A data warehouse architecture system for nuclear power digital twins
By constructing a data warehouse structure system for nuclear power digital twins, the problems of data quality and correlation in nuclear power digital twin systems have been solved, enabling efficient data integration and analysis, supporting the optimization of equipment operation and maintenance and business decision-making, and promoting the data-driven transformation of nuclear power.
Patent Information
- Application Number
- CN202511183022.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-22
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-08-22
AI Technical Summary
The existing nuclear power digital twin system suffers from inconsistent data quality, sparse value density, and limited correlation, resulting in incomplete data asset management, loss of high-quality data assets, and low utilization efficiency.
Design a data warehouse structure system for nuclear power digital twins, including a data processing layer, a data integration layer, a semantic enhancement layer, and an analysis and application layer. Through data standardization, encoding mapping, data fusion, feature labeling, and algorithm analysis, construct digital twin thematic libraries at the equipment level, system level, and structure level to achieve efficient data integration and analysis.
This has enabled system optimization of equipment operation and maintenance, shifting from scheduled inspections to predictive maintenance. It has improved data analysis efficiency and the data positioning capabilities of business departments, and promoted the transformation of nuclear power from experience-driven to data-driven.
Smart Images

Figure CN121166822B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automated management technology for nuclear power data assets, and in particular to a data warehouse structure system for nuclear power digital twins. Background Technology
[0002] With the deepening of digital transformation, the application of digital twin technology in the nuclear power industry is gradually becoming an important means to improve operational efficiency and safety. Digital twins can enable refined management of nuclear power plant equipment, systems, and structures, real-time monitoring and fault diagnosis of nuclear power systems and equipment, support adjustments to operational strategies and decision-making, provide early warnings and rapid responses, ensure the stable operation of nuclear power plant units, and improve the economic benefits and safety level of nuclear power plants. The digital twin system collects and analyzes real-time unit operation data, nuclear power plant digital control system (DCS) data, test instrument data, equipment sensor data, inspection data, simulator simulation data, and digital twin operation data. Utilizing multi-source heterogeneous data information, it establishes relevant analytical and mechanistic models, forming a virtual mapping of actual equipment and systems, simulating the operation of the physical power plant. Currently, the nuclear power industrial internet platform has integrated massive amounts of data from industrial systems and equipment, achieving centralized data management across multiple regions and sites, supporting the full lifecycle business data connectivity of multiple reactor types and units, and serving as a data aggregation and management foundation covering all nuclear power scenarios.
[0003] However, in existing technologies, the massive multimodal data resources required in digital twin systems often suffer from problems such as inconsistent data quality, sparse value density, and limited correlation, making it difficult to effectively tap their potential value. Furthermore, in current nuclear power plant operation and maintenance practices, there is a lack of comprehensive inventory and effective management of data assets. The methods for opening and sharing data assets are singular and lack flexibility, making data retrieval and utilization inconvenient, resulting in the loss of high-quality data assets and low utilization efficiency. Summary of the Invention
[0004] This invention provides a data warehouse structure system for nuclear power digital twins, which addresses the problems of inconsistent sample data quality, sparse value density, and limited correlation in existing digital twin technologies.
[0005] The technical solution of the present invention is as follows:
[0006] This invention proposes a data warehouse structure system for nuclear power digital twins. The system includes a data processing layer, a data integration layer, a semantic enhancement layer, and an analysis and application layer. The data processing layer reads the raw data from the nuclear power industrial internet platform and performs data standardization processing on the raw data to form standard data. Based on the standard data, a nuclear power industry theme library is constructed. The data integration layer performs data encoding mapping, data classification, and data fusion on the data in the nuclear power industry theme library to form equipment-level digital twin theme libraries, system-level digital twin theme libraries, and structure-level digital twin theme libraries. The semantic enhancement layer labels the data in the nuclear power industry theme library and the data in the theme libraries formed by the data integration layer as data association feature tags, and constructs a structured feature tag library from the data with associated feature tags. The analysis and application layer reads the data in the feature tag library, and constructs a digital twin algorithm library and a digital twin sample library through algorithm analysis and training samples.
[0007] In some embodiments, the data processing layer maps the raw data to standard data through pre-defined mapping rules, mapping the raw data to unified business terms, parameter naming rules, and equipment codes; the data processing layer divides the standard data into business domains such as subject domain grouping, subject domain, business object, and data entity, and divides them into equipment, operation, safety and quality, maintenance, control, detection, fuel, and design business domains, and uses relational database, time series database, and object database subject modeling and storage for the standard data.
[0008] In some embodiments, the data integration layer will associate the SSC codes with data objects of various types, such as system operation, equipment monitoring and analysis, and structure-level monitoring and management, based on the SSC codes and equipment codes extracted from the subject library, and associate the equipment codes with the component level of the SSC codes; and establish mapping rules between SSC codes and equipment codes to realize cross-library association of the three libraries: the equipment-level digital twin subject library, the system-level digital twin subject library, and the structure-level digital twin subject library.
[0009] In some embodiments, the semantic enhancement layer is the data of the nuclear power industry theme library and the data integration layer form the theme library, which are associated with the feature tags of equipment status, mechanism, operating condition and risk emergency response, and a tag knowledge graph is formed based on the feature tag library.
[0010] In some embodiments, the analysis application layer includes a data analysis module and a data sample creation module. The data analysis module provides analysis algorithms to analyze and process data from the feature tag library and construct a digital twin algorithm library. The data analysis module includes a time-series data analysis unit, a spatial-physical correlation analysis unit, a multimodal data fusion analysis unit, and a business scenario analysis unit. The time-series data analysis unit reads real-time sensor data of nuclear power equipment from the feature tag library, and uses methods such as time series decomposition, trend prediction, and anomaly detection to analyze the changing patterns of equipment parameters over time, predict future changes in equipment parameters, and determine the health trend of the equipment. The spatial-physical correlation analysis unit reads 3D model data of structures and spatial location information of equipment from the feature tag library, performs spatial topology analysis and physical field coupling analysis on the 3D model data of structures and spatial location information of equipment, correlates the spatial location and physical state data of structures and equipment, and analyzes the performance degradation parameters of structures and equipment. The multimodal data fusion analysis unit reads multimodal data from the feature tag library and performs cross-modal feature alignment on the multimodal data.
[0011] In some embodiments, the business scenario analysis unit reads real-time data of the nuclear power plant system from the feature tag library, performs fault diagnosis and root cause analysis on the real-time data based on historical fault cases, and uses association rule mining algorithms and causal inference algorithms to locate the system fault propagation path; the business scenario analysis unit performs system operation optimization analysis on the real-time data by combining multi-objective optimization algorithms with physical constraints, analyzes the impact of system operating parameters on energy consumption and efficiency, and outputs parameter adjustment schemes; the business scenario analysis unit reads equipment health, environmental parameters, historical accident data, etc. from the feature tag library, performs risk warning analysis, and uses risk matrices and risk algorithm models to calculate the probability of potential risks to equipment.
[0012] In some embodiments, the data sample creation module constructs a digital twin sample library by processing data from the feature label library. The data sample creation module includes a sample data processing unit, a sample enhancement unit, a sample annotation unit, and a sample partitioning unit. The sample data processing unit forms sample data for the digital twin and performs data preparation and preprocessing operations according to the business requirements of the digital twin. The sample enhancement unit performs noise addition, time shifting, stretching and compression operations on the normal operating time-series data of the feature label library, as well as correlation and fusion of multimodal data. The sample annotation unit combines the feature labels of the semantic enhancement layer to annotate the sample data processed by the data processing unit with business labels. The sample partitioning unit uses stratified sampling to divide the data samples annotated by the sample annotation unit into training set, validation set, and test set.
[0013] In some embodiments, the sample data processing unit includes a time-series data sample management subunit and an anomaly data processing unit. The time-series data sample management subunit has data access, benchmark data selection, sample data processing, and sample management functions. The data access function reads data with associated feature tags from a feature tag library. The benchmark data selection function selects data from the data access function based on a time range, verifies the selected data, and uses the verified data determined by the user as benchmark data. The sample data processing function prepares and preprocesses the benchmark data according to business requirements, extracts key features from the benchmark data, and forms sample data. The sample management function retrieves the sample data, encapsulates the sample dataset according to requirements, forms sample files, and sends them to a specified path. The time-series data sample management subunit stores data independently for each project, with each project corresponding to multiple sample data sets. The time-series data sample management subunit sets up a shared database, and the sample data in the shared database is shared by all projects. The time-series data sample management subunit sets up four data-isolated storage areas to store raw data, benchmark data, intermediate data, and sample data, respectively. The sample data processing function reads the data of the associated feature labels from the feature label library as the original data, the data selected by the benchmark data selection function as the benchmark data, the data generated in the sample label unit annotation as the intermediate data, and the data processed by the sample data processing function as the sample data.
[0014] In some embodiments, the abnormal data processing unit uses statistical methods, model methods, or time series downsampling methods to process the sample data, identify and remove abnormal data, write the abnormal data identification results into the sample data database, and perform linear interpolation and normalization processing on the sample data after removing abnormal data; the statistical methods include the normal distribution method and the box plot method, and identify data that exceed 3 times the standard deviation of the normal distribution and data that exceed 1.5 times the box length of the quantile as abnormal data; the model methods identify outliers through density clustering algorithm, isolated forest algorithm, and random cut forest algorithm; the time series downsampling method downsamples the sample data from the time series and outputs time series data with reduced sampling frequency.
[0015] In some embodiments, the data verification of the benchmark data selection function specifically includes data correlation test, multicollinearity test, normality test, time series stationarity test, dead value test, limit value test, and repeatability test; the data preprocessing of the sample data processing function includes missing value handling, outlier detection, data augmentation, data imbalance handling, data normalization, and categorical data encoding.
[0016] In some embodiments, the sample labeling unit of the data sample creation module includes a label rule management function and an automatic labeling function for business data. The label rule management function configures label rule templates, provides rule preview and rule copying, checks rule logic, and binds rule templates to specified label categories. The automatic labeling function for business data is set by the user according to the rule template library. The rule engine is responsible for parsing the label rules, generating a directed acyclic graph of tasks, and performing filtering, calculation, and labeling on the sample data through user-defined functions.
[0017] The implementation of this invention has the following beneficial effects:
[0018] This invention proposes a data warehouse structure system for nuclear power digital twins. This system is based on the underlying support of the nuclear power industrial internet platform and a four-layer progressive architecture, and is designed around the business characteristics of nuclear power digital twins. Through the collaborative integration of data resources in the four-layer structure, it supports equipment operation and maintenance from regular inspection to predictive maintenance, system optimization from experience-based control to intelligent collaboration, and structural safety from post-detection to long-term early warning.
[0019] This invention proposes a data warehouse structure system for nuclear power digital twins. This system forms a data thematic library through a data integration layer and links the data between thematic libraries to achieve data connectivity between "systems, equipment and structures". Business departments do not need to deal with data heterogeneity issues and can directly and quickly locate the entire life cycle data through standard coding, thereby improving analysis efficiency.
[0020] This invention proposes a data warehouse structure system for nuclear power digital twins. This system adds decision-making knowledge information to the data through a semantic enhancement layer and sample annotation function. Through the closed loop of data, knowledge and decision, it promotes the transformation of nuclear power from experience-driven to data-driven. Attached Figure Description
[0021] Figure 1 This is a system architecture diagram of a data warehouse structure for nuclear power digital twins proposed in this invention;
[0022] Figure 2 This is a structural diagram of the time-series data sample management subunit of a data warehouse structure system for nuclear power digital twins proposed in this invention;
[0023] Figure 3 This is a flowchart of the time-series data sample management subunit of a data warehouse structure system for nuclear power digital twins proposed in this invention;
[0024] Figure 4 This is a schematic diagram of data processing in a data warehouse structure system for nuclear power digital twins proposed in this invention. Detailed Implementation
[0025] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Digital twins are digital representations of target entities, requiring data connections to achieve synchronization and integration with the target entity at an appropriate frequency. To meet the digitalization needs of nuclear power plants in design, engineering construction, operation and maintenance, aging management, decommissioning, and environmental governance, digital twins can be used to represent various entity objects in systems, equipment, and structures within the nuclear power plant. The nuclear power digital twin data warehouse structure proposed in this invention accesses multi-source heterogeneous data from the nuclear power industrial internet platform, performing layered processing such as data standardization, data fusion, data annotation, data analysis, and data sample creation. It organizes equipment master data, digital twin business domain knowledge, expert experience, and combines them with existing data resources to digitize business knowledge, forming high-quality digital twin data resources such as quantitative expert rules, business knowledge models, subject libraries, thematic libraries, and sample libraries. This solves the difficulties in integrating and analyzing multi-source heterogeneous data in digital twin applications.
[0027] This invention discloses a data warehouse architecture system for nuclear power digital twins. Based on the standardized, high-quality data resources provided in the nuclear power digital twin data warehouse, it achieves the discovery, screening, and access of data assets in the digital twin data warehouse through a data asset access layer; it enables the visualization of asset catalogs and asset information management through a data asset management layer, providing asset lists and enabling asset release, application, and approval; and it provides various service methods according to the needs of applications and business personnel through a data asset service layer, integrating data value into business products and services, and fully leveraging the role of data assets in promoting business implementation.
[0028] like Figures 1 to 4 As shown, this invention proposes a data warehouse structure system for nuclear power digital twins. This system constructs an industrial data acquisition, cleaning, transformation, transmission and storage base covering the entire nuclear power scenario through a nuclear power industrial internet platform, realizing the aggregation of nuclear power industrial data across regions and multiple bases, and supporting the connection of business data throughout the entire life cycle of multiple reactor types and multiple units.
[0029] The system comprises a data processing layer, a data integration layer, a semantic enhancement layer, and an analysis and application layer. The data processing layer reads raw data from the nuclear power industrial internet platform. This raw data includes unit operation data, test instrument data, DCS system data, equipment sensor data, production management data, inspection data, temperature and wastewater flow data, environmental pollutant concentration distribution data, simulator data, structure aging status data, operating procedures, operating outlines, power plant system data, inspection reports, and other data. It accesses the raw data from the nuclear power industrial internet platform using Extract, Transform, Load (ETL) tools and performs data standardization processing. The raw data consists of industrial data required for the nuclear power digital twin in the nuclear power industrial internet platform, including equipment, operation, safety and quality, maintenance, control, testing, fuel, and design. The system deduplicates, completes, and converts the format of the raw data. Pre-defined mapping rules map the raw data to standard data, and a nuclear power industry theme library is constructed using this standard data. The mapping rules, based on digital twin business domain knowledge and expert experience, map the raw data to unified business terminology, parameter naming rules, and equipment codes.
[0030] The nuclear power industry thematic library progressively divides standard data into levels such as thematic domain grouping, thematic domain, business objects, and data entities. Based on nuclear power business processes, the standard data is broken down into business domains such as equipment, operation, safety and quality, maintenance, control, testing, fuel, and design. The data processing layer defines the thematic domain boundaries of the nuclear power industry thematic library, clarifies the field range of each data model, and allocates the divided standard data into data structure models. The nuclear power industry thematic library uses thematic modeling and storage of standard data using relational databases, time-series databases, and object databases to ensure business alignment and traceability of the data.
[0031] The data integration layer is used to extract data from the nuclear power industry thematic library. The data integration layer uses coding mapping, data classification, hierarchical fusion and topology modeling to build a digital twin thematic library of equipment, system and structure in three dimensions.
[0032] For coding mapping, the data integration layer uses SSC codes and equipment codes (SSC stands for Structure, System, Component; a general term for structures, systems, and equipment in a nuclear power plant) extracted from the thematic library to associate data objects of various types in the nuclear power industry thematic library, such as system operation, equipment monitoring and analysis, and structure-level monitoring and management, through SSC codes. It also associates equipment codes with the component level of the SSC codes, constructing a four-level coding system of "structure-system-component-equipment" to achieve data connectivity and full lifecycle management of equipment, facilitating dynamic maintenance of equipment information. Simultaneously, mapping rules between SSC codes and equipment codes are established to enable cross-library association among the equipment-level digital twin thematic library, the system-level digital twin thematic library, and the structure-level digital twin thematic library. These rules are periodically reviewed by nuclear power business personnel to ensure unique codes and accurate hierarchical relationships.
[0033] The data integration layer, based on standardized nuclear power industry thematic database data, categorizes the database data into three dimensions: equipment, systems, and structures. It then integrates this data according to these dimensions, using data fusion technologies such as multi-source data association and time-series data alignment to form equipment-level, system-level, and structure-level digital twin thematic databases. The data integration layer connects the equipment-level, system-level, and structure-level data links, achieving deep integration of individual equipment data, system-level data, and structure data to form the multi-dimensional thematic datasets required for digital twins.
[0034] The semantic enhancement layer annotates data from the nuclear power industry thematic library and the thematic library formed by the data integration layer. Combining the business knowledge of digital twin experts, it associates data with feature tags related to equipment status, mechanisms, operating conditions, and risk emergency response. Semantic annotation processing is performed on the data to establish a semantic index oriented towards equipment status, operating mechanisms, operating conditions, risks, and emergency responses. This endows the data with machine-interpretable semantic meaning, and a structured feature tag library is constructed from the data with associated feature tags. The semantic enhancement layer forms a tag knowledge graph, providing a unified semantic benchmark for digital twin simulation, analysis, and decision-making. Data assets enhanced with semantic tags can be used for cross-system and cross-level data analysis and model training in digital twins.
[0035] The analysis application layer reads data from the feature label library, provides analysis algorithms and training samples to support prediction, optimization, and decision support for business applications, and builds a digital twin algorithm library and a digital twin sample library. The analysis application layer is the core link connecting data assets and business value. It mines data patterns through data analysis and builds analysis algorithm libraries and sample library resources to support digital twin scenarios.
[0036] The analysis application layer includes a data analysis module and a data sample creation module. The data analysis module provides analysis algorithms to analyze and process data from the feature tag library and build a digital twin algorithm library. The data analysis module includes a time-series data analysis unit, a spatial-physical correlation analysis unit, a multimodal data fusion analysis unit, and a business scenario analysis unit. The time-series data analysis unit reads real-time sensor data (temperature, pressure, vibration, flow, etc.) from the feature tag library of nuclear power equipment. It uses methods such as time series decomposition, trend prediction, and anomaly detection to analyze the changing patterns of equipment parameters over time, predict future changes in equipment parameters, and determine the health trend of the equipment. The spatial-physical correlation analysis unit reads 3D model data of structures and spatial location information of equipment from the feature tag library. The spatial-physical correlation analysis unit performs spatial topology analysis and physical field coupling analysis on the above data to correlate the spatial location and physical state data of structures and equipment, and analyzes the performance degradation parameters of structures and equipment. This is achieved by correlating three types of parameters according to the unique identifier of the equipment or structure. Spatial location includes: region, floor elevation, three-dimensional coordinates of equipment / structure and its functional zone; physical state includes: vibration, temperature, strain, cracks, etc.; performance degradation includes efficiency reduction, vibration intensity exceeding limits, temperature anomalies or increased inter-story drift angle, etc., thereby obtaining the mapping relationship of "spatial location-physical state-performance degradation" for structural integrity assessment.
[0037] The multimodal data fusion and analysis unit reads multimodal data from the feature label library, including sensor data, maintenance visual images, and operation log text data. It then fuses this data and applies cross-modal feature alignment to address the incomplete information in a single data dimension. Specifically, the multimodal data fusion and analysis unit includes three levels of alignment: First, coarse-grained alignment based on timestamps—unifying sensor time-series data, image capture time, and log timestamps to the same clock source, ensuring millisecond-level consistency; second, association based on spatial coordinates—using the device's unique ID to link target bounding boxes in visual images with sensor measurement points and log entries; and third, alignment based on shared latent features—using a pre-trained structure to map image patches and text tokens to the same semantic space, considering entities with a cosine similarity greater than a preset value as the same entity.
[0038] The business scenario analysis unit reads business feature tag data from the feature tag library for nuclear power plants. Based on historical failure cases, it performs fault diagnosis and root cause analysis on the aforementioned real-time data. The unit employs association rule mining and causal inference algorithms to locate system fault propagation paths, supporting rapid fault localization. The unit also uses multi-objective optimization algorithms combined with physical constraints to perform system operation optimization analysis on the aforementioned real-time data, analyzing the impact of system operating parameters on energy consumption and efficiency, and outputting parameter adjustment plans. Finally, the unit reads equipment health, environmental parameters, and historical accident data from the feature tag library. Based on this data, it performs risk warning analysis, using risk matrices and risk algorithm models to calculate the probability of potential equipment risks, supporting preventative maintenance.
[0039] The data sample creation module constructs a digital twin sample library by processing data from the feature label library. This module includes a sample data processing unit, a sample augmentation unit, a sample annotation unit, and a sample segmentation unit. The sample data processing unit generates sample data for the digital twin, performing data preparation and preprocessing operations according to the business requirements of the digital twin.
[0040] The sample augmentation unit addresses the scarcity of samples for extreme nuclear power plant operating conditions (such as transient accidents and severe accidents) by expanding the sample data volume through data augmentation operations such as noise addition, time shifting, stretching and compression on normal operating condition time-series data, and correlation and fusion of multimodal data. The sample augmentation unit performs noise addition (adding noise data to simulate sensor errors), time shifting (translating the time axis to simulate start-up / shutdown delays), and stretching and compression operations on normal operating condition time-series data. For multimodal data augmentation, the unit can correlate equipment information with real-time equipment status data to generate data under different equipment operating conditions; and fuse the spatial location data of the 3D model with physical field data to generate data of different locations and states.
[0041] The sample labeling unit, combining the feature labels from the semantic enhancement layer, automatically labels the sample data processed by the data processing unit with business tags, which are then verified by business personnel to ensure accuracy. The sample partitioning unit uses stratified sampling to divide the labeled data into training, validation, and test sets, ensuring a consistent proportion of faulty and normal samples, as well as samples from different equipment types.
[0042] In practice, for sample data, users need to perform processing (using algorithms such as alignment, slicing, missing data imputation, outlier handling, and data augmentation) and automatic labeling (batch labeling of data according to logic), among other big data processing techniques. For example... Figure 2 and Figure 3As shown, the sample data processing unit includes a time-series data sample management subunit, which has functions such as data access, benchmark data selection, sample data processing, and sample management, as detailed below:
[0043] The data access function reads data associated with feature tags from the feature tag library. The data includes: time series data accessed from the time series database, data files exported from the simulator, accident sequence files in the Modular Accident Analysis Program for Nuclear Emergency (MAAP), and initial condition (IC) related data generated using nuclear emergency dedicated software, etc.
[0044] The baseline data selection function reads data from the data input function, selects data within a specified time range, and performs a validity check on the input data content as the baseline data for subsequent annotation and processing. To avoid the introduction of noise data and maintain the data content unchanged, only data validation is performed. Data validation includes format processing, removal of blank rows, duplicate row detection, and data validity checks. The user decides whether to remove duplicate rows and whether to use the input data based on the data quality. The data validity check provides verification of the quantity and logical relationships between data columns. The conditions to be met for verification are given according to the business requirements for time-series data, including but not limited to: material balance relationships, energy conservation laws, electrical laws, thermodynamic laws, and other related equation constraints; the range of process parameters for equipment operation under normal conditions; and other equations or inequalities related to equipment and operating conditions. After validity checks and deduplication, the data confirmed by the user is used as the baseline data for the sample data.
[0045] like Figure 4 As shown, the data validation of the benchmark data selection function specifically includes: data correlation test, which calculates the correlation coefficient between benchmark data variables to determine whether the input features are related to the target variable and removes irrelevant or redundant features; multicollinearity test, which uses statistical tests to find that the benchmark data are collinear and then uses the variance inflation coefficient to quantify it, removing redundant features that cause instability in the benchmark data model; normality test, which uses K-Sigma or Jarque-Bera tests to determine whether the benchmark data follows a normal distribution, providing a basis for subsequent parameter testing and outlier threshold setting; time series stationarity test, which uses visualization and ADF test to determine whether the benchmark data sequence is stationary, and if it is non-stationary, it needs to be differencing or transformed to meet the model requirements; dead value test, limit value test, and duplicate row test, which quickly scans whether the values exceed the physical range or appear repeatedly, ensuring that the data is logically correct and unique.
[0046] The sample data processing function performs several optional data preparation and preprocessing operations on the selected benchmark data content according to business requirements, extracting key features from the data to form sample data. These include time-domain and frequency-domain features of time-series data, keyword features of text data, and topological features of structure data, thus reducing data dimensionality.
[0047] like Figure 4 As shown, the data preprocessing part of the sample data processing function includes: missing value handling, manually filling key missing values in the sample data, and imputing consecutive missing values in forward or backward or smoothly to maintain temporal continuity and trend consistency; outlier detection, processing the sample data through statistical methods, model methods, or time series downsampling methods in the outlier processing unit to identify outliers and decide whether to correct or remove them; data augmentation, generating new samples through jittering, window transformation, dynamic time warping (DTW), etc., to expand small sample data or improve the robustness of sample data; data imbalance handling, adjusting the proportion of sample data categories using undersampling, oversampling, or interpolation to prevent sample data from being biased towards the majority class; data normalization, scaling sample data features to a uniform dimension using Min-Max or Z-Score to accelerate sample data convergence and reduce the influence of dimensions; and categorical data encoding, establishing a mapping table from sample data to numbers to safely convert discrete categories into a numerical form that the sample data can process.
[0048] The sample management function retrieves sample datasets and encapsulates them according to the requirements of model training and data analysis, outputting adapted sample files that can be published to specified databases, folders, external application platforms, etc.
[0049] The time-series data sample management subunit breaks down its functionality into independently running microservices. Sample data across different projects is completely independent and isolated. However, after processing by the sample data processing function, the sample data is sent to a shared database, allowing all projects to share sample data within the shared database. Baseline data is named using the format of project ID + measurement point ID and stored in separate tables; sample data is named using a globally unique sample ID and stored in separate tables. The time-series data sample management subunit is deployed as a microservice on a Kubernetes cluster. Each project has its own namespace, and each project contains multiple sample data sets stored in CSV files. Cross-project sharing is achieved through MinIO object storage. The time-series data sample management subunit has four isolated storage areas: raw data, baseline data, intermediate data, and sample data. The raw data is data read from the feature label library by the sample data processing function; the baseline data is data selected by the baseline data selection function; the intermediate data is data generated during annotation by the sample labeling unit; and the sample data is data processed by the sample data processing function.
[0050] The sample data processing unit also includes an anomaly data processing unit, used for anomaly data processing and data cleaning. It performs linear interpolation imputation and normalization on the sample data after removing anomalies, thereby eliminating missing and outlier values, making the data more complete and ensuring sample data quality. Anomalies refer to values in the sample data that significantly deviate from other data in the same data column. The anomaly data processing unit supports the identification and removal of anomalies, writing the anomaly identification results into the metadata database of the relevant sample data. The methods used by the anomaly data processing unit to process anomalies include statistical methods, model methods, and time series downsampling, specifically including:
[0051] Statistical methods include assuming that the sample data follows a normal distribution, based on the 3σ (3 times the standard deviation) principle, where 99.73% of the data samples are distributed within (μ-3σ, μ-3σ), and data outside this range can be considered outliers; using the box plot method, generally taking the 25th percentile Q1 and 75th percentile Q3 of all sample values, with the distance between the two being the box length IQR, and considering sample data less than Q1-1.5IQR or greater than Q3+1.5IQR as outliers;
[0052] The model methods identify outliers using algorithms such as density-based clustering (DBSCAN), isolated forest (IF), and randomly cut forest (RCF). These methods require external modeling platforms.
[0053] Time series downsampling methods process sample data by downsampling the time series data, meeting the requirements for modeling using sequence data with different sampling frequencies. The downsampling method filters the data by specifying a new sampling interval based on the original sequence sampling frequency, outputting new time series data with a reduced sampling frequency.
[0054] Taking reactor coolant flow measurement points as an example, the abnormal data processing unit first uses the isolated forest algorithm to identify outliers, then performs linear interpolation to complete the data after removing outliers, and finally normalizes it using Z-Score to ensure that the sample distribution conforms to N(0,1).
[0055] The sample labeling unit of the data sample creation module includes label rule management and automatic labeling of business data. The label rule management function configures label rule templates, provides rule previews and copying, checks rule logic, and binds rule templates to specified label categories. The automatic labeling of business data automatically labels sample data according to its characteristics. The sample labeling unit uses the DolphinScheduler platform to run automatic labeling rules, enabling the system to automatically complete the batch data labeling process. In the automatic labeling of business data, users set label rules based on the rule template library. The rule engine is responsible for parsing the label rules, generating a directed acyclic graph (DAG), and performing filtering, calculation, and labeling on multimodal sample data such as time series, Boolean values, and images through User-Defined Functions (UDFs), allowing users to upload their own custom rule scripts. The sample labeling unit has multiple data-isolated storage areas to store raw data, baseline data, intermediate data, and sample data respectively.
[0056] The automatic labeling business tagging function specifically includes: tag rule configuration, where users configure tag rules in the rule template library, including the tag name and tagging logic, such as the tag name being "full power" and the parameters being the factory, unit, and logical parameters; creating an automatic labeling task, where users select the tag rule for the sample labeling unit and the factory, unit, and logical parameters are displayed, and users can modify the factory and unit parameters in the task interface; and running the task, where the automatic labeling business tagging function generates the ID code of the measurement point associated with the equipment based on the business logic and input parameter combination, and breaks down the logic into multiple conditions, filtering and tagging the sample data according to the conditions.
[0057] The rule template library contains 28 templates across 6 categories: T-class (Time-align): used to downsample second-level time series to minute-level; S-class (State): Boolean state inversion identification, such as "voltage regulator heater ON / OFF"; V-class (Value): analog quantity limit detection, supporting hysteresis thresholds; D-class (Derivative): identification of heating and pressure ramp rates based on first-order differential; C-class (Composite): multi-condition AND, OR, and NOT, used for initial LOCA judgment; F-class (Failure): fault mode matching based on expert knowledge. Specifically, it includes: time alignment (gap-fill, resampling), threshold filtering (single threshold, double threshold, hysteresis threshold), trend identification (pure rise, pure fall, plateau), state machine (number of Boolean inversions, duration), statistical features (moving mean, variance, kurtosis), and composite operating conditions (multi-condition AND / OR / NOT), etc.
[0058] User-defined rules are described in two layers. For example, the expression layer supports SQL-like Boolean expressions such as ${RCS_P}>15.5&${RCS_T}>300. UDF calculation supports referencing user-uploaded functions, such as UDF_LOCA(ds,{"slope_th":2.1}).
[0059] The node types of the Directed Acyclic Graph (DAG) generated by the rules engine include: DataReadNode: reads baseline data; UDFNode: calls built-in or custom UDFs; LabelWriteNode: writes labels to intermediate data; AuditNode: if the label contains sensitive fields (related to kernel defects), it is transferred to the manual review queue.
[0060] As a preferred embodiment, the standard rule operation logic of the automatic annotation function is as follows:
[0061] 1. Verify the validity of the set tag rules through the tag verification interface;
[0062] 2. Parse the business logic in the tag rules, combine the input parameters, generate the ID code of the measurement point associated with the device, parse the >, &, || and other logic in the user-defined rules, and generate the judgment logic for each measurement point.
[0063] 3. Label the time periods and measurement points that meet the requirements to generate labeled data.
[0064] 4. Call the data saving interface to save the tag data.
[0065] The automatic annotation process is illustrated below through a specific example:
[0066] The user-selected data sources and data types include: S1: primary coolant temperature, pressure, and flow rate, with a sampling period of 1 second; S2: eddy current detection image of the heat transfer tube of the steam generator, resolution 1024×512, JPG format; S3: ultrasonic thickness measurement data of the voltage regulator oscillation tube, CSV format, with a sampling period of 1 hour.
[0067] User-defined automatic annotation rules consist of two parts: a Boolean expression and a UDF call. An example is shown below:
[0068] RULE_ID = R001
[0069] NAME = "Main pump bearing overheating too quickly"
[0070] EXPR=${BP_TEMP}>80&UDF_GRAD(${BP_TEMP})>3
[0071] TAG_NAME = "bearing_overheat"
[0072] CONFIDENCE = 0.92
[0073] UDF is calculated using a built-in first-order difference function.
[0074] The rules engine orchestrates tasks based on the rule template library, and the resulting DAG nodes are as follows:
[0075] N1: DataReadNode, reads baseline data;
[0076] N2: UDFNode, calls UDF_GRAD;
[0077] N3: LabelWriteNode, used for writing intermediate data;
[0078] N4: AuditNode (optional), for manually reviewed sensitive tags.
[0079] Next, run the task.
[0080] The hash of "Project ID + Measurement Point ID" is split into 32 Workers, with a maximum concurrency of 32 per task. The task can resume from where it fails.
[0081] Read the temperature measurement points of the first loop within the time window [T0, T1], with a data volume of approximately 86,400 points;
[0082] Memory usage is less than 512MiB when using streaming reads.
[0083] Next, the UDF is invoked to calculate the sliding gradient over one minute; the gradient calculation formula is:
[0084] grad[i]=(temp[i+60]-temp[i]) / 60
[0085] Matching is performed according to user-preset rules:
[0086] If grad[i] > 3 and temp[i] > 80, then label the minute interval as bearing_overheat.
[0087] Finally, the data label structure is as follows:
[0088] Table 1 Example of Tag Data
[0089]
[0090] During execution, the sample annotation unit generates a globally unique label version number for each annotation result.
[0091] It should be noted that, in addition to automatic annotation, sample annotation units can also be manually annotated. During manual annotation, users can directly select the target interval or region from the nuclear power equipment time-series curve, 3D model, or detection image in the sample annotation unit's data using either point selection or region selection. Manual labeling is then completed on the pop-up preset label tree. After secondary confirmation by the reviewer, the annotation results are written to the intermediate data area and a globally unique label version number is generated, enabling traceability and version control of manual annotation.
[0092] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.
Claims
1. A data warehouse structure system for nuclear power digital twins, characterized in that, The system includes a data processing layer, a data integration layer, a semantic enhancement layer, and an analysis and application layer; The data processing layer reads the raw data from the nuclear power industrial internet platform and performs data standardization processing on the raw data to form standard data, and constructs a nuclear power industry theme library based on the standard data; the data integration layer performs data encoding mapping, data classification and data fusion on the data in the nuclear power industry theme library to form equipment-level digital twin theme library, system-level digital twin theme library and structure-level digital twin theme library. The semantic enhancement layer annotates the data in the nuclear power industry theme library and the data in the thematic library formed by the data integration layer, associating the data with feature tags, and constructing a structured feature tag library from the data with associated feature tags. The analysis and application layer reads the data from the feature tag library, and constructs a digital twin algorithm library and a digital twin sample library through algorithm analysis and training samples. The semantic enhancement layer associates feature tags of equipment status, mechanism, operating condition, and risk emergency response with the data in the nuclear power industry theme library and the data in the thematic library formed by the data integration layer, and forms a tag knowledge graph based on the feature tag library. The analysis and application layer includes a data analysis module and a data sample production module. The data analysis module provides analysis algorithms to analyze and process the data in the feature tag library and construct a digital twin algorithm. The data analysis module includes a time-series data analysis unit, a spatial-physical correlation analysis unit, a multimodal data fusion analysis unit, and a business scenario analysis unit. The time-series data analysis unit reads real-time sensor data of nuclear power equipment from the feature tag library, and uses time series decomposition, trend prediction, and anomaly detection methods to analyze the changing patterns of equipment parameters over time, predict future changes in equipment parameters, and determine the health trend of the equipment. The spatial-physical correlation analysis unit reads 3D model data of structures and spatial location information of equipment from the feature tag library, performs spatial topology analysis and physical field coupling analysis on the 3D model data of structures and spatial location information of equipment, correlates the spatial location and physical state data of structures and equipment, and analyzes the performance degradation parameters of structures and equipment. The multimodal data fusion analysis unit reads multimodal data from the feature label library and performs cross-modal feature alignment on the multimodal data.
2. The data warehouse structure system for nuclear power digital twins according to claim 1, characterized in that, The data processing layer maps the raw data into standard data through pre-defined mapping rules, and maps the raw data into unified business terms, parameter naming rules, and equipment codes. The data processing layer divides the standard data into business domains such as equipment, operation, safety and quality, maintenance, control, detection, fuel, and design by hierarchically from subject domain grouping, subject domain, business object, and data entity. The standard data is stored using relational database, time-series database, and object database subject modeling.
3. The data warehouse structure system for nuclear power digital twins according to claim 1, characterized in that, The data integration layer will associate the SSC codes with system operation, equipment monitoring and analysis, and structure-level monitoring and management data objects based on the SSC codes and equipment codes extracted from the thematic library, and associate the equipment codes with the component level of the SSC codes; and establish mapping rules between SSC codes and equipment codes to realize cross-library association of the three libraries: equipment-level digital twin thematic library, system-level digital twin thematic library, and structure-level digital twin thematic library.
4. The data warehouse structure system for nuclear power digital twins according to claim 1, characterized in that, The business scenario analysis unit reads real-time data of the nuclear power plant system from the feature tag library, performs fault diagnosis and root cause analysis based on historical fault cases, and uses association rule mining algorithm and causal inference algorithm to locate the system fault propagation path. The business scenario analysis unit uses a multi-objective optimization algorithm combined with physical constraints to perform system operation optimization analysis on the real-time data, analyzes the impact of system operating parameters on energy consumption and efficiency, and outputs parameter adjustment schemes. The business scenario analysis unit reads the device health status, environmental parameters, and historical accident data from the feature tag library, performs risk warning analysis, and uses risk matrix and risk algorithm model to calculate the probability of potential risks to the device.
5. The data warehouse structure system for a nuclear power digital twin according to claim 4, characterized in that, The data sample creation module constructs a digital twin sample library by processing data from the feature label library. This module includes a sample data processing unit, a sample enhancement unit, a sample annotation unit, and a sample partitioning unit. The sample data processing unit generates sample data for the digital twin, performing data preparation and preprocessing operations according to the digital twin business requirements. The sample enhancement unit adds noise, time-shifts, stretches, and compresses normal operating time-series data from the feature label library, and performs correlation and fusion on multimodal data. The sample annotation unit, in conjunction with the feature labels of the semantic enhancement layer, annotates the sample data processed by the data processing unit with business labels. The sample partitioning unit uses stratified sampling to divide the data samples annotated by the sample annotation unit into training, validation, and test sets.
6. The data warehouse structure system for a nuclear power digital twin according to claim 5, characterized in that, The sample data processing unit includes a time-series data sample management subunit and an anomaly data processing unit. The time-series data sample management subunit has data access function, benchmark data selection function, sample data processing function, and sample management function. The data access function reads data of associated feature tags from the feature tag library. The benchmark data selection function selects data from the data access function according to the time range, performs data verification on the selected data, and the verified data is determined by the operator and used as benchmark data. The sample data processing function performs data preparation and preprocessing operations on the benchmark data according to business needs, extracts key features from the benchmark data, and forms sample data. The sample management function retrieves sample data and encapsulates sample datasets according to requirements, forming sample files and sending them to a specified path; the time series data sample management subunit stores data independently for each project, with each project corresponding to multiple sample data sets; the time series data sample management subunit sets up a shared database, and the sample data in the shared database is shared by all projects.
7. A data warehouse structure system for nuclear power digital twins according to claim 6, characterized in that, The abnormal data processing unit uses statistical methods, model methods, or time series downsampling methods to process the sample data, identify and remove abnormal data, write the abnormal data identification results into the sample data database, and perform linear interpolation and normalization processing on the sample data after removing abnormal data. The statistical methods include the normal distribution method and the box plot method, which identify data that exceed a certain number of standard deviations of the normal distribution and data that exceed the fractional points by a preset multiple of the box length as abnormal data. The model methods identify outliers using density clustering algorithms, isolated forest algorithms, and randomized cut forest algorithms. The time series downsampling method downsamples the sample data from the time series and outputs time series data with reduced sampling frequency.
8. The data warehouse structure system for nuclear power digital twins according to claim 7, characterized in that, The data verification of the benchmark data selection function specifically includes data correlation test, multicollinearity test, normality test, time series stationarity test, dead value test, limit value test, and repeatability test; The data preprocessing functions of the sample data processing function include: missing value handling, outlier detection, data augmentation, data imbalance handling, data normalization, and categorical data encoding.
9. A data warehouse structure system for nuclear power digital twins according to claim 8, characterized in that, The sample labeling unit of the data sample creation module includes a label rule management function and an automatic labeling business label function. The label rule management function configures label rule templates, provides rule preview and rule copying, checks rule logic, and binds rule templates to specified label categories. The automatic labeling function allows users to set label rules based on a rule template library. The rule engine is responsible for parsing the label rules, generating a directed acyclic graph of tasks, and performing filtering, calculation, and labeling on the sample data through user-defined functions.
Citation Information
Patent Citations
System for integrating monitoring data of key equipment of nuclear power station
CN113179319A
DCS health management data warehouse system
CN119396931A
