Method for automatically classifying and generating plant condition monitoring model of nuclear power plant

Abstract

A method for automatically classifying and generating a plant condition monitoring model of a nuclear power plant comprises the steps of: collecting first monitoring signals received by a first facility model and second monitoring signals received by a second facility model; classifying, among the first monitoring signals, certain signals having strong signal correlation into a first independent component analysis condition monitoring model, and classifying other signals having weak signal correlation into a first approximate nearest neighbor condition monitoring model; and classifying, among the second monitoring signals, certain signals into a second independent component analysis condition monitoring model and classifying other signals into a second approximate nearest neighbor condition monitoring model.

Description

Method for automatically classifying and generating a nuclear power plant condition monitoring model

[0001] This paper relates to a method for automatically classifying and generating a nuclear power plant condition monitoring model.

[0002] In general, the plant condition monitoring model is a model that monitors the condition of a nuclear power plant by using signals received from sensors installed in the facilities included in the nuclear power plant.

[0003] Plant condition monitoring models are generated using signals received from various sensors.

[0004] The conventional method of creating a plant condition monitoring model involves combining signals received from the sensors of the nuclear power plant based on the know-how and experience of engineers based on the physical classification system of the nuclear power plant and optimizing the model through repeated experiments, which increases the model creation time and cost.

[0005] In addition, the conventional method of creating a plant condition monitoring model constructs a condition monitoring model through repeated experiments only for monitoring signals with strong signal correlation among the monitoring signals received from the nuclear power plant to the plant condition monitoring model, which makes it difficult to respond to the various operating environments of the nuclear power plant. At the same time, when a new monitoring signal is added from the nuclear power plant, the prediction accuracy of the entire plant condition monitoring model decreases, which causes false alarms.

[0006] One embodiment provides a method for automatically classifying and generating a nuclear power plant condition monitoring model, which automatically classifies and generates a plant condition monitoring model with improved prediction accuracy while reducing model generation time and generation cost, even when the plant condition monitoring model includes low-relevance monitoring signals, such as new monitoring signals, among the monitoring signals received from the nuclear power plant while responding to various operating environments of the nuclear power plant.

[0007] One aspect provides a method for automatically classifying and generating a plant condition monitoring model, the method comprising the steps of collecting a plurality of first monitoring signals received in a first equipment model of a plant condition monitoring model of a nuclear power plant and a plurality of second monitoring signals received in a second equipment model of the plant condition monitoring model, classifying signals having strong signal correlations among the plurality of first monitoring signals into a first independent component analysis (ICA) condition monitoring model and classifying other signals having weak signal correlations among the plurality of first monitoring signals into a first approximate nearest neighbor condition monitoring model, and classifying signals having strong signal correlations among the plurality of second monitoring signals into a second ICA condition monitoring model and classifying other signals having weak signal correlations among the plurality of second monitoring signals into a second approximate nearest neighbor condition monitoring model.

[0008] The first independent component analysis state monitoring model and the second independent component analysis state monitoring model can each perform independent component analysis based on dimension reduction on the first monitoring signals and the second monitoring signals, respectively, to form a state monitoring model based on principal components based on the physical classification system of the nuclear power plant.

[0009] Each of the first independent component analysis state monitoring model and the second independent component analysis state monitoring model can perform signal restoration prediction based on each of the first equipment model and the second equipment model.

[0010] Each of the first approximate neighbor condition monitoring model and the second approximate neighbor condition monitoring model can configure adjacent signals as condition monitoring models based on the physical classification system of the nuclear power plant for each of the other signals among the first monitoring signals and each of the other signals among the second monitoring signals.

[0011] Each of the first approximate neighbor condition monitoring model and the second approximate neighbor condition monitoring model can perform approximate neighbor search prediction based on the physical classification system of the nuclear power plant.

[0012] The method may further include a step of classifying other signals having a weak signal correlation among the plurality of first monitoring signals and other signals having a weak signal correlation among the plurality of second monitoring signals into a third approximate nearest neighbor state monitoring model.

[0013] According to one embodiment, a method for automatically classifying and generating a nuclear power plant condition monitoring model is provided, which automatically classifies and generates a plant condition monitoring model that responds to various operating environments of a nuclear power plant and improves prediction accuracy while reducing model creation time and creation cost, even when low-relevance monitoring signals, such as new monitoring signals, are included among the monitoring signals received from the nuclear power plant in the plant condition monitoring model.

[0014] FIG. 1 is a flowchart illustrating a method for automatically classifying and generating a plant status monitoring model according to one embodiment.

[0015] FIG. 2 is a diagram for explaining a method for automatically classifying and generating a plant status monitoring model according to one embodiment.

[0016] FIG. 3 is a graph showing an example of independent component analysis used in a method for automatically classifying and generating a plant condition monitoring model according to one embodiment.

[0017] FIG. 4 is a diagram illustrating an example of approximate nearest neighbor search used in a method for automatically classifying and generating a plant condition monitoring model according to one embodiment.

[0018] FIG. 5 is a flowchart illustrating an example of a method for automatically classifying and generating a plant status monitoring model according to one embodiment.

[0019] Hereinafter, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily implement the present invention. The present invention may be implemented in various different forms and is not limited to the embodiments described herein.

[0020] Additionally, throughout the specification, whenever a part is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated.

[0021] Hereinafter, a method for automatically classifying and generating a plant status monitoring model according to one embodiment will be described with reference to FIGS. 1 to 5.

[0022] A method for automatically classifying and generating a plant condition monitoring model according to one embodiment can automatically classify and generate a plant condition monitoring model for monitoring the condition of a nuclear power plant, but is not limited thereto.

[0023] The method for automatically classifying and generating a plant condition monitoring model according to one embodiment may be performed by artificial intelligence driven by a computing device, but is not limited thereto.

[0024] Fig. 1 is a flowchart illustrating a method for automatically classifying and generating a plant condition monitoring model according to one embodiment. Fig. 2 is a diagram for explaining a method for automatically classifying and generating a plant condition monitoring model according to one embodiment.

[0025] Referring to FIGS. 1 and 2, first, a plurality of first monitoring signals (MS1) received by a first equipment model (100) of a nuclear power plant status monitoring model (1000) and a plurality of second monitoring signals (MS2) received by a second equipment model (200) of the plant status monitoring model (1000) are collected (S100).

[0026] For example, a plurality of first monitoring signals (MS1) are received by a first equipment model (100) of a plant condition monitoring model (1000) of a nuclear power plant. The plurality of first monitoring signals (MS1) may include existing monitoring signals received by the first equipment model (100). A plurality of second monitoring signals (MS2) are received by a second equipment model (200) of a plant condition monitoring model (1000) of a nuclear power plant. The plurality of second monitoring signals (MS2) may include existing monitoring signals received by the second equipment model (200). A plurality of first monitoring signals (MS1) and a plurality of second monitoring signals (MS2) are collected.

[0027] Next, among the plurality of first monitoring signals (MS1), signals with strong signal correlation are classified into the first independent component analysis state monitoring model (300), and among the plurality of first monitoring signals (MS1), other signals with weak signal correlation are classified into the first approximate nearest neighbor state monitoring model (400) (S200).

[0028] For example, independent component analysis (ICA) of multiple first surveillance signals (MS1) is performed. The ICA may include various known ICA methods. For example, various known ICA methods may be performed by artificial intelligence driven by a computing device. The step of performing ICA of multiple first surveillance signals (MS1) may be performed by applying ICA based on dimensionality reduction.

[0029] FIG. 3 is a graph showing an example of independent component analysis used in a method for automatically classifying and generating a plant condition monitoring model according to one embodiment.

[0030] Referring to FIG. 3, for example, the artificial intelligence of the first independent component analysis (ICA) condition monitoring model (300) can automatically classify signal components by applying independent component analysis (ICA) based on dimensionality reduction. The artificial intelligence of the first independent component analysis (ICA) condition monitoring model (300) groups only signals with strong signal correlation among the plurality of first monitoring signals (MS1) and performs independent component analysis based on dimensionality reduction on each of the signals to configure the principal components as a condition monitoring model based on the physical classification system of the nuclear power plant and performs signal restoration prediction based on the first facility model (100), thereby performing precise prediction. The independent component analysis based on dimensionality reduction of the first independent component analysis (ICA) condition monitoring model (300) can provide superior prediction results compared to existing statistical techniques. The artificial intelligence can automatically separate other signals with weak signal correlation among the plurality of first monitoring signals (MS1) into the first approximate nearest neighbor condition monitoring model (400) and perform separate prediction. The independent component analysis based on dimensionality reduction can be more efficient and secure objectivity compared to the conventional manual separation by engineers.

[0031] When signal correlation is strong, independent component analysis (ICA) can be used to classify the signal data into a signal model based on the plant's physical system structure. This highly correlated signal structure provides higher prediction accuracy compared to conventional statistical-based predictions. Based on the prediction results, residual validation can be used to generate a final monitoring signal model.

[0032] The first approximate nearest neighbor condition monitoring model (400) can perform an Approximate Nearest Neighbor Search (ANNS) prediction based on the physical classification system of the nuclear power plant. The first approximate nearest neighbor condition monitoring model (400) can configure adjacent signals as condition monitoring models based on the physical classification system of the nuclear power plant for other signals with weak signal correlation among the first monitoring signals (MS1). For example, the first approximate nearest neighbor condition monitoring model (400) can automatically reorganize adjacent signals, such as vibration and pressure signals of the same pump or temperature and pressure signals of the same system line, based on the physical classification system of the nuclear power plant. The first approximate nearest neighbor condition monitoring model (400) can perform an Approximate Nearest Neighbor Search prediction that quickly searches for the most similar value from the most recent past information based on the physical classification system of the nuclear power plant rather than a conventional correlation-based statistical technique.

[0033] FIG. 4 is a graph showing an example of approximate nearest neighbor search used in a method for automatically classifying and generating a plant condition monitoring model according to one embodiment.

[0034] Referring to FIG. 4, for example, the artificial intelligence of the first approximate nearest neighbor condition monitoring model (400) can perform prediction using approximate nearest neighbor search. The artificial intelligence of the first approximate nearest neighbor condition monitoring model (400) can perform approximate nearest neighbor search by vectorizing past signal information, dividing a hyperplane to index the data distribution area into a tree structure, vectorizing signal measurement information, performing an index search, and deriving adjacent values. Since nuclear power plants repeat operation and planned preventive maintenance (OH) at regular intervals, the artificial intelligence of the first approximate nearest neighbor condition monitoring model (400) can easily utilize the latest data of the previous cycle. The first approximate nearest neighbor condition monitoring model (400) can increase the correlation from a physical model perspective and have excellent prediction accuracy and computation speed even with low correlation. The Approximate Nearest Neighbor Search (ANNS) prediction of the first approximate nearest neighbor state monitoring model (400) can have higher prediction accuracy than the conventional Auto Associative Kernel Regression (AAKR) statistical prediction.

[0035] Approximate nearest neighbor search (ANN) can include various known ANN methods. For example, various ANN predictions can be performed by AI powered by a computing device. AI can perform sophisticated predictions for highly correlated signals through signal reconstruction using independent component analysis (ICA) of multivariate signals, compared to statistical techniques, and automatically separate out signals with low correlation. The separated signals can then be combined into signal clusters from adjacent equipment signals based on the plant's physical system structure. Using ANN methods, AI can then find the most similar values ​​from past data to validate the prediction quality.

[0036] When signal correlation is weak, an approximate nearest neighbor search (ANS) condition monitoring system can be used to classify adjacent signals and construct a signal model based on the plant's physical system structure. In statistical prediction, when correlation is weak, residuals increase significantly, making it difficult to detect anomalies. However, ANS prediction learns from recent past data and derives the most similar values ​​as predicted values, enabling precise steady-state predictions and enabling the detection of anomalies through residual detection.

[0037] Next, referring to FIGS. 1 and 2, among the plurality of second monitoring signals (MS2), signals having strong signal correlation are classified into a second independent component analysis state monitoring model (500), and among the plurality of second monitoring signals (MS2), other signals having weak signal correlation are classified into a second approximate nearest neighbor state monitoring model (600) (S300).

[0038] For example, independent component analysis (ICA) of multiple second surveillance signals (MS2) is performed. The ICA may include various known ICA methods. For example, various known ICA methods may be performed by artificial intelligence driven by a computing device. The step of performing ICA of multiple second surveillance signals (MS2) may be performed by applying ICA based on dimensionality reduction.

[0039] For example, the artificial intelligence of the second independent component analysis (ICA) condition monitoring model (500) can automatically classify signal components by applying independent component analysis (ICA) based on dimensionality reduction. The artificial intelligence of the second independent component analysis (ICA) condition monitoring model (500) groups only signals with strong signal correlation among the plurality of second monitoring signals (MS2) and performs independent component analysis based on dimensionality reduction on each of the signals to configure the principal components as a condition monitoring model based on the physical classification system of the nuclear power plant and performs signal restoration prediction based on the second facility model (200), thereby performing precise prediction. The independent component analysis based on dimensionality reduction of the second independent component analysis (ICA) condition monitoring model (500) can provide superior prediction results compared to existing statistical techniques. The artificial intelligence can automatically separate other signals with weak signal correlation among the plurality of second monitoring signals (MS2) into the second approximate nearest neighbor condition monitoring model (600) and perform separate prediction. The independent component analysis based on dimensionality reduction can be more efficient and secure objectivity compared to the conventional manual separation by engineers.

[0040] When signal correlation is strong, independent component analysis (ICA) can be used to classify the signal data into a signal model based on the plant's physical system structure. This highly correlated signal structure provides higher prediction accuracy compared to conventional statistical-based predictions. Based on the prediction results, residual validation can be used to generate a final monitoring signal model.

[0041] The second approximate nearest neighbor condition monitoring model (600) can perform an Approximate Nearest Neighbor Search (ANNS) prediction based on the physical classification system of the nuclear power plant. The second approximate nearest neighbor condition monitoring model (600) can configure adjacent signals as condition monitoring models based on the physical classification system of the nuclear power plant for other signals among the second monitoring signals (MS2) with weak signal correlation. For example, the second approximate nearest neighbor condition monitoring model (600) can automatically reorganize adjacent signals such as vibration and pressure signals of the same pump or temperature and pressure signals of the same system line based on the physical classification system of the nuclear power plant. The second approximate nearest neighbor condition monitoring model (600) can perform an Approximate Nearest Neighbor Search prediction that quickly searches for the most similar value from the latest past information based on the physical classification system of the nuclear power plant rather than a conventional correlation-based statistical technique.

[0042] For example, the artificial intelligence of the second approximate nearest neighbor condition monitoring model (600) can perform prediction using approximate nearest neighbor search. The artificial intelligence of the second approximate nearest neighbor condition monitoring model (600) can perform approximate nearest neighbor search by vectorizing past signal information, dividing the hyperplane to index the data distribution area into a tree structure, vectorizing signal measurement information, performing index search, and deriving adjacent values. Since nuclear power plants repeat operation and planned preventive maintenance (OH) at regular intervals, the artificial intelligence of the second approximate nearest neighbor condition monitoring model (600) can easily utilize the latest data of the previous cycle. The second approximate nearest neighbor condition monitoring model (600) can increase the correlation from the physical model perspective and have excellent prediction accuracy and computation speed even at low correlation. The Approximate Nearest Neighbor Search (ANNS) prediction of the second approximate nearest neighbor state monitoring model (600) can have higher prediction accuracy than the conventional Auto Associative Kernel Regression (AAKR) statistical prediction.

[0043] Approximate nearest neighbor search (ANN) can include various known ANN methods. For example, various ANN predictions can be performed by AI powered by a computing device. AI can perform sophisticated predictions for highly correlated signals through signal reconstruction using independent component analysis (ICA) of multivariate signals, compared to statistical techniques, and automatically separate out signals with low correlation. The separated signals can then be combined into signal clusters from adjacent equipment signals based on the plant's physical system structure. Using ANN methods, AI can then find the most similar values ​​from past data to validate the prediction quality.

[0044] When signal correlation is weak, an approximate nearest neighbor search (ANS) condition monitoring system can be used to classify adjacent signals and construct a signal model based on the plant's physical system structure. In statistical prediction, when correlation is weak, residuals increase significantly, making it difficult to detect anomalies. However, ANS prediction learns from recent past data and derives the most similar values ​​as predicted values, enabling precise steady-state predictions and enabling the detection of anomalies through residual detection.

[0045] Next, among the plurality of first monitoring signals (MS1), other signals having a weak signal correlation and among the plurality of second monitoring signals (MS2), other signals having a weak signal correlation are classified into a third approximate nearest neighbor state monitoring model (700) (S400).

[0046] For example, the artificial intelligence can automatically separate other signals with weak signal correlation among the plurality of first monitoring signals (MS1) and other signals with weak signal correlation among the plurality of second monitoring signals (MS2) into a third approximate nearest neighbor state monitoring model (700) and perform separate predictions.

[0047] The third approximate nearest neighbor condition monitoring model (700) can perform an Approximate Nearest Neighbor Search (ANNS) prediction based on the physical classification system of the nuclear power plant. The third approximate nearest neighbor condition monitoring model (700) can configure adjacent signals as condition monitoring models based on the physical classification system of the nuclear power plant for other signals with weak signal correlation among the first monitoring signals (MS1) and other signals with weak signal correlation among the second monitoring signals (MS2). For example, the third approximate nearest neighbor condition monitoring model (700) can automatically reorganize adjacent signals, such as vibration and pressure signals of the same pump or temperature and pressure signals of the same system line, based on the physical classification system of the nuclear power plant. The third approximate nearest neighbor condition monitoring model (700) can perform an Approximate Nearest Neighbor Search prediction that quickly searches for the most similar value from the latest past information based on the physical classification system of the nuclear power plant, rather than a conventional correlation-based statistical technique.

[0048] For example, the artificial intelligence of the third approximate nearest neighbor condition monitoring model (700) can perform prediction using approximate nearest neighbor search. The artificial intelligence of the third approximate nearest neighbor condition monitoring model (700) can perform approximate nearest neighbor search by vectorizing past signal information, dividing the hyperplane to index the data distribution area into a tree structure, vectorizing signal measurement information, performing index search, and deriving adjacent values. Since nuclear power plants repeat operation and planned preventive maintenance (OH) at regular intervals, the artificial intelligence of the third approximate nearest neighbor condition monitoring model (700) can easily utilize the latest data of the previous cycle. The third approximate nearest neighbor condition monitoring model (700) can increase the correlation from the physical model perspective and have excellent prediction accuracy and computation speed even at low correlation. The Approximate Nearest Neighbor Search (ANNS) prediction of the third approximate nearest neighbor state monitoring model (700) can have higher prediction accuracy than the conventional Auto Associative Kernel Regression (AAKR) statistical prediction.

[0049] Approximate nearest neighbor search (ANN) can include various known ANN methods. For example, various ANN predictions can be performed by AI powered by a computing device. AI can perform sophisticated predictions for highly correlated signals through signal reconstruction using independent component analysis (ICA) of multivariate signals, compared to statistical techniques, and automatically separate out signals with low correlation. The separated signals can then be combined into signal clusters from adjacent equipment signals based on the plant's physical system structure. Using ANN methods, AI can then find the most similar values ​​from past data to validate the prediction quality.

[0050] When signal correlation is weak, an approximate nearest neighbor search (ANS) condition monitoring system can be used to classify adjacent signals and construct a signal model based on the plant's physical system structure. In statistical prediction, when correlation is weak, residuals increase significantly, making it difficult to detect anomalies. However, ANS prediction learns from recent past data and derives the most similar values ​​as predicted values, enabling precise steady-state predictions and enabling the detection of anomalies through residual detection.

[0051] FIG. 5 is a flowchart illustrating an example of a method for automatically classifying and generating a plant status monitoring model according to one embodiment.

[0052] Referring to FIG. 5, an example of a method for automatically classifying and generating a plant condition monitoring model according to one embodiment is illustrated. When signal correlation is strong, the system can be classified as an independent component analysis (ICA) condition monitoring system, and the principal components can be constructed into a signal model based on the plant physical system structure. Because the signal is composed of highly correlated signals, the prediction accuracy is higher than that of conventional statistical-based predictions when predicting signal restoration. Based on the prediction results, residual validation can be performed to ultimately generate a monitoring signal model.

[0053] When signal correlation is weak, an approximate nearest neighbor search (ANS) condition monitoring system can be used to classify adjacent signals and construct a signal model based on the plant's physical system structure. In statistical prediction, when correlation is weak, residuals increase significantly, making it difficult to detect anomalies. However, ANS prediction learns from recent past data and derives the most similar values ​​as predicted values, enabling precise steady-state predictions and enabling the detection of anomalies through residual detection.

[0054] For example, a method for automatically classifying and generating a nuclear power plant condition monitoring model is provided, which automatically classifies and generates a plant condition monitoring model that responds to various operating environments of a nuclear power plant and improves prediction accuracy while reducing model creation time and creation cost, even when new monitoring signals with low correlation are included among the monitoring signals received from the nuclear power plant in the plant condition monitoring model.

[0055] Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by those skilled in the art using the basic concept of the present invention defined in the following claims also fall within the scope of the present invention.

[0056] [Explanation of symbols]

[0057] Plant condition monitoring model (1000), first equipment model (100), second equipment model (200), first independent component analysis condition monitoring model (300), first approximate nearest neighbor condition monitoring model (400), second independent component analysis condition monitoring model (500), second approximate nearest neighbor condition monitoring model (600), third approximate nearest neighbor condition monitoring model (700)

Claims

1. A step of collecting a plurality of first monitoring signals received in a first equipment model of a plant status monitoring model of a nuclear power plant and a plurality of second monitoring signals received in a second equipment model of the plant status monitoring model; A step of classifying signals having strong signal correlation among the plurality of first monitoring signals into a first independent component analysis state monitoring model, and classifying other signals having weak signal correlation among the plurality of first monitoring signals into a first approximate nearest neighbor state monitoring model; and A step of classifying signals having strong signal correlation among the plurality of second monitoring signals into a second independent component analysis state monitoring model, and classifying other signals having weak signal correlation among the plurality of second monitoring signals into a second approximate nearest neighbor state monitoring model. A method for automatically classifying and generating a plant condition monitoring model including:

2. In paragraph 1, A method for automatically classifying and generating a plant condition monitoring model, wherein each of the first independent component analysis condition monitoring model and the second independent component analysis condition monitoring model performs independent component analysis based on dimension reduction for each of the first monitoring signals and each of the second monitoring signals to construct a condition monitoring model based on principal components based on the physical classification system of the nuclear power plant.

3. In paragraph 2, A method for automatically classifying and generating plant condition monitoring models, wherein each of the first independent component analysis condition monitoring model and the second independent component analysis condition monitoring model performs signal restoration prediction based on each of the first equipment model and the second equipment model.

4. In paragraph 1, A method for automatically classifying and generating a plant condition monitoring model, wherein each of the first approximate nearest neighbor condition monitoring model and the second approximate nearest neighbor condition monitoring model constructs a condition monitoring model for adjacent signals based on the physical classification system of the nuclear power plant for each of the other signals among the first monitoring signals and each of the other signals among the second monitoring signals.

5. In paragraph 4, A method for automatically classifying and generating plant condition monitoring models, wherein each of the first approximate nearest neighbor condition monitoring model and the second approximate nearest neighbor condition monitoring model performs approximate nearest neighbor search prediction based on the physical classification system of the nuclear power plant.

6. In paragraph 1, A method for automatically classifying and generating a plant condition monitoring model, further comprising a step of classifying other signals having a weak signal correlation among the plurality of first monitoring signals and other signals having a weak signal correlation among the plurality of second monitoring signals into a third approximate nearest neighbor condition monitoring model.

Images