Automatic analysis system and method for failure mode and influence analysis of nuclear power plant

Abstract

The invention relates to the field of nuclear power simulation, in particular to an automatic analysis system and method for failure mode and influence analysis of a nuclear power plant. The method comprises the following steps: step 1, constructing a nuclear power equipment knowledge graph, and integrating equipment structures, functions, failure modes and historical data; 2, automatically identifying a potential failure mode based on a machine learning model, and associating an influence path of the potential failure mode; integrating the machine learning model with digital twinborn simulation software to realize failure deduction and verification in a simulation environment; 3, after failure classification and influence path analysis are carried out on the accessed real-time data, a multi-criterion decision algorithm is adopted to calculate the risk priority, and optimization suggestions are output. According to the method, the automatic identification of the failure mode, the dynamic reasoning of the influence path and the quantitative evaluation of the risk priority are realized by constructing an intelligent analysis framework combining the knowledge graph and the machine learning.

Description

Technical Field

[0001] This invention relates to the field of nuclear power simulation, and in particular to an automatic analysis system and method for failure mode and effects analysis of nuclear power plants. Background Technology

[0002] Failure Mode and Effects Analysis (FMEA) is a key method for safety assessment in nuclear power plants, used to identify potential failure modes in equipment, analyze their impact, and develop preventative measures. Traditional FMEA primarily relies on manual analysis, using expert experience combined with historical data to evaluate system functional units item by item. Its basic principle is qualitative or semi-quantitative analysis based on structured tables (such as failure mode lists, severity classifications, etc.). In recent years, with the development of digital technology, some studies have attempted to introduce computer-aided analysis tools, but most remain limited to data management or limited automation, failing to achieve fully intelligent processing throughout the entire process.

[0003] Traditional methods have significant limitations: manual analysis is inefficient, especially for complex systems like nuclear power plants, involving a large workload and prone to oversights; analysis results are influenced by subjective factors, resulting in poor consistency; and dynamic update capabilities are insufficient, making it difficult to adapt to real-time changes in equipment status or operating environment. Furthermore, existing automation tools largely rely on fixed rule bases, lacking flexibility and unable to effectively integrate multi-source data (such as sensor monitoring and maintenance records) for adaptive optimization. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an automatic analysis system and method for failure mode and impact analysis of nuclear power plants. By constructing an intelligent analysis framework that combines knowledge graph and machine learning, it realizes automatic identification of failure modes, dynamic reasoning of impact paths, and quantitative assessment of risk priorities.

[0005] This invention provides an automated analysis method for failure mode and effects analysis (FEM) of nuclear power plants, comprising the following steps:

[0006] Step 1: Construct a knowledge graph for nuclear power equipment, integrating equipment structure, function, failure modes, and historical data;

[0007] Step 2: Automatically identify potential failure modes based on machine learning models and associate their impact paths; integrate the machine learning models with digital twin simulation software to realize failure simulation and verification in a simulation environment;

[0008] Step 3: After classifying failures and analyzing impact paths based on the accessed real-time data, a multi-criteria decision-making algorithm is used to calculate risk priorities and output optimization suggestions.

[0009] In a specific embodiment of the present invention, step 1 specifically includes:

[0010] Step 1-1: Perform data integration. For structured data, use ETL tools to map it to ontology attributes. For unstructured data, use NLP technology to extract entity relationships. Expert experience is entered through structured interview templates.

[0011] Steps 1-2: Perform ontology modeling, using the Protégé tool to construct the ontology, define core classes and relationships, and establish attribute constraints;

[0012] Steps 1-3: Perform knowledge storage, using Neo4j graph database to store triples, supporting Cypher queries; back up as RDF files compatible with SPARQL.

[0013] In a specific embodiment of the present invention, step 2 specifically includes:

[0014] Step 2-1: Develop a machine learning model;

[0015] Step 2-2: Integrate the machine learning model with the digital twin simulation software. When the machine learning model predicts a certain failure situation, call the digital twin simulation software to simulate it and write back the simulation indicators to the threshold attribute of the nuclear power equipment knowledge graph.

[0016] Steps 2-3: Compare the simulation results with real accident records. If the error is greater than 20%, trigger the machine learning model to be retrained.

[0017] In a specific embodiment of the present invention, step 2-1 specifically comprises:

[0018] The static and dynamic characteristics were analyzed from the integrated data.

[0019] The static and dynamic features are associated with historical facts and domain knowledge to generate a dataset including failure labels and impact paths; the dataset is then divided into training, validation, and test sets according to time order.

[0020] Analyze the distribution of various types of samples in the dataset, identify rare failure modes, and formulate targeted SMOTE data augmentation strategies accordingly; apply SMOTE only to the training set to generate synthetic samples, and after verifying their rationality, form a class-balanced training dataset, while ensuring that the validation set and test set maintain their original distribution for reliable evaluation.

[0021] We design a shared encoding layer that integrates static features and dynamic temporal features, extract deep representations through an LSTM network, and configure regularization to prevent overfitting.

[0022] A dual-branch output architecture is constructed based on shared features, with a Softmax branch for failure classification and an impact path prediction branch based on GAT designed respectively.

[0023] Configure a multi-task training environment, including a balanced combination of loss functions, optimizer and learning rate scheduling strategies, and hyperparameter settings to prevent overfitting;

[0024] The constructed machine learning model is trained using the dataset.

[0025] In a specific embodiment of the present invention, step 3 specifically includes:

[0026] Step 3-1: Use the AHP (Analytic Hierarchy Process) to assign weights and determine the relative importance weights of the three criteria: safety, economy, and environment.

[0027] Step 3-2: Convert the linguistic evaluation of each alternative into triangular fuzzy numbers and defuzzify them. Combine the AHP weights to construct a standardized weighted decision matrix to prepare for the ranking of the alternatives.

[0028] Step 3-3: By calculating the relative closeness of each scheme to the positive and negative ideal solutions, prioritize the alternative schemes and finally output a decision list of risk handling order;

[0029] Steps 3-4: Based on the decision list, use SPARQL to intelligently retrieve, filter, and sort a list of matching maintenance measures from the maintenance knowledge base;

[0030] Steps 3-5: Using the retrieved specific measures data, automatically generate a standardized report with a complete structure, standardized format, and including risk levels and specific recommendations through a template method;

[0031] Steps 3-6: Automatically encapsulate the core content of the generated suggestion report into a maintenance work order, create a task by calling the JIRA API and pre-fill key fields, and simultaneously realize the tracking and management of work order status.

[0032] Steps 3-7: Construct a streaming data processing pipeline. For the edge computing layer, deploy a lightweight LSTM model to calculate the device health index in real time; perform initial screening on the rule engine, where the health index is <0.6 and the temperature is >90°C.

[0033] When C is selected, it is marked as high risk.

[0034] Steps 3-8: Conduct central system analysis, construct equipment dependency graph based on the association relationship of nuclear power equipment knowledge graph, use PageRank algorithm to identify key nodes and perform cascading risk detection; the dynamic probability update method uses Kalman filter to fuse predicted probability and real-time health index, and updates the failure probability in Neo4j at preset time intervals;

[0035] Steps 3-9: Receive external system alarms based on the REST API, map them to the EnvironmentalHazard class of the knowledge graph; define the joint risk value R_combined = max(R_local, 0.3*R_external), and output the cross-system risk matrix.

[0036] This invention provides an automated analysis system for failure mode and effects analysis (FEM) of nuclear power plants, comprising:

[0037] Nuclear power equipment knowledge graph, used to integrate equipment structure, function, failure modes and historical data;

[0038] Failure Mode Predictor (FMRP) is used to automatically predict failure modes of nuclear power equipment and perform impact analysis of these failure modes.

[0039] The dynamic data updater is used to collect sensor data and operation and maintenance logs in real time, and adjust the failure probability by combining time series analysis.

[0040] The risk priority calculator is used to calculate the risks that may be caused by failure modes of nuclear power equipment, sort the risks, and provide optimization suggestions for the risks.

[0041] A visual interactive interface is used to process human commands and receive analysis results from system feedback.

[0042] In one specific embodiment of the present invention, the failure mode predictor includes:

[0043] The core prediction and identification module is used to automatically identify potential failure modes and construct impact paths from device time-series data. It extracts deep features through TCN temporal convolutional network and attention mechanism, and outputs multidimensional failure probability distribution and propagation topology.

[0044] The digital twin integration and verification module is used to perform failure simulation and verification in a simulation environment. It interacts with the digital twin platform in real time through the API interface to simulate the failure propagation process and compare the prediction accuracy to ensure that the model output conforms to physical laws and engineering reality.

[0045] The uncertainty reasoning and optimization module is used to quantify the uncertainty of prediction results and achieve adaptive optimization. It integrates deep learning output and expert prior knowledge through a Bayesian network to provide confidence assessment and trigger online learning and model updates based on feedback.

[0046] The data support and quality assurance module is used to provide high-quality training data and rigorous performance verification. It builds a failure case library based on the knowledge graph, ensures that the F1-score is greater than 0.85 through cross-validation, and continuously monitors the knowledge graph coverage and data quality.

[0047] The system integration and interface module is used to achieve end-to-end system integration and business interface. It connects to real-time data streams and external systems through standardized interfaces, provides early warning notifications and decision support services, and monitors system performance to ensure SLA < 500ms response requirements.

[0048] In one specific embodiment of the present invention, the risk priority calculator introduces fuzzy logic to handle the quantification of qualitative indicators.

[0049] The present invention provides a computer-readable storage medium storing program instructions for performing the method.

[0050] This invention provides an electronic device for automatic analysis of nuclear power plant failure modes and effects, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the automatic analysis method for nuclear power plant failure modes and effects.

[0051] Compared with existing technologies, the automatic analysis system and method for nuclear power plant failure mode and effects analysis (FMEA) of the present invention reconstructs the FMEA process through intelligent technology, providing core methodological support for the digital transformation of nuclear power plant safety management, and has the following beneficial effects:

[0052] (1) Automated analysis significantly reduces manual time consumption, covers hidden failure modes that are easily missed by traditional methods, and improves the efficiency and comprehensiveness of analysis;

[0053] (2) Combine real-time data to dynamically adjust risk assessment, enhance the ability to respond to changes in equipment status and new types of failures, and improve the dynamic adaptability of the analysis process;

[0054] (3) Use knowledge graphs and machine learning to reduce subjective bias, provide quantitative and traceable analysis results, and support more scientific and accurate operation and maintenance decisions. Attached Figure Description

[0055] Figure 1 A flowchart illustrating the automated analysis method for failure modes and effects analysis of nuclear power plants;

[0056] Figure 2 This diagram shows the module composition of the automatic analysis system for failure modes and effects analysis of nuclear power plants. Detailed Implementation

[0057] To further understand the present invention, embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the present invention.

[0058] Embodiments of the present invention disclose an automated analysis method for failure mode and effects analysis of nuclear power plants, such as... Figure 1 As shown, it includes the following steps:

[0059] Step 1: Construct a knowledge graph for nuclear power equipment, integrating equipment structure, function, failure modes, and historical data;

[0060] Specifically, it includes:

[0061] Step 1-1: Perform data integration. For structured data, use ETL tools to map it to ontology attributes. For unstructured data, use NLP technology to extract entity relationships. Expert experience is entered through structured interview templates.

[0062] The method for processing and integrating the structured data, such as SAP maintenance work orders (CSV), SCADA sensor metadata (JSON), and equipment BOM tables (Excel), uses Apache Spark to perform field mapping. For example, the equipment code in the work order is associated with the serial number in the BOM table, and missing values ​​are filled with expert rules, such as "default lifespan of pump equipment = 15 years".

[0063] The unstructured data includes technical documents and expert experience. When parsing technical documents, the LayoutLM model is used to identify tables and diagrams in the PDF and extract equipment hierarchy relationships, such as "main cooling pump → impeller assembly". Expert experience is digitized, and OWL templates are designed, such as <Expert A>: suggestionsInspectionInterval: centrifugal pump "once a month" and entered into the database.

[0064] Steps 1-2: Perform ontology modeling, using the Protégé tool to construct the ontology, define core classes and relationships, and establish attribute constraints;

[0065] The core classes include device classes, function classes, and failure mode classes.

[0066] The relationship includes having functionality and causing failure;

[0067] Attribute constraints can be established by referring to IEEE Std 1584 and nuclear power industry standards (such as RCM);

[0068] Steps 1-3: Perform knowledge storage, using Neo4j graph database to store triples, supporting Cypher queries; back up as RDF files compatible with SPARQL.

[0069] Neo4j is used to create indexes to accelerate queries and optimize the graph database; Git-LFS is used to manage OWL file changes, and difference comparison tools, such as OWLDiff, are used to assist expert review. 。

[0070] Step 2: Automatically identify potential failure modes based on machine learning models and associate their impact paths; integrate the machine learning models with digital twin simulation software to realize failure simulation and verification in a simulation environment;

[0071] Specifically, it includes:

[0072] Step 2-1: Develop a machine learning model;

[0073] Specifically:

[0074] Static and dynamic characteristics were extracted from the integrated data; the static characteristics included equipment age and material properties, and the dynamic characteristics included the vibration RMS value over the past 30 days.

[0075] The static and dynamic features are associated with historical facts and domain knowledge to generate a dataset including failure labels and impact paths; the dataset is then divided into training, validation, and test sets according to time order.

[0076] Analyze the distribution of various types of samples in the dataset, identify rare failure modes, and formulate targeted SMOTE data augmentation strategies accordingly; apply SMOTE only to the training set to generate synthetic samples, and after verifying their rationality, form a class-balanced training dataset, while ensuring that the validation set and test set maintain their original distribution for reliable evaluation.

[0077] We design a shared encoding layer that integrates static features and dynamic temporal features, extract deep representations through an LSTM network and configure regularization to prevent overfitting;

[0078] A dual-branch output architecture is constructed based on shared features, with a Softmax branch for failure classification and an impact path prediction branch based on GAT designed respectively.

[0079] Configure a multi-task training environment, including a balanced combination of loss functions, optimizer and learning rate scheduling strategies, and hyperparameter settings to prevent overfitting;

[0080] The constructed machine learning model is trained using the dataset.

[0081] Step 2-2: Integrate the machine learning model with the digital twin simulation software. When the machine learning model predicts a certain failure situation, call the digital twin simulation software to simulate it and write back the simulation indicators to the threshold attribute of the nuclear power equipment knowledge graph.

[0082] Digital twin simulation software can be Unity3D or ANSYS;

[0083] Steps 2-3: Compare the simulation results with real accident records. If the error is greater than 20%, trigger the machine learning model to be retrained.

[0084] Step 3: After classifying failures and analyzing impact paths based on the accessed real-time data, a multi-criteria decision-making algorithm is used to calculate risk priorities and output optimization suggestions.

[0085] Specifically, it includes:

[0086] Step 3-1: Use the AHP (Analytic Hierarchy Process) to assign weights and determine the relative importance weights of the three criteria: safety, economy, and environment.

[0087] Experts compare the "safety / economy / environment" criteria in pairs, and the criteria take effect when the conformity ratio (CR) is less than 0.1.

[0088] Step 3-2: Convert the linguistic evaluation of each alternative into triangular fuzzy numbers and defuzzify them. Combine the AHP weights to construct a standardized weighted decision matrix to prepare for the ranking of the alternatives.

[0089] Step 3-3: By calculating the relative closeness of each scheme to the positive and negative ideal solutions, prioritize the alternative schemes and finally output a decision list of risk handling order;

[0090] Steps 3-4: Based on the decision list, use SPARQL to intelligently retrieve, filter, and sort a list of matching maintenance measures from the maintenance knowledge base;

[0091] Steps 3-5: Using the retrieved specific measures data, automatically generate a standardized report with a complete structure, standardized format, and including risk levels and specific recommendations through a template method;

[0092] Steps 3-6: Automatically encapsulate the core content of the generated suggestion report into a maintenance work order, create a task by calling the JIRA API and pre-fill key fields, and simultaneously realize the tracking and management of work order status.

[0093] Steps 3-7: Construct a streaming data processing pipeline. For the edge computing layer, deploy a lightweight LSTM model to calculate the device health index in real time; perform initial screening on the rule engine, where the health index is <0.6 and the temperature is >90°C.

[0094] When C is selected, it is marked as high risk.

[0095] Steps 3-8: Conduct central system analysis, construct equipment dependency graph based on the association relationship of nuclear power equipment knowledge graph, use PageRank algorithm to identify key nodes, and perform cascading risk detection; the dynamic probability update method uses Kalman filter to fuse predicted probability and real-time health index, and updates the failure probability in Neo4j at preset time intervals; for example, update failureProbability in Neo4j every 5 minutes.

[0096] Steps 3-9: Receive external system alarms based on the REST API, map them to the EnvironmentalHazard class of the knowledge graph; define the joint risk value R_combined = max(R_local, 0.3*R_external), and output the cross-system risk matrix.

[0097] When implementing the method, a Prometheus tracking model is used to predict latency (SLA < 500ms), and a Grafana dashboard displays the knowledge graph coverage in real time for performance monitoring. An industrial firewall is used to strictly isolate operational and information technology network data flows, and PKI digital signature verification is implemented for knowledge graph update operations, thus constructing an end-to-end data security protection system.

[0098] Embodiments of the present invention disclose an automated analysis system for failure mode and effects analysis of nuclear power plants, such as... Figure 2 As shown, it includes:

[0099] A nuclear power equipment knowledge graph is used to integrate equipment structure, function, failure modes, and historical data; it adopts an ontology modeling approach and integrates nuclear power equipment technical documents, maintenance records, and expert experience.

[0100] Failure Mode Predictor (FMRP) is used to automatically predict failure modes of nuclear power equipment and perform impact analysis of these failure modes.

[0101] The failure mode predictor includes:

[0102] The core prediction and identification module is used to automatically identify potential failure modes and construct impact paths from device time-series data. It extracts deep features through TCN temporal convolutional network and attention mechanism, and outputs multidimensional failure probability distribution and propagation topology.

[0103] The digital twin integration and verification module is used to perform failure simulation and verification in a simulation environment. It interacts with the digital twin platform in real time through the API interface to simulate the failure propagation process and compare the prediction accuracy to ensure that the model output conforms to physical laws and engineering reality.

[0104] The uncertainty reasoning and optimization module is used to quantify the uncertainty of prediction results and achieve adaptive optimization. It integrates deep learning output and expert prior knowledge through a Bayesian network to provide confidence assessment and trigger online learning and model updates based on feedback.

[0105] The data support and quality assurance module is used to provide high-quality training data and rigorous performance verification. It builds a failure case library based on the knowledge graph, stores at least 100,000 samples, ensures F1-score > 0.85 through cross-validation, and continuously monitors the knowledge graph coverage and data quality.

[0106] The system integration and interface module is used to achieve end-to-end system integration and business interface. It connects to real-time data streams and external systems through standardized interfaces, provides early warning notifications and decision support services, and monitors system performance to ensure SLA < 500ms response requirements.

[0107] The dynamic data updater is used to collect sensor data and operation and maintenance logs in real time, and adjust the failure probability by combining time series analysis.

[0108] The real-time processing pipeline data sources include sensors and operation logs. Sensor data is collected via the OPC UA protocol at a frequency ≥1Hz (Apache Kafka streaming buffer). Operation logs are parsed in real-time using a Natural Language Processing (NLP) model (Logstash + custom regular expression templates). Then, time-series analysis is performed, and anomalies are detected using a sliding window (STL decomposition + isolated forest) to trigger probability adjustment events. The update strategy is to smooth the prediction results using Kalman filtering, and write them to the knowledge graph only when the confidence level is >90%. The disaster recovery design of this module is to cache locally (SQLite) during network outages and incrementally synchronize after recovery.

[0109] The risk priority calculator is used to calculate the risks that may be caused by failure modes of nuclear power equipment, sort the risks, and provide optimization suggestions for the risks.

[0110] The risk priority calculator introduces fuzzy logic to handle the quantification of qualitative indicators.

[0111] Its specific operation mode is as follows: the three risk indicators of severity (S), probability of occurrence (O) and detectability (D) are fuzzified, and the precise numerical values ​​are converted into fuzzy linguistic variables by using triangular membership functions in order to better simulate the expert decision-making thinking;

[0112] Risk assessment is performed by building an "IF-THEN" type rule base based on Mamdani fuzzy reasoning. For example, when a rule detects the condition combination of "severity = high AND probability of occurrence > 0.3", it automatically triggers the conclusion of "risk priority = critical", realizing a reasoning process similar to that of human experts.

[0113] Using case-based reasoning (CBR) technology, specific and feasible optimization suggestions are generated by matching current risk characteristics with similar patterns in a historical case database. It automatically retrieves validated measures from similar historical risk scenarios, such as targeted suggestions like "increasing infrared detection frequency" and "strengthening vibration monitoring."

[0114] The module ultimately outputs analysis results in two forms: on the one hand, it dynamically generates an intuitive risk matrix heatmap using Matplotlib, visually displaying the level distribution of each risk point; on the other hand, it outputs a detailed list of recommendations in standard JSON format, facilitating integration with other systems and further processing.

[0115] A visual interactive interface is used to process human commands and receive analysis results from system feedback.

[0116] The manual intervention interface uses a drag-and-drop graph editing (D3.js + React), allowing users to intuitively edit knowledge elements such as device relationships and failure paths. All manual modifications are fully recorded through the Git version control system to ensure the traceability and rollbackability of the knowledge evolution process.

[0117] The prediction results cover manually labeled error cases. When there is a discrepancy between the system's predictions and the actual situation, experts can annotate the error cases. This labeled data automatically triggers the model's active learning process, enabling continuous optimization based on human feedback and constantly improving prediction accuracy.

[0118] The system integrates Three.js technology to dynamically render the 3D model of the device and maps real-time risk data to model components. Users can view detailed failure chain analysis by clicking on specific components. Furthermore, the timeline tool built using Plotly Dash supports visual comparison and analysis of predicted results and actual failure scenarios.

[0119] Based on role-based differentiated operational permissions: the engineer level is primarily responsible for correcting and verifying basic data to ensure data quality; the expert level has the authority to modify the ontology model and can optimize the system's core cognitive structure based on domain knowledge. This hierarchical mechanism ensures both the system's professionalism and the standardization of knowledge updates.

[0120] The entire system is integrated using a microservice architecture, including Kubernetes and Docker. Inter-module communication uses gRPC to achieve front-end and back-end separation. The back-end uses FastAPI, and the front-end uses Vue3 and TypeScript. The deployment method is industrial-grade servers (meeting the IEC 62264 standard), and lightweight prediction models (TensorFlow Lite) are deployed on edge nodes.

[0121] An embodiment of the present invention discloses a computer-readable storage medium storing program instructions for performing the methods described above.

[0122] An embodiment of the present invention discloses an electronic device for automatic analysis of nuclear power plant failure modes and effects analysis, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform an automatic analysis method for nuclear power plant failure modes and effects analysis as described above.

[0123] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0124] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An automated analysis method for failure mode and effects analysis (FEM) of nuclear power plants, characterized in that, Includes the following steps: Step 1: Construct a knowledge graph for nuclear power equipment, integrating equipment structure, function, failure modes, and historical data; Step 2: Automatically identify potential failure modes based on machine learning models and associate their impact paths; Integrating machine learning models with digital twin simulation software enables failure simulation and verification in a simulation environment; Step 3: After classifying failures and analyzing impact paths based on the accessed real-time data, a multi-criteria decision-making algorithm is used to calculate risk priorities and output optimization suggestions.

2. The automatic analysis method for nuclear power plant failure mode and effects analysis according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: Perform data integration. For structured data, use ETL tools to map it to ontology attributes. For unstructured data, use NLP technology to extract entity relationships. Expert experience is entered through structured interview templates. Steps 1-2: Perform ontology modeling, using the Protégé tool to construct the ontology, define core classes and relationships, and establish attribute constraints; Steps 1-3: Perform knowledge storage, using Neo4j graph database to store triples, supporting Cypher queries; back up as RDF files compatible with SPARQL.

3. The automatic analysis method for nuclear power plant failure mode and effects analysis according to claim 1, characterized in that, Step 2 specifically includes: Step 2-1: Develop a machine learning model; Step 2-2: Integrate the machine learning model with the digital twin simulation software. When the machine learning model predicts a certain failure situation, call the digital twin simulation software to simulate it and write back the simulation indicators to the threshold attribute of the nuclear power equipment knowledge graph. Steps 2-3: Compare the simulation results with real accident records. If the error is greater than 20%, trigger the machine learning model to be retrained.

4. The automatic analysis method for nuclear power plant failure mode and effects analysis according to claim 3, characterized in that, Step 2-1 specifically involves: The static and dynamic characteristics were analyzed from the integrated data. The static and dynamic features are associated with historical facts and domain knowledge to generate a dataset including failure labels and impact paths; the dataset is then divided into training, validation, and test sets according to time order. Analyze the distribution of various types of samples in the dataset, identify rare failure modes, and formulate targeted SMOTE data augmentation strategies accordingly; apply SMOTE only to the training set to generate synthetic samples, and after verifying their rationality, form a class-balanced training dataset, while ensuring that the validation set and test set maintain their original distribution for reliable evaluation. We design a shared encoding layer that integrates static features and dynamic temporal features, extract deep representations through an LSTM network, and configure regularization to prevent overfitting. A dual-branch output architecture is constructed based on shared features, with a Softmax branch for failure classification and an impact path prediction branch based on GAT designed respectively. Configure a multi-task training environment, including a balanced combination of loss functions, optimizer and learning rate scheduling strategies, and hyperparameter settings to prevent overfitting; The constructed machine learning model is trained using the dataset.

5. The automatic analysis method for nuclear power plant failure mode and effects analysis according to claim 1, characterized in that, Step 3 specifically includes: Step 3-1: Use the AHP (Analytic Hierarchy Process) to assign weights and determine the relative importance weights of the three criteria: safety, economy, and environment. Step 3-2: Convert the linguistic evaluation of each alternative into triangular fuzzy numbers and defuzzify them. Combine the AHP weights to construct a standardized weighted decision matrix to prepare for the ranking of the alternatives. Step 3-3: By calculating the relative closeness of each scheme to the positive and negative ideal solutions, prioritize the alternative schemes and finally output a decision list of risk handling order; Steps 3-4: Based on the decision list, use SPARQL to intelligently retrieve, filter, and sort a list of matching maintenance measures from the maintenance knowledge base; Steps 3-5: Using the retrieved specific measures data, automatically generate a standardized report with a complete structure, standardized format, and including risk levels and specific recommendations through a template method; Steps 3-6: Automatically encapsulate the core content of the generated suggestion report into a maintenance work order, create a task by calling the JIRA API and pre-fill key fields, and simultaneously realize the tracking and management of work order status. Steps 3-7: Construct a streaming data processing pipeline. For the edge computing layer, deploy a lightweight LSTM model to calculate the device health index in real time; perform initial screening on the rule engine, and mark it as high risk when the health index is <0.6 and the temperature is >90℃. Steps 3-8: Conduct central system analysis, construct equipment dependency graph based on the association relationship of nuclear power equipment knowledge graph, use PageRank algorithm to identify key nodes and perform cascading risk detection; the dynamic probability update method uses Kalman filter to fuse predicted probability and real-time health index, and updates the failure probability in Neo4j at preset time intervals; Steps 3-9: Receive external system alarms based on the REST API, map them to the EnvironmentalHazard class of the knowledge graph; define the joint risk value R_combined = max(R_local, 0.3*R_external), and output the cross-system risk matrix.

6. An automated analysis system for failure modes and effects analysis of nuclear power plants, characterized in that, include: Nuclear power equipment knowledge graph, used to integrate equipment structure, function, failure modes and historical data; Failure Mode Predictor (FMRP) is used to automatically predict failure modes of nuclear power equipment and perform impact analysis of these failure modes. The dynamic data updater is used to collect sensor data and operation and maintenance logs in real time, and adjust the failure probability by combining time series analysis. The risk priority calculator is used to calculate the risks that may be caused by failure modes of nuclear power equipment, sort the risks, and provide optimization suggestions for the risks. A visual interactive interface is used to process human commands and receive analysis results from system feedback.

7. The automatic analysis system for nuclear power plant failure mode and effects analysis according to claim 6, characterized in that, The failure mode predictor includes: The core prediction and identification module is used to automatically identify potential failure modes and construct impact paths from device time-series data. It extracts deep features through TCN temporal convolutional network and attention mechanism, and outputs multidimensional failure probability distribution and propagation topology. The digital twin integration and verification module is used to perform failure simulation and verification in a simulation environment. It interacts with the digital twin platform in real time through the API interface to simulate the failure propagation process and compare the prediction accuracy to ensure that the model output conforms to physical laws and engineering reality. The uncertainty reasoning and optimization module is used to quantify the uncertainty of prediction results and achieve adaptive optimization. It integrates deep learning output and expert prior knowledge through a Bayesian network to provide confidence assessment and trigger online learning and model updates based on feedback. The data support and quality assurance module is used to provide high-quality training data and rigorous performance verification. It builds a failure case library based on the knowledge graph, ensures that the F1-score is greater than 0.85 through cross-validation, and continuously monitors the knowledge graph coverage and data quality. The system integration and interface module is used to achieve end-to-end system integration and business interface. It connects to real-time data streams and external systems through standardized interfaces, provides early warning notifications and decision support services, and monitors system performance to ensure SLA < 500ms response requirements.

8. The automatic analysis system for nuclear power plant failure mode and effects analysis according to claim 6, characterized in that, The risk priority calculator introduces fuzzy logic to handle the quantification of qualitative indicators.

9. A computer-readable storage medium, characterized in that, It stores program instructions for executing the automatic analysis method for nuclear power plant failure mode and effects analysis as described in any one of claims 1-5.

10. An electronic device for automatic analysis of failure modes and effects in nuclear power plants, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the automatic analysis method for nuclear power plant failure mode and effects analysis as described in any one of claims 1 to 5.

Images