cluster plant management requirement validity determination agent, a list of questions for determining the validity of management requirements, and rules for converting validity question answers into validity determinations. The method comprises the following steps:

[0062] Step 1: Build a structured management requirements database

[0063] By manually sorting and extracting information from structured fields, management requirements are extracted to construct a structured management requirements database. This specifically includes the following steps:

[0064] 1) Data Acquisition and Preprocessing. Obtain the raw text information of management requirements from the source system, and perform data cleaning and standardization to lay the foundation for subsequent structured processing. Data acquisition can be achieved programmatically through the standardized RESTful API provided by the source system to obtain management requirement data for a specified year in batches. If the above method is not feasible, export PDF or Excel reports and use document parsing libraries (such as pdfplumber, pypdf2 for PDF; pandas for Excel) to extract text and tabular information. After data acquisition, data cleaning and standardization are required, including checking and handling missing values, removing irrelevant symbols and garbled characters from the text, standardizing date formats, and establishing a standard dictionary to map synonyms and variant words in the raw data to a unified standard.

[0065] 2) Machine-Assisted Structured Information Extraction. Utilizing machine automation technology, standardized field information is extracted from management requirement text to achieve information structuring. Natural language processing technology and a rule engine are employed to automatically identify and extract key metadata such as the release time of management requirements (accurate to year, month, and day) and applicable objects (e.g., specific power plants, unit models, or the entire group of plants). Simultaneously, other core attributes of the management requirements are extracted and solidified, including but not limited to: management requirement number, requirement type, corresponding event title, release date, management requirement description, and responsible unit.

[0066] 3) Data Pattern Definition and Manual Review. Define the data structure pattern for management requirements, and based on this, conduct crucial manual review. First, manually remove interfering data. Domain experts review the original set of management requirements, removing obsolete, replaced, or irrelevant entries to ensure the accuracy and timeliness of the data source. Second, manually review validity judgment questions. For each management requirement, domain experts analyze its core purpose and the problem it aims to solve, transforming it into multiple specific, judgeable "validity judgment questions." These questions form the core basis for subsequent automated judgments by the large-scale model.

[0067] 4) Database Integration and Storage. Information that has been manually processed and extracted by machines is integrated and structured into a database of management requirements according to a predefined data model. Each management requirement record includes its corresponding "validity judgment question set" and all structured fields, providing an accurate and standardized data foundation for subsequent automated comparison with event feature vectors.

[0068] Step 2: Extraction and Storage of Event Report Information

[0069] This step aims to transform unstructured nuclear power plant incident report text into machine-readable, understandable, and computable structured event feature information through natural language processing, providing standardized data input for subsequent validity assessment. Specifically, it includes the following steps:

[0070] 1) Event Report Preprocessing. Various monitoring event reports (including operational event reports and internal event reports) obtained from the experience feedback database are preprocessed to provide high-quality input text for subsequent large-scale model analysis. The preprocessing process includes: (1) Format parsing: Using document parsing tools (such as Apache PDFBox for PDF, python-docx for Word) to extract the raw text stream from the document. This process needs to handle complex formats such as columns, headers and footers, and tables to ensure the correctness of the text order. (2) Text cleaning: Removing irrelevant characters, such as special symbols, garbled characters, excessive spaces, and line breaks. Converting full-width characters to half-width characters, unifying the case of English letters, and standardizing the formats of dates, numbers, etc. (3) Key Chapter Segmentation: Based on the standard structure of nuclear power event reports, key chapters in the report are identified and segmented through keyword matching or rule engines, including but not limited to "Event Overview," "Cause Analysis (Direct Cause, Root Cause)," "Corrective Actions," and "Lessons Learned." This lays the foundation for subsequent accurate information extraction.

[0071] 2) Structured extraction of key information. Entities and content crucial to the effectiveness of management requirements are precisely extracted from the cleaned text and stored in a structured event database. The specific process is as follows: (1) Using a Named Entity Recognition (NER) model or predefined regular expression rules, key entities and core content are extracted, including: power plant name, unit number, system / equipment, event occurrence time, event summary, basic information, description, consequences, failure / fault point determination, analysis conclusions, and other related information. (2) The extracted information is filled into the corresponding fields according to the predefined database schema to form structured records. For example, the database table may contain the following fields: event ID, report name, power plant, unit, system, equipment, occurrence time, event summary, failure point text, root cause text, consequence description, etc. This facilitates efficient and accurate querying and retrieval of event information by the large language model in subsequent steps.

[0072] 3) Text Vectorization and Indexing. Convert the full text or core paragraphs of the event report into numerical vectors (embeddings) and establish an efficient vector index for subsequent rapid semantic similarity matching and pre-screening with management requirements. The specific process is as follows: (1) Document Slicing. Since event reports are usually long and exceed the context window limit of most vectorization models, slicing is required. Usually, the text is divided according to the chapters identified in the preprocessing, and each slice retains the metadata information of its respective chapter. For example, the chapters such as "Event Description" and "Analysis Conclusion" are treated as independent text blocks. (2) Use a pre-trained language model (such as text2vec-large-chinese, Sentence-BERT, etc.) to convert each text slice into a high-dimensional floating-point vector (e.g., 768-dimensional or 1024-dimensional). This vector can capture the deep semantic information of the text. Texts with similar semantics (such as both describing "valve leakage") will have a closer distance (such as cosine similarity) in the high-dimensional space. (3) Store the vectors of all event report text slices, along with their corresponding metadata (such as event ID, slice source chapter, etc.), in a dedicated vector database (such as Chroma, Milvus, Pinecone). The vector database can achieve millisecond-level fast retrieval of massive vector data through the Approximate Nearest Neighbor (ANN) algorithm.

[0073] Step 3: Screening the Scope of Management Requirement Validity Judgment

[0074] This step aims to accurately filter a subset of management requirements related to the "current event" and requiring validity judgment from a structured management requirements database based on pre-defined business rules. This allows for focused computing resources and improved accuracy and efficiency in the judgment process. Its core is the automatic filtering of management requirements through a set of configurable rule engines. These rules specifically include the following three dimensions:

[0075] 1) Time range rule: The time of the event must be later than the effective time required by management.

[0076] 2) Coverage of power plant: The power plant where the incident occurred must be located within the scope of the applicable management requirements.

[0077] 3) Power plant completion status rule: The corrective action taken by the power plant in response to the management requirement must be marked as "completed".

[0078] The above rules must be executed sequentially using an "AND" logical relationship. The system will automatically and streamline the filtering of the attribute set of the "current event" and all records in the structured management requirements database according to the above rules. Finally, a list of candidate management requirements will be output, in which each management requirement fully satisfies all preset rules and serves as the direct object for deep semantic matching and validity scoring in subsequent steps.

[0079] Step 4: Management Requirements Validity Questions and Calculations

[0080] This step is the core intelligent reasoning process. By leveraging the deep semantic understanding and logical reasoning capabilities of a large language model, it accurately correlates and judges the pre-screened management requirements with the events. Specifically, it consists of the following three steps:

[0081] 1) Contextual Enhancement and Retrieval Based on RAG. For each management requirement question to be judged, the most relevant evidence fragments are accurately located from massive event information to provide sufficient decision-making basis for the large model. The specific engineering is as follows: (1) The "judgment question" in the management requirement (e.g., "Is there a main pump seal leak?") is converted into a query vector and an approximate nearest neighbor search is performed in the event vector database constructed in step 2. This step is based on the semantic understanding ability of the deep learning model to recall event text slices (such as event description, failure point, analysis conclusion, etc.) that are related to the question in "meaning". (2) The results of semantic retrieval are merged. In order to avoid information overload, a threshold is set to retain only the Top-K most relevant text slices (e.g., K=3 or 5). When merging, weighted fusion is performed based on similarity score, BM25 score, etc., and the final score is rearranged to ensure that the most critical evidence is placed first. (3) The rearranged Top-K event text slices, together with the complete description of management requirements and specific judgment questions, are combined into a structured prompt context, which constitutes the "reference material" for the large model to make judgments, effectively solving the knowledge blind spots or illusion problems that may exist in the large model.

[0082] 2) Large-scale intelligent question answering and reasoning. This guides the large-scale model to perform logical analysis based on the provided context, much like a seasoned expert, and to provide clear judgments for specific questions. The specific process is as follows:

[0083] (1) Prompt word construction: For each management requirement in the "Candidate Management Requirement List" and its corresponding "Validity Judgment Question Set (Q1, Q2, ..., Qi)", a structured prompt word is constructed. This prompt word contains the following parts:

[0084] System role setting: Assign a clear role to the large model, such as "You are a senior nuclear power safety analyst, responsible for rigorously judging whether there are specific problems pointed to by management requirements in the incident based on the provided incident report and management requirements".

[0085] Task definition: Give the large model a clear task, such as "Your task is to strictly answer the 'question' based on the following 'background information'".

[0086] Background Information / Context: [Management Requirement Description]: {Full text of management requirement} [Related Event Report Fragments]: {Top-K text slices retrieved from the previous RAG step}

[0087] Specific issue: Question: {Management requirement validity assessment issue, such as "Does this incident indicate a material aging problem in system XX at high temperatures?"}

[0088] Answer format and rules constraints:

[0089] Please make your judgment based solely on the provided background information and do not rely on external knowledge.

[0090] If your analysis suggests that the background information is sufficient to support a "yes" or "no" answer, please answer in the following format: **Result: Yes / No**.

[0091] If the background information does not mention the relevant issues at all, or if the information is insufficient to draw the following conclusions, please answer:

[0092] **Judgment result: Insufficient information**.

[0093] After your answer, please briefly explain your **understanding process**, citing key evidence from the background information.

[0094] (2) Large Model API Call: Using the constructed prompts, call the API of the large language model (such as Tongyi Qianwen, Deepseek, etc.). By setting appropriate parameters (such as temperature=0.1 to ensure the certainty of the answer rather than creativity), obtain the model's output.

[0095] (3) Output parsing: The large model returns a piece of natural language text. Regular expressions or rule-based parsers are needed to accurately extract the content (yes, no, or insufficient information) after "Judgment result:" from the returned text.

[0096] 3) Rule-based conversion from answers to scores. To ensure the objectivity and computability of the judgment results, a set of mapping rules needs to be predefined to convert the qualitative answers (Answer) of the large model into quantitative scores (Q). The rules are as follows:

[0097] (1) Predefined scoring rules: During system initialization, a weight and scoring rule are predefined for each management requirement validity judgment question (Q1, Q2, ..., Qi).

[0098] Weight (Wi): Represents the importance of the question in the overall validity judgment. For example, a question about "root cause" might have a weight of 0.8, while a question about "phenomenon" might have a weight of 0.4. The sum of the weights of all questions is usually 1.

[0099] Scoring rule: An explicit mapping dictionary.

[0100] Answer = "Yes", Score = Full marks Si (e.g., Si = 100, indicating that the problem no longer exists and management requirements may be valid)

[0101] Answer = "No", Score = 0 (because the problem still exists in the incident, indicating that management requirements failed to effectively prevent it).

[0102] Answer = "Uncertain", Score = 50 or marked as requiring manual review

[0103] (2) Score calculation: For each question Qi, the corresponding score Score_i is found from the scoring rule library based on the parsed large model answer.

[0104] For example, for question Q1 (weight W1 = 0.3), if the large model answers "yes", its score Score_1 = 100. The weighted score for this question is Weighted_Score_1 = W1 * Score_1 = 0.3 * 100 = 30.

[0105] (3) Results recording: Record the judgment results (answers, raw scores, weighted scores) of all questions to form a preliminary validity judgment vector of the management requirement for the current event: (Q1, Q2, ..., Qi) and its corresponding scores, and pass it to the final step 5 for aggregation and decision-making.

[0106] Step 5: Output the results of the management requirement validity assessment

[0107] This step transforms quantitative calculation results into intuitive management insights and drives intervention actions by setting dynamic thresholds and multi-level early warning mechanisms. Specifically, it includes the following two stages.

[0108] 1) Setting and Adjusting the Judgment Threshold. Based on historical data or expert experience, an initial judgment threshold T (e.g., T = 30) is set for the effectiveness score E. When E ≥ T, the management requirement is judged as "insufficiently effective". Simultaneously, the system provides a management interface that allows administrators to fine-tune the threshold T for the entire system or specific categories based on changes in the overall safety performance of the plant, the category of the management requirement (e.g., category A or category B), or the distribution of judgment results over a period of time, in order to optimize the sensitivity and accuracy of the early warning.

[0109] 2) Generation and display of early warning information. Multiple threshold levels can be set to achieve tiered early warning systems. For example:

[0110] Red alert (high risk): E≥50 indicates severe inadequacy of effectiveness, requiring immediate intervention.

[0111] Yellow alert (attention): T≤E<50, indicating potential risks. It is recommended to pay close attention and arrange for a review.

[0112] Green (Effective): 0 ≤ E < T, indicating that the management requirements are effective and no intervention is required.

[0113] The system automatically scans the calculation results of all candidate management requirements and identifies all management requirements that meet E ≥ T. On the graphical interface of the group plant experience feedback management system, they are centrally displayed through the warning list or dashboard. The displayed information shall at least include: management requirement number, management requirement description, corresponding event title, release date, warning information of different power plants. At the same time, a drill-down function is provided. When the user clicks on any warning item, they can immediately view the judgment basis (original evidence) provided by the large model in step 4, realizing the transparency and traceability of the judgment process and assisting the management in making final decisions.

[0114] Embodiment:

[0115] Such as Figure 1 shown, taking the effectiveness judgment of management requirements such as "preventing misoperation of valves" by a certain nuclear power group as an example, the effectiveness judgment of management requirements includes the following steps:

[0116] Step 1: Construct a structured management requirement database

[0117] Collect the existing management requirement data in the source system and perform text preprocessing. Through manual sorting and information extraction of structured fields, extract information such as the release time, object (power plant), and effectiveness judgment questions of the management requirements, and construct a structured management requirement database.

[0118] In the management requirement database, there is a management requirement MR-001, whose title is "Requirements on strengthening the isolation operation management of high-risk valves".

[0119] Its applicable object is "the whole group of power plants", and the effective date is January 1, 2023.

[0120] Domain experts sorted out two effectiveness judgment questions for it:

[0121] Q1: Does the root cause of the event include "insufficient valve isolation operation procedure or risk analysis"?

[0122] Q2: Did misoperation caused by unclear valve identification occur again in the event?

[0123] Step 2: Event report information extraction and storage

[0124] In May 2024, an internal event occurred in Power Plant A, and the event report was input into the source system.

[0125] The event report is retrieved from the source system, processed by natural language, and key information is extracted and stored in a structured event database. This database contains information such as: power plant name, unit number, system / equipment, event occurrence time, event summary, basic information, description, consequences, failure / fault point identification, and analysis conclusions.

[0126] Simultaneously, the event report is processed through document slicing and vectorization to output an event feature vector, which includes:

[0127] Root cause: ["Insufficient risk analysis of work instructions, failure to identify potential impacts on adjacent operating systems", "Unclear field valve labels"].

[0128] Keywords: ["false isolation", "valves with unclear labels"].

[0129] It is stored in a vector database to support subsequent semantic retrieval.

[0130] Step 3: Screening the Scope of Management Requirement Validity Judgment

[0131] Rule engine startup:

[0132] Time rule: Event time (May 2024) > Required effective time (January 2023) → Meets the requirements.

[0133] Scope rule: Event power plant (power plant A) ∈ Applicable objects (all plants) → Meets the criteria.

[0134] Status rule: According to the query, the corrective action for MR-001 by power plant A has been closed → Compliant.

[0135] Management requirement MR-001 has been included in the list of candidate management requirements.

[0136] Step 4: Management Requirements Validity Questions and Calculations

[0137] Based on the candidate management requirement list, for each management requirement question to be judged, the RAG retrieval enhancement service is invoked to accurately locate the most relevant evidence fragments from the vector event database. The most relevant current events are pre-screened according to topk=3 and threshold=0.1. The text slices of the rearranged top-3 events are combined with the complete description of the management requirement and the specific judgment question to form a structured prompt context dataset. This dataset contains the current events most relevant to management requirement MR-001, such as power plant A, power plant B, and power plant C. Specific fields include: event number, management requirement description, event content, validity judgment question 1, question score 1, validity judgment question 2, question score 2, validity judgment question 3, question score 3, and total score.

[0138] Using large language model tools such as Tongyi Qianwen 2.5 or Deepseek, based on the constructed prompt words, question-answering and reasoning analysis are performed on the two questions in MR-001.

[0139] The large model outputs the following after analyzing the provided context:

[0140] In power plant incident IR-2024-0156, MR-001, A1: For validity judgment question 1, the answer is "yes", and the reasoning process is "the root cause includes 'insufficient risk analysis of work instructions'".

[0141] In power plant incident IR-2024-0156, MR-001, A2: Validity judgment question 2, the answer is "yes". The reasoning process is that "both the root cause and the keywords mention 'valves with unclear labels' or 'valves with unclear labels'".

[0142] According to the conversion rule ("Yes" = 100 points), the system obtains the score set: {Q1 = 100, Q2 = 100}.

[0143] Result score calculation:

[0144] The default weighting coefficients for MR-001 are: given that both issues are very critical, set x1 = 0.5 and x2 = 0.5.

[0145] Calculate the validity score E: E = 1 * 0.5 + 1 * 0.5 = 100.

[0146] Step 5: Output the results of the management requirement validity assessment

[0147] The system preset threshold T = 30.

[0148] Judgment: E>T=30.

[0149] Red Alert: In the list of management requirement validity judgment results of the source system, Power Plant A is marked in red under this management requirement clause, indicating that: "Management requirement MR-001 (Requirement on Strengthening the Management of High-Risk Valve Isolation Operation) has insufficient validity score and too high (100 points). It is recommended to review it immediately!"

[0150] When managers click on the alarm, they can immediately view the judgment results and reasoning process provided by the big data model (i.e., the above-mentioned "insufficient risk analysis of work instructions" and "unclear valve labels"), thereby quickly understanding the reason for the warning and initiating the investigation and intervention process.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for judging the validity of management requirements of a nuclear power company's group of plants based on event characteristics, comprising a method for designing a group of plant management requirement validity judgment agent, a list of management requirement validity judgment questions, and conversion rules for validity question answers and validity judgments, characterized in that, Includes the following steps: Step 1: Build a structured management requirements database; By manually sorting and extracting information from structured fields, information on management requirements is extracted, and a structured management requirements database is constructed. Step 2: Extraction and storage of event report information; Through natural language processing, unstructured nuclear power incident report texts are transformed into machine-readable, understandable, and computable structured event feature information, providing standardized data input for subsequent validity assessments; Step 3: Screening the scope of management requirement validity assessment; Based on pre-defined business rules, a subset of management requirements that are associated with the "current event" and require validity judgment is selected from the structured management requirements database; Step 4: Management requirements validity Q&A and calculation; By utilizing large language models, pre-screened management requirements are accurately correlated with events; Step 5: Output the results of the validity assessment of management requirements; By setting dynamic thresholds and multi-level early warning mechanisms, quantitative calculation results are transformed into intuitive management and drive intervention actions.

2. The method for judging the effectiveness of nuclear power company group plant management requirements based on event characteristics as described in claim 1, characterized in that, Step 1 includes: 1) Data Acquisition and Preprocessing The system obtains the original text information of management requirements from the source system, and performs data cleaning and standardization. Data collection is carried out through the standardized RESTful API provided by the source system to obtain management requirement data for a specified year in batches in a programmatic manner. If the above method is not feasible, PDF or Excel reports are exported, and text and table information are extracted using a document parsing library. After data collection, data cleaning and standardization are performed, including checking and handling missing values, removing irrelevant symbols and garbled characters from the text, unifying the date format, and establishing a standard dictionary to map synonyms and variant words in the original data to a unified standard. 2) Machine-assisted structured information extraction Standardized field information is extracted from management requirement text to achieve information structuring. Natural language processing technology and rule engine are used to automatically identify and extract the release time and applicable objects of management requirements, and extract and solidify other core attributes of management requirements, including: management requirement number, requirement type, corresponding event title, release date, management requirement description, and responsible unit. 3) Data pattern definition and manual sorting Define the data structure schema for management requirements, and then conduct manual review. First, manually remove interfering data. Domain experts review the original set of management requirements and remove obsolete, replaced, or irrelevant management requirement items to ensure the accuracy and timeliness of the data source. Second, manually identify validity judgment questions. For each management requirement, domain experts analyze its core purpose and the problem it is expected to solve, and transform it into multiple specific and measurable "validity judgment questions". 4) Database integration and storage Information that has been manually sorted and extracted by machines is integrated and structured into a database of management requirements according to a predefined data pattern.

3. The method for judging the effectiveness of nuclear power company group plant management requirements based on event characteristics as described in claim 1, characterized in that, Step 2 includes: 1) Incident Report Preprocessing Preprocessing of various monitoring event reports obtained from the experience feedback database provides high-quality input text for subsequent large-scale model analysis; 2) Structured extraction of key information; Extract entities and content that are crucial to the validity of management requirements from the cleaned text and store them in a structured event library; 3) Text Vectorization and Indexing Convert the full text or core paragraphs of the event report into numerical vectors and build an efficient vector index for subsequent rapid semantic similarity matching and pre-screening with management requirements.

4. The method for judging the effectiveness of nuclear power company group plant management requirements based on event characteristics as described in claim 3, characterized in that, The preprocessing process in step 2 includes: (1) Format parsing: Use document parsing tools to extract the raw text stream from the document; (2) Text cleaning: Remove irrelevant characters, including special symbols, garbled characters, excessive spaces and newlines, convert full-width characters to half-width characters, unify the case of English letters, and standardize the date and number formats; (3) Key chapter segmentation: Based on the standard structure of nuclear power incident reports, key chapters in the report are identified and segmented through keyword matching or rule engine, including "incident overview", "cause analysis", "corrective actions" and "lessons learned".

5. The method for judging the validity of nuclear power company group plant management requirements based on event characteristics as described in claim 3, characterized in that, The key information extraction process in step 2 is as follows: (1) Use the named entity recognition model or predefined regular expression rules to extract key entities and core content, including: power plant name, unit number, system / equipment, event occurrence time, event summary, basic information, description, consequences, failure / fault point determination, and analysis conclusions; (2) The extracted information is filled into the corresponding fields according to the predefined database schema to form structured records.

6. The method for judging the validity of nuclear power company group plant management requirements based on event characteristics as described in claim 3, characterized in that, The text vectorization and indexing process in step 2 is as follows: (1) Document slicing The slices are divided according to the chapters identified in the preprocessing, and each slice retains the metadata information of the chapter to which it belongs; (2) Use a pre-trained language model to convert each text slice into a high-dimensional floating-point vector; (3) Store the vectors of all event report text slices, along with their corresponding metadata, in a dedicated vector database.

7. The method for judging the validity of nuclear power company group plant management requirements based on event characteristics as described in claim 1, characterized in that, The business rules in step 3 include the following three dimensions: 1) Time range rule: The event must occur later than the effective time required by management; 2) Coverage of power plant: The power plant where the incident occurred must be located within the scope of the applicable management requirements; 3) Power plant completion status rule: The corrective action taken by the power plant in response to the management requirement must be marked as "completed".

8. The method for judging the validity of nuclear power company group plant management requirements based on event characteristics as described in claim 1, characterized in that, Step 4 consists of the following three parts: 1) Based on RAG context enhancement and retrieval, for each management requirement question to be judged, the most relevant evidence fragments are accurately located from massive event information to provide decision-making basis for large models; 2) Large-scale intelligent question answering and reasoning guides large-scale models to perform logical analysis based on the provided context and make clear judgments for specific questions; 3) Rule-based conversion from answer to score: A set of mapping rules is predefined to convert qualitative answers from large models into quantitative scores.

9. The method for judging the validity of nuclear power company group plant management requirements based on event characteristics as described in claim 1, characterized in that, Step 5 includes the following two steps: 1) Setting and adjusting the judgment threshold; Based on historical data or expert experience, an initial judgment threshold T is set for the effectiveness score E. When E≥T, the management requirement is judged to be "insufficiently effective". A management interface is provided, allowing administrators to fine-tune the threshold T for the whole or specific categories according to changes in the overall safety performance of the group plant, the category of management requirements, or the distribution of judgment results over a period of time, in order to optimize the sensitivity and accuracy of the early warning. 2) The generation and display of early warning information will be further improved by setting multi-level thresholds to achieve graded early warning.

Images