KR102253230B1 - Predictive diagnosis method and system of nuclear power plant equipment - Google Patents

Abstract

The present invention is to provide a nuclear power plant predictive diagnosis method and system that performs machine learning based on information acquired from nuclear power plant facilities to perform automatic predictive diagnosis on nuclear power plants, based on the data provided from the nuclear power plant. Monitoring the state of the facility and diagnosing a first fault value for the nuclear power plant by performing rule-based machine learning based on the data, and performing statistics-based machine learning based on the data Diagnosing a second fault value for the nuclear power plant, performing a predictive diagnosis for the nuclear power plant based on the first and second fault values, and visualizing the predicted diagnostic value, and information on the nuclear power plant And outputting. Accordingly, the present invention has the effect of increasing the accuracy of the defect value by performing rule-based machine learning and statistics-based machine learning together, and enabling the inspector to intuitively determine the presence or absence of an abnormality by visualizing the information on the combined value. .

Description

Predictive diagnosis method and system of nuclear power plant facilities {PREDICTIVE DIAGNOSIS METHOD AND SYSTEM OF NUCLEAR POWER PLANT EQUIPMENT}

The present invention relates to a predictive diagnosis method and system for nuclear power plants, and more particularly, to a predictive diagnosis method and system for nuclear power plants provided in a nuclear power plant environment.

In general, nuclear power plants are performing preventive maintenance (PM), in which a maintenance plan is established based on operating hours, etc., and regularly performs a fixed maintenance plan. Here, the monitoring status of the facility is by setting the alarm and emergency stop value according to the vibration evaluation standard of the root mean square (RMS) or peak to peak value of the vibration generated from the facility. Vibration trends are managed for safe facility operation until period.

However, in the conventional maintenance method, excessive maintenance may occur due to maintenance performed unconditionally over a certain period of time. In addition, the conventional maintenance method maintains a facility that has been stably operated due to unnecessary maintenance, thereby causing a problem that was not before.

In addition, in the conventional maintenance method, only the vibration trend is detected due to the stable operation of the facility. Accordingly, the conventional maintenance method has a problem in that it is difficult to predict when an abnormality occurs in a facility, and a detailed analysis is required when an abnormality occurs.

Korean Patent Publication No. 1258017 (Pump performance diagnosis device, 2013.04.18)

An object of the present invention is to provide a nuclear power plant predictive diagnosis method and system that performs machine learning based on information acquired from a nuclear power plant and automatically predicts and diagnoses nuclear power plants.

The predictive diagnosis method of a nuclear power plant according to the present invention includes the steps of monitoring the state of the nuclear power plant based on data provided from the nuclear power plant, and performing rule-based machine learning on the nuclear power plant. Diagnosing a first fault value for the nuclear power plant and diagnosing a second fault value for the nuclear power plant by performing statistical machine learning based on the data, and the nuclear power plant facility based on the first and second fault values And performing a predictive diagnosis on and visualizing the predicted diagnosis value and outputting information on the nuclear power plant.

In the outputting step, an area or part that has a high probability of occurrence of an abnormality or requires replacement or maintenance may be marked.

In the step of diagnosing the first fault value, the rule-based machine learning may be performed based on information including at least one of fault information of the nuclear power plant facility, maintenance knowledge, and experience and knowledge of an expert.

Diagnosing the second fault value includes acquiring data from the nuclear power plant, selecting a feature representing the state of the nuclear power plant by applying a feature selection algorithm to the data, and selecting the feature selection algorithm. And diagnosing the second fault value of the nuclear power plant by learning the selected feature through the learning feature and classifying the learned feature through a classification algorithm.

The predictive diagnosis method of the nuclear power plant further includes a step of pre-processing the data before the step of selecting, and the step of pre-processing includes the steps of dividing the data, removing the unstable nature of the data, and reducing the size of the data. It may include the step of reducing.

In the dividing step, the data may be divided at intervals of 500 ms.

In the dividing step, information between adjacent data may be included by overlapping the divided data.

The predictive diagnosis method of the nuclear power plant further includes a step of normalizing the data before the step of selecting, and the step of normalizing includes characteristics having statistical characteristics and geometric characteristics for time domain and data domain data, respectively. It may include the step of extracting, and the step of matching the range of the extracted data and adjusting the distribution to be similar.

Meanwhile, the predictive diagnosis system of a nuclear power plant according to the present invention monitors the state of the nuclear power plant based on data provided from the nuclear power plant, and performs rule-based machine learning based on the data. Diagnosing a first fault value for, and performing statistics-based machine learning based on the data to diagnose a second fault value for the nuclear power plant, and to the nuclear power plant based on the first and second fault values. Predictive diagnosis is performed, the predictive diagnosis value is visualized, and information on the nuclear power plant is output.

On the other hand, the calculation process of the predictive diagnosis system according to the present invention monitors the state of the nuclear power plant based on data provided from the nuclear power plant, and performs rule-based machine learning based on the data. Diagnosing a first fault value for, and performing statistics-based machine learning based on the data to diagnose a second fault value for the nuclear power plant, and to the nuclear power plant based on the first and second fault values. Predictive diagnosis is performed, the predictive diagnosis value is visualized, and information on the nuclear power plant is output.

The predictive diagnosis method and system of a nuclear power plant according to the present invention improves the accuracy of the defect value by performing rule-based machine learning and statistics-based machine learning together, and visualizes the information on the combined value so that the inspector intuitively determines whether or not there is an abnormality It has the effect of being able to perform.

The technical effects of the present invention as described above are not limited to the above-mentioned effects, and other technical effects that are not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a block diagram showing a predictive diagnosis system for a nuclear power plant according to the present embodiment.
2 is a block diagram showing an operation processor of the predictive diagnosis system according to the present embodiment.
3 is a flowchart showing a method for predicting and diagnosing nuclear power plants according to the present embodiment.
4 is a flowchart illustrating a second defect diagnosis step of the predictive diagnosis method of a nuclear power plant according to the present embodiment.
5 is a flowchart showing a pre-processing step of a method for predicting and diagnosing a nuclear power plant according to the present embodiment.
6 is a conceptual diagram showing a visualization of a nuclear power plant predictive diagnosis method according to the present embodiment.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, this embodiment is not limited to the embodiments disclosed below, but may be implemented in various forms, only this embodiment makes the disclosure of the present invention complete, and the scope of the invention is given to those of ordinary skill in the art. It is provided to be fully informed. The shapes of elements in the drawings may be exaggerated for a more clear description, and elements indicated by the same reference numerals in the drawings refer to the same elements.

1 is a block diagram showing a predictive diagnosis system of a nuclear power plant according to the present embodiment, and FIG. 2 is a block diagram showing an operation processor of the predictive diagnosis system according to the present embodiment.

As shown in FIGS. 1 and 2, the predictive diagnosis system 100 of a nuclear power plant according to the present embodiment performs predictive diagnosis on various facilities provided in a nuclear power plant environment. Here, the nuclear power plant 10 may be various rotating devices such as a pump, a compressor, and a fan. However, this is for explaining the present embodiment, and the type of the nuclear power plant 10 is not limited.

On the other hand, the predictive diagnosis system 100 may acquire data from the nuclear power plant 10 to build the data, and perform automatic predictive diagnosis by using the constructed data when the predictive diagnosis of the nuclear power plant 10 is required again. . In addition, the predictive diagnosis system 100 outputs the predicted and diagnosed defect values so that the inspector intuitively identifies whether there is an abnormality in the nuclear power plant 10 and determines the replacement and maintenance timing.

First, the nuclear power plant 10 may be equipped with a sensor for acquiring various information including operation information of the nuclear power plant 10. In addition, the sensor may be interlocked with the predictive diagnosis system 100 to provide the acquired data to the predictive diagnosis system 100. However, this is for explaining the present embodiment, and it should be noted that data on the nuclear power plant 10 may not be acquired by a sensor, but may be directly acquired by an operator and input to the predictive diagnosis system.

In addition, the predictive diagnosis system 100 includes a storage unit 110 in which data provided from the nuclear power plant 10 is stored, an operation processor 120 that performs predictive diagnosis based on data acquired from the nuclear power plant 10, and It may include an output unit 130 for displaying defect information. Here, the operation processor 120 may include an operation device provided in a computer, software for operation, a computer language for operation, and the like.

Meanwhile, hereinafter, a method for predicting and diagnosing a nuclear power plant according to the present embodiment will be described in detail. However, for the above-described components, detailed descriptions are omitted and the same reference numerals are assigned to describe them.

3 is a flow chart showing a nuclear power plant predictive diagnosis method according to the present embodiment, and FIG. 4 is a flow chart showing a second defect diagnosis step of the nuclear power plant predictive diagnosis method according to the present embodiment. And Figure 5 is a flow chart showing the pre-processing step of the nuclear power plant predictive diagnosis method according to the present embodiment, Figure 6 is a conceptual diagram showing a visualization of the nuclear power plant predictive diagnosis method according to the present embodiment.

3 to 5, the predictive diagnosis system 100 according to the present embodiment includes a trend monitoring step (S100), a first defect diagnosis step (S200), a second defect diagnosis step (S300), and a verification step. (S400) and visualization step (S500) may be performed.

First, in the trend monitoring step (S100), the state of the nuclear power plant 10 is monitored in real time. In this case, data of the nuclear power plant 10 may be continuously accumulated in the predictive diagnosis system 100. In addition, data on the nuclear power plant in the past may be stored in advance.

Accordingly, the predictive diagnosis system 100 monitors the state change trend of the nuclear power plant 10 by comparing the accumulated or stored past data with the data provided from the current nuclear power plant 10. At this time, the predictive diagnosis system 100 can monitor the fluctuation trend using information such as amplitude, frequency band, and CR (Consistency Ratio) provided from the nuclear power plant 10, and the characteristics of each part of the nuclear power plant 10 The narrow-band technique for early detection of defects caused by can be applied.

In addition, in the first defect diagnosis step S200, the defect of the nuclear power plant 10 is diagnosed based on the change trend compared in the trend monitoring step S100. In the first defect diagnosis step S200, a rule-based diagnosis method may be used. That is, the predictive diagnosis system 100 may store expert experience and knowledge, defect information and maintenance knowledge of the nuclear power plant 10 in advance.

Accordingly, the predictive diagnosis system 100 may perform spectrum analysis based on a rule-based diagnosis method, thereby primarily diagnosing defects in the nuclear power plant 10. At this time, the predictive diagnosis system 100 may be equipped with an optimal diagnosis algorithm for each nuclear power plant, and the predictive diagnosis system 100 calls a diagnosis algorithm suitable for the nuclear power plant 10 to use a rule-based machine learning method. The defect of the nuclear power plant 10 is primarily diagnosed.

On the other hand, in the second defect diagnosis step (S300), the defect of the nuclear power plant 10 is secondaryly diagnosed based on data provided from the nuclear power plant 10. In the second defect diagnosis step (S300), an algorithm installed in the predictive diagnosis system 100 is called to diagnose a defect in the nuclear power plant 10 using a machine learning method.

That is, in the first defect diagnosis step (S200), defects were diagnosed using a machine learning method in which expert experience and knowledge, defect information and maintenance knowledge of nuclear power plants are combined. However, in the second defect diagnosis step (S300), the defect of the nuclear power plant 10 is diagnosed using a machine learning method based on statistics. Accordingly, the reliability of predictive diagnosis may be improved by comparing the defect values derived respectively in the first defect diagnosis step S200 and the second defect diagnosis step S300 later.

In the second defect diagnosis step S300, data on the fluctuation trend compared in the trend monitoring step S100 is provided. In addition, the predictive diagnosis system 100 performs a data preprocessing step (S310) in consideration of the characteristics of the nuclear power plant 10.

In the preprocessing step S310, a sample overlapping step S311, a filtering step S312, an averaging step S313, and a subband designating step S314 may be sequentially performed.

First, in the sample superimposition step S311, the predictive diagnosis system 100 may perform sample superimposition in consideration of data of the nuclear power plant 10, for example, vibration characteristic data. In this case, the predictive diagnosis system 100 may apply an optimal data sample division method. That is, the conventional data sample length is 1000 ms, that is, 1 second. The signal of such a conventional data sample has one sample in the sample division method.

However, in the case of using the conventional data sample method, components such as fluid inducing characteristics of the nuclear power plant 10 or operation outside of an appropriate efficiency point, and vibrations of other facilities may be included. Accordingly, in the conventional data sample method, status information in the data sample may be lost, and the analysis result may be affected, resulting in misdiagnosis.

Accordingly, in the sample overlapping step S311, the interval per data sample is set to 500 ms, about 0.5 seconds. In this case, the data samples are overlapped so that information between adjacent data samples may be included and divided into multiple pieces. In the case of overlapping data samples, a range in which rotational synchronization and defect components in the data samples are included can be automatically set. In addition, the overlapped data sample may be configured as a mechanism for repeatedly analyzing the state of the nuclear power plant 10.

Thereafter, in the filtering step (S312), the predictive diagnosis system 100 performs filtering to prevent removal of defect information included in the data sample. Here, the predictive diagnosis system 100 performs pass filtering such as low frequency, medium frequency, and high frequency with respect to a frequency band in which noise may be included. At this time, the predictive diagnosis system 100 converts the analog information collected from the nuclear power plant 10 into digital information. In addition, noise components generated due to quantization errors are removed to prevent a convergence phenomenon of vibration defect components generated in the nuclear power plant 10. That is, the predictive diagnosis system 100 in the filtering step S312 may remove and reduce noise inevitable in the digitization process in advance.

Thereafter, in the averaging step (S313), the predictive diagnosis system 100 automatically performs an averaging step on the filtered data to reduce the size of the data. At this time, the predictive diagnosis system 100 removes the unstable nature of the signal that degrades the defect detection capability such as mass imbalance due to improper screw loosening and blade corrosion of the nuclear power plant 10. The size of the data from which the unstable nature has been removed reduces the size, enabling efficient cumulative management of data. In this case, the predictive diagnosis system 100 divides and calculates the data according to the averaging interval and the number of times. That is, a plurality of pieces of data divided according to the degree of averaging are output.

Thereafter, in the subband designation step (S314), the predictive diagnosis system 100 performs subband designation in which the data passed through the averaging step (S313) is divided into a set of specific basis functions. Here, the set of the basis functions to be used is secured through expansion and reduction and parallel movement in the time axis direction for one basic wavelet basis function. In this case, the arbitrary section is provided with nodes that can be selected according to the set level, and is calculated by dividing the shock waves included in the signal waveform.

That is, the predictive diagnosis system 100 decomposes the data into arbitrary sections. Therefore, it is possible to detect a defect component having a high dependence on time, which is difficult to detect with conventional signal processing techniques. In other words, among the defects that may occur in the nuclear power plant 10, minor defects in bearings due to normal wear that are difficult to detect early, fatigue damage of non-metallic parts due to excessive axial thrust, and micro distortion due to thermal shock are prevented early. Can be detected.

This may be a differentiated signal processing technology for improving diagnostic performance of the nuclear power plant 10. Here, the calculated result is divided into a plurality of time intervals, divided into an evaluation that applies the envelope analysis for each section, and divided into several frequency bands, and is divided into an evaluation that applies the envelope analysis for each section and output.

Meanwhile, in the second defect diagnosis step (S300), a feature expression and normalization step (S320) of extracting/normalizing statistical or geometric features from the data is performed after the preprocessing step (S310). Here, the predictive diagnosis system 100 extracts statistical features and geometric features for time domain and data domain data through equations to be described later. Then, normalization is performed to make the distribution similar by matching the range of the extracted feature data.

At this time, the predictive diagnosis system 100 extracts features having respective statistical characteristics or morphological characteristics with respect to the output time-domain and frequency-domain data. Here, the extraction of features having statistical characteristics or morphological characteristics is set as many as a set number through the following equation. In addition, each extracted feature value is provided as data for a machine learning algorithm by deriving a correlation coefficient for each feature vector in the predictive diagnosis system 100.

는 x(n)의 평균을 나타내고,

은 x(n)의 표준편차를 나타낸다. In the following equation, x(n) denotes a time domain signal used for fault feature extraction, and N denotes the number of samples of x(n). And

Represents the average of x(n),

Represents the standard deviation of x(n).

First, the predictive diagnosis system 100 may calculate the square root of the squared average through Equation 1. The square root mean square is a value for the magnitude of a change in a signal, and means the degree or magnitude of the yin and yang of a continuous waveform such as a sine wave.

[Equation 1]

In addition, the predictive diagnosis system 100 allows the asymmetry to be calculated through Equation 2. In statistics, asymmetry is an indicator of the asymmetry of the probability distribution. The asymmetry increases as the bias of the signal (the degree to which the distribution of signal values is concentrated on one side based on the average of the signal) increases.

[Equation 2]

In addition, the predictive diagnosis system 100 allows the impulse factor to be calculated through Equation 3. The impulse factor represents the largest impulse in the signal, that is, the size of the point where the waveform rises.

[Equation 3]

In addition, the predictive diagnosis system 100 allows the kurtosis to be calculated through Equation 4. In statistics, kurtosis is an index indicating the sharpness of the shape of the probability distribution. At this time, the kurtosis increases as the distribution of signal values becomes sharper as the distribution of signal values is concentrated near a specific value.

[Equation 4]

In addition, the predictive diagnosis system 100 allows the kurtosis vector to be calculated through Equation (5). The kurtosis vector is a modified value of the kurtosis and can compensate for the disadvantage that the kurtosis is sensitive to the magnitude of the entire signal.

[Equation 5]

In addition, the predictive diagnosis system 100 allows the average square root to be calculated through Equation 6. The mean square root means the degree or size of the yin and yang of a continuous waveform that is the same as the square mean square root. And the mean square root may be more sensitive to the signal size than the square mean square root.

[Equation 6]

In addition, the predictive diagnosis system 100 allows the peak-to-peak to be calculated through Equation 7. Peak-to-peak is an index representing the overall width of a signal, and is the difference between the smallest value and the largest value in the signal.

[Equation 7]

In addition, the predictive diagnosis system 100 allows the margin vector to be calculated through Equation 8. The margin vector refers to the difference between the minimum and maximum values compared to the average signal size.

[Equation 8]

In addition, the predictive diagnosis system 100 allows the crest factor to be calculated through Equation 9. The crest factor, like the margin, refers to the difference between the minimum and maximum values compared to the average size of the signal. In this case, the crest factor may use the square root average square instead of the average square root.

[Equation 9]

In addition, the predictive diagnosis system 100 allows the shape coefficient to be calculated through Equation 10. The shape factor is an index indicating the ratio of the DC component and the AC component in electronic engineering. The shape factor refers to the ratio of the size of a continuous waveform that goes back and forth between the yin and the yang relative to the average of the signal.

[Equation 10]

In addition, the predictive diagnosis system 100 allows the frequency center to be calculated through Equation 11. The frequency center means the average of the frequency domain.

[Equation 11]

In addition, the predictive diagnosis system 100 allows the RMS of Frequency to be calculated through Equation 12. RMS of Frequency means the square root mean squared value in the frequency domain.

[Equation 12]

In addition, the predictive diagnosis system 100 allows the Root Variance Frequency to be calculated through Equation 13. Root Variance Frequency refers to the variance of frequency values in the frequency domain of a signal.

[Equation 13]

In this way, the predictive diagnosis system 100 normalizes the features calculated through the above equation. In this normalization, the range of the extracted feature data is located. In addition, it makes the distribution similar and can be provided as an automatic feature of artificial intelligence-based machine learning analysis. This normalization aims to match the data range. If the normalization is not performed, the reliability of the diagnosis result is lowered because the characteristic values for each state of the nuclear power plant 10 are underestimated or overestimated in the cluster range.

In general, in the evaluation of data, the state is classified by evaluating the extent of the area of each cluster or the distance between other clusters. However, when using data that has not undergone normalization for analysis, the values of each variable have a significant effect on the calculation of distance measurements because the units of each variable are statistically and morphologically different from ??100 to ??10, 1 to 10, and 1000 to 10000 Goes crazy. And the correlation between variables can be misevaluated. Therefore, normalization is performed to prevent this and effectively learn each state of the nuclear power plant 10.

Thereafter, the predictive diagnosis system 100 forms analysis data based on the data passed through the feature expression and normalization step (S320). At this time, the predictive diagnosis system 100 may enable a three-dimensional analysis by using the operation information of the nuclear power plant 10. That is, the predictive diagnosis system 100 may form weighted analysis data by linking operation information such as flow rate, pressure, and temperature of the nuclear power plant 10 to data that has been subjected to the feature expression and normalization step (S320). .

Thereafter, the predictive diagnosis system 100 performs a feature selection step (S330) for the analysis data. At this time, the predictive diagnosis system 100 applies a feature selection algorithm such as a genetic algorithm and a principal component analysis algorithm to the analysis data.

For example, in the application of a genetic algorithm, a set of n solution codes separated as a group is searched in parallel with initial values. Each object in the population is called a chromosome, and the chromosomes that make up the population are improved through an ongoing iterative process called generation. In this genetic algorithm, at the end of a generation, each chromosome is evaluated using the value to which it is adapted. And when the selection process is completed, the genetic algorithm improves the adaptability of the object by using crossover and mutation.

At this time, if the selection method, the crossing probability, and the mutation probability are set according to the objective function, the above process is repeated as many as the number of generations or until an optimal solution that satisfies a predetermined error range is obtained.

Accordingly, in the present embodiment, the objective function of the genetic algorithm is set as a class, that is, the average of the distance between the classes and the cluster in a set of patterns having a common property in order to utilize the optimized feature learning and classification. Accordingly, in the genetic algorithm, a feature capable of effectively expressing the state of the nuclear power plant 10 can be selected from among the features utilized. In addition, since the selected features can detect minute defects in the state of the nuclear power plant 10 in advance, learning and classification performance to be performed in a lower stage can be maximized.

In addition, in the application of the principal component analysis algorithm, a new feature suitable for the state of the nuclear power plant 10 can be constructed by reducing the dimension of the feature vector. However, in the present embodiment, a principal component analysis algorithm is described as a dimensional reduction method used for dimensional reduction and extraction, but this is for explaining the present embodiment, and various algorithms may be used for dimensional reduction and extraction.

On the other hand, the principal component analysis algorithm is a statistical technique that reduces the dimension of the feature vector to compensate for the disadvantages that the classification performance is degraded and the learning and recognition speed of the classifier is slowed by including noise characteristics when there are many features.

로부터 평균 μ, 분산

, 공분산(covariance) cov를 계산한다.In the application of this principal component analysis algorithm, the principal components are ordered irrespective of the k-th largest variance among all principal components. At this time, the k-th principal component is set in a method of maximizing the projection deviation of the data points so as to be orthogonal to the k-1 th principal component. Therefore, in the principal component analysis algorithm, the original data as shown in Equation 14 below as a feature extraction method

From mean μ, variance

, Calculate the covariance cov.

[Equation 14]

Accordingly, the predictive diagnosis system 100 calculates a covariance matrix S composed of variance and covariance, and calculates an eigenvalue of the covariance matrix. Then, the calculated eigenvalues are arranged in order of size, and the principal components for the eigenvalues are selected. Here, the predictive diagnosis system 100 may extract three main components to closely grasp the state of the nuclear power plant.

On the other hand, when the feature selection for the nuclear power plant 10 is completed from the analysis data, the predictive diagnosis system 100 performs a defect classification step (S340) on the data subjected to the feature selection. Here, the predictive diagnosis system 100 learns the feature vectors selected in feature selection and grasps the state of the nuclear power plant 10 based on this.

Hereinafter, the application of the SVM algorithm and the K-NN algorithm in the predictive diagnosis system 100 will be described.

First, the SVM algorithm can be used not only for precedence but also for classification of nonlinear data. This SVM algorithm basically performs binary classification. However, in practical application, the SVM algorithm corresponding to the classification of each class is trained in order to distinguish several states.

In addition, in the determination of new data, after applying all SVM algorithms, a multi-SVM algorithm that determines the affiliation of the class with the most voting values can be applied.

And the K-NN algorithm is a nonparametric classification method. Accordingly, in the K-NN algorithm, when new data are encountered, a distance measure between all the learned data is learned, and k data specified in order from the nearest one is searched to create a set of candidates (state of nuclear power plant). Then, after finding the label value of which class each element of the candidate set belongs to, the class that occupies the highest frequency among the label values is found, and new data is assigned to the class.

Therefore, it is important to set the number of neighbors to be searched for in the K-NN algorithm. To this end, a k value with a low misclassification rate is set by using the learning data and verification data of the nuclear power plant 10, and a weight according to the distance is imposed to improve the accuracy of classification. However, since distance scale learning of the K-NN algorithm is disadvantageous to learning of high-dimensional data due to the characteristics of the classifier, it is desirable to compensate for the disadvantages of the classifier. Accordingly, it may be desirable for the predictive diagnosis system 100 to apply the K-NN algorithm to data to which the principal component analysis algorithm is applied.

This K-NN algorithm enables consistent classification, and if pre-learned abnormal information is sufficient, it is possible to efficiently and accurately grasp the state of the nuclear power plant 10. Accordingly, the K-NN algorithm may be suitable for predictive diagnosis of the nuclear power plant 10 where possible abnormalities are clear, and since it is less affected by noise, stable classification is possible even if the noise is not completely removed in the averaging step (S313). It can be possible.

Meanwhile, in the verification step S400, the accuracy of the predictive diagnosis is improved based on the fault value output from the first fault diagnosis step S200 and the fault value output from the second fault diagnosis step S300. At this time, the predictive diagnosis system 100 compares the data in the first defect diagnosis step S200 and the second defect diagnosis step S300 or collects the data to perform predictive diagnosis for the nuclear power plant 10. .

Meanwhile, in this embodiment, an embodiment in which the predictive diagnosis system 100 performs predictive diagnosis on the nuclear power plant 10 itself has been described. However, this is for explaining the present embodiment, and the predictive diagnosis system 100 may perform predictive diagnosis not only for the nuclear power plant 10 but also for each component constituting the nuclear power plant 10.

Meanwhile, in the visualization step S500, a diagnostic value based on the defect value is displayed on the output unit 130.

In this case, the output unit 130 may output a 2D or 3D image and text. For example, an image of the nuclear power plant 10 (see FIG. 6A) or an exploded view of the parts constituting the nuclear power plant 10 (see FIG. 6B) are output to the output unit 130 as a 2D or 3D image.

In this case, an area or component having a high probability of occurrence of an abnormality or requiring replacement or maintenance may be displayed or highlighted in red on the output unit 130. In addition, when the inspector selects a corresponding region or part, the predictive diagnosis system 100 may output various abnormal information such as an abnormality degree, a probability of occurrence of an abnormality, and an expected time of occurrence of an abnormality as text. Accordingly, the inspector may determine whether a corresponding area and part is replaced and maintenance is performed by using the information output from the predictive diagnosis system 100, and further, a replacement timing of the corresponding area and part may be set.

As described above, the predictive diagnosis method and system for nuclear power plants according to the present invention improves the accuracy of the defect value by performing rule-based machine learning and statistics-based machine learning together, and visualizes the information on the combined value so that the inspector is intuitive. There is an effect of making it possible to judge the presence or absence of an abnormality.

One embodiment of the present invention described above and illustrated in the drawings should not be construed as limiting the technical idea of the present invention. The protection scope of the present invention is limited only by the matters described in the claims, and those of ordinary skill in the technical field of the present invention can improve and change the technical idea of the present invention in various forms. Therefore, such improvements and changes will fall within the scope of the present invention as long as it is apparent to those of ordinary skill in the art.

10: nuclear power plant equipment
100: predictive diagnosis system of nuclear power plant facilities
110: storage unit
120: computation processor
130: output

Claims (10)

Monitoring the state of the nuclear power plant based on data provided from the nuclear power plant;
Diagnosing a first fault value for the nuclear power plant by performing rule-based machine learning based on the data;
Diagnosing a second fault value for the nuclear power plant by performing statistical machine learning based on the data;
Performing predictive diagnosis on the nuclear power plant based on the first and second fault values; And
Visualizing the predicted diagnosis value and outputting information on the nuclear power plant,
Diagnosing the second fault value
Acquiring data from the nuclear power plant facility, selecting a feature representing the state of the nuclear power plant by applying a feature selection algorithm to the data, learning the feature selected through the feature selection algorithm, and learning Classifying features through a classification algorithm to diagnose the second fault value of the nuclear power plant,
Prior to the step of selecting, the step of pre-processing the data, the step of pre-processing further comprises the step of dividing the data, removing the unstable nature of the data and reducing the size of the data The predictive diagnosis method of the facility.
The method of claim 1,
In the outputting step
A predictive diagnosis method of nuclear power plant facilities that mark areas or parts that are likely to cause abnormalities or require replacement or maintenance.
The method of claim 1,
In the step of diagnosing the first fault value
A predictive diagnosis method of a nuclear power plant facility for performing the rule-based machine learning based on information including at least one of fault information of the nuclear power plant facility, maintenance knowledge, and experience and knowledge of an expert.
delete delete The method of claim 1,
In the dividing step
A predictive diagnosis method for a nuclear power plant that divides the data into 500ms intervals.
The method of claim 1,
In the dividing step
A predictive diagnosis method of a nuclear power plant facility in which information between adjacent data is included by overlapping the divided data.
The method of claim 1,
Prior to the screening step,
Further comprising the step of normalizing the data,
The normalizing step
Extracting features having statistical features and geometric features, respectively, for time domain and data domain data,
And matching the range of the extracted data and adjusting the distribution to be similar.
Monitoring the state of the nuclear power plant based on the data provided from the nuclear power plant,
Diagnosing a first fault value for the nuclear power plant by performing rule-based machine learning based on the data,
Diagnosing a second fault value for the nuclear power plant facility by performing statistics-based machine learning based on the data,
Perform predictive diagnosis on the nuclear power plant based on the first and second fault values,
Visualize the predicted diagnosis value and output information on the nuclear power plant,
Diagnosing the second fault value comprises:
Acquiring data from the nuclear power plant facility, selecting a feature representing the state of the nuclear power plant by applying a feature selection algorithm to the data, learning the feature selected through the feature selection algorithm, and learning Classifying features through a classification algorithm to diagnose the second fault value of the nuclear power plant,
Before the step of selecting, the step of pre-processing the data further comprises the step of pre-processing, the step of dividing the data, removing the unstable nature of the data and reducing the size of the Facility predictive diagnosis system.
Monitoring the state of the nuclear power plant based on the data provided from the nuclear power plant,
Diagnosing a first fault value for the nuclear power plant by performing rule-based machine learning based on the data,
Diagnosing a second fault value for the nuclear power plant facility by performing statistical machine learning based on the data,
Perform predictive diagnosis on the nuclear power plant facility based on the first and second fault values,
Visualize the predicted diagnosis value and output information on the nuclear power plant,
Diagnosing the second fault value
Acquiring data from the nuclear power plant facility, selecting a feature representing the state of the nuclear power plant by applying a feature selection algorithm to the data, learning the feature selected through the feature selection algorithm, and learning Classifying features through a classification algorithm to diagnose the second fault value of the nuclear power plant,
Before the step of selecting, the step of pre-processing the data, the step of pre-processing further comprises the step of dividing the data, removing the unstable nature of the data and reducing the size of the data prediction. The computational processor of the diagnostic system.

Images