ONLINE MONITORING METHOD OF NUCLEAR POWER PLANT SYSTEM BASED ON ISOLATION FOREST METHOD AND SLIDING WINDOW METHOD

Abstract

The present disclosure relates to an online monitoring method of a nuclear power plant system based on an isolation forest method and a sliding window method. An isolation forest method used in the present disclosure is an abnormal detection model based on the idea of binary tree division, and has no requirements on the dimension and linear characteristics of monitoring data. In view of the characteristics of strong nonlinearity and high dimension of operation data of the nuclear power plant system, in the process of state monitoring, system abnormalities can be detected more quickly and accurately. In the present disclosure, a sliding window method is used to improve an isolation forest model, so that the improved isolation forest model has the functions of model online updating and real-time state monitoring, and the usability of an isolation forest state monitoring method is enhanced.

Claims

1. An online monitoring method of a nuclear power plant system based on an isolation forest method and a sliding window method, comprising the following steps: step 1: acquiring historical operation data of a nuclear power plant in a normal state, and performing standardized preprocessing on the historical operation data of the nuclear power plant in a normal state by a maximum and minimum normalization method, so as to obtain historical operation dimensionless sample data X of the nuclear power plant in a normal state; step 2: acquiring real-time operation data of the nuclear power plant, performing standardized preprocessing on the real-time operation data of the nuclear power plant by the maximum and minimum normalization method, and adding the real-time operation data of the nuclear power plant after the standardized preprocessing to X by a sliding window method, so as to form training data X′, wherein assuming that a length of the sliding window is T, the historical operation dimensionless sample data X of the nuclear power plant in a normal state is expressed as:
X={x.sub.1,x.sub.2, . . . ,x.sub.T−1,x.sub.T}; after acquiring real-time data x.sub.t of the nuclear power plant, deleting the first data of the sliding window, and adding the new data x.sub.t to the end of the sliding window at the same time, so as to form the training data X′:
X′={x.sub.2,x.sub.3, . . . ,x.sub.Tx.sub.t}; step 3: performing state monitoring by an isolation forest method, inputting the training data X′ into an isolation forest model for abnormal detection training, and calculating an abnormal score of the real-time data, so as to realize accurate monitoring of the state of the nuclear power plant system; step 3.1: performing random sampling on the training data X′, constructing an isolated tree model by using the data obtained by random sampling each time, setting a maximum depth of the isolated tree model as l, and integrating all isolated trees into an isolation forest model:
l=ceiling [log.sub.2(φ)], wherein φ represents a number of subsamples, and y=ceiling(x) represents a round-up function, that is, the smallest integer greater than or equal to x is taken; step 3.2: calculating an average path length c(n) of each isolated tree, calculating a path length h(x) of the real-time data x.sub.t of the nuclear power plant in the isolated tree, and calculating an abnormal score s of the real-time data: $c\left(n\right)=\{\begin{array}{cc}2H\left(n-1\right)-\frac{2\left(n-1\right)}{n}& n>1\\ 0& \mathrm{otherwise}\end{array},s\left(x,n\right)=2^{-\frac{E\left[h\left(x\right)\right]}{c\left(n\right)}},$ wherein n represents a sample size contained in a root node of an isolated tree, H(n) represents a harmonic function H (n)=ln(n)+ε, ε=0.5772156649, represents an Euler's constant, h(x) represents a path length of the real-time data x.sub.t of the nuclear power plant in the isolated tree, and E[h(x)] represents an expected value of the path length of the real-time data x.sub.t in all isolated trees of the isolation forest; step 3.3: if the abnormal score is greater than 0.5, determining that the state of the nuclear power plant system is abnormal at the current moment, and issuing an alarm; otherwise, determining that the state of the nuclear power plant system is normal at the current moment, acquiring the real-time operation data of the next moment, and returning to step 2, so as to realize online real-time monitoring of the nuclear power plant system.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0020] FIG. 1 is a flow chart of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The present disclosure will be further described below with reference to the accompanying drawings.

[0022] The present disclosure relates to an online state monitoring method of a nuclear power plant system, and particularly relates to an online monitoring method of a nuclear power plant system based on an isolation forest method and a sliding window method. An objective of the present disclosure is to provide a real-time and accurate online monitoring method of a nuclear power plant system. The method can solve the problems of model online updating and real-time online monitoring of the isolation forest state monitoring method.

[0023] The objective of the present disclosure is realized as follows:

[0024] step 1: historical operation data of a nuclear power plant in a normal state is acquired as training data X;

[0025] step 2: standardized preprocessing is performed on the training data by a maximum and minimum normalization method, so as to obtain dimensionless sample data of an online monitoring model;

[0026] step 3: standardized preprocessing is performed on the real-time operation data of the nuclear power plant according to maximum and minimum values in step 2, and the real-time operation data is added to the training data by a sliding window method, so as to form training data X′;

[0027] step 4: state monitoring is performed by an isolation forest method, the training data X′ is input into an isolation forest model for abnormal detection training, and an abnormal score of the real-time data is calculated, so as to realize accurate monitoring of the state of the nuclear power plant system; and

[0028] step 5: the real-time operation data of the nuclear power plant at the next moment is acquired, steps 3, 4 and 5 are repeated, and online real-time monitoring of the nuclear power plant system is realized by the sliding window method in step 3.

[0029] The sliding window method in step 3 is as follows:

[0030] assuming that a length of the sliding window is T, the training data X composed of normal data can be expressed as:

X={x.sub.1,x.sub.2, . . . ,x.sub.T−1,x.sub.T}; and

[0031] after real-time data x.sub.t of the nuclear power plant is acquired, the first data of the sliding window is deleted, and the new data x.sub.t is added to the end of the sliding window at the same time, so as to form the training data X′:

X′={x.sub.2,x.sub.3, . . . ,x.sub.Tx.sub.t};

[0032] The isolation forest state monitoring method in step 4 is as follows:

[0033] 1) random sampling is performed on the training data, an isolated tree model is constructed by using the data obtained by random sampling each time, a maximum depth of the isolated tree model is set as l, and all isolated trees are integrated into an isolation forest model:

l=ceiling [log.sub.2(φ)],

[0034] where φ represents a number of subsamples, and y=ceiling(x) represents a round-up function, that is, the smallest integer greater than or equal to x is taken;

[0035] 2) an average path length c(n) of each isolated tree is calculated, a path length h(x) of the real-time data x.sub.t of the nuclear power plant in the isolated tree is calculated, and an abnormal score s of the real-time data is calculated:

[00002]$c\left(n\right)=\{\begin{array}{cc}2H\left(n-1\right)-\frac{2\left(n-1\right)}{n}& n>1\\ 0& \mathrm{otherwise}\end{array},s\left(x,n\right)=2^{-\frac{E\left[h\left(x\right)\right]}{c\left(n\right)}},$

[0036] where n represents a sample size contained in a root node of an isolated tree, H(n) represents a harmonic function H(n)=ln(n)+ε, ε=0.5772156649, represents an Euler's constant, h(x) represents a path length of the real-time data x.sub.t of the nuclear power plant in the isolated tree, and E[h(x)] represents an expected value of the path length of the real-time data x.sub.t in all isolated trees of the isolation forest; and

[0037] 3) if the abnormal score is greater than 0.5, it is determined that the state of the nuclear power plant system is abnormal, and if the abnormal score is less than or equal to 0.5, the system is normal.

[0038] The present disclosure has the following beneficial effects:

[0039] The isolation forest method used in step 4 in the technical solution of the present disclosure is an abnormal detection model based on the idea of binary tree division, and has no requirements on the dimension and linear characteristics of monitoring data. In view of the characteristics of strong nonlinearity and high dimension of operation data of the nuclear power plant system, in the process of state monitoring, system abnormalities can be detected more quickly and accurately. The sliding window method is used in step 3 in the technical solution of the present disclosure to improve the isolation forest model, so that the improved isolation forest model has the functions of model online updating and real-time state monitoring, and the usability of the isolation forest state monitoring method is enhanced.

Embodiment 1

[0040] The software of the present disclosure takes PyCharm as a development platform and is compiled by Python3.6 language, and main functions are:

[0041] After a system is connected, historical data of a nuclear power plant during normal operation and real-time online operation data are input and trained to obtain an improved isolation forest online state monitoring model, and then, real-time online monitoring of the nuclear power plant system is performed. Monitoring results are displayed in a main interface for state monitoring in real time in the form of text and curves.

[0042] As shown in FIG. 1, an online monitoring method of a nuclear power plant system based on an improved isolation forest method in the present disclosure includes the following steps:

[0043] (1) PCTRAN simulation software is used to acquire historical data of a nuclear power plant during steady-state normal operation as training data X;

[0044] (2) in order to reduce the influence of noise and dimension, dimensionless normalization is performed on the training data by a maximum and minimum standardized method;

[0045] (3) standardized preprocessing is performed on the real-time operation data of the nuclear power plant according to maximum and minimum values in step (2), and the real-time operation data is added to the training data by a sliding window method, so as to form training data X′;

[0046] the realization process of the sliding window method is as follows: assuming that a length of the sliding window is T, the training data X composed of normal data can be expressed as:

X={x.sub.1,x.sub.2, . . . ,x.sub.T−1,x.sub.T};

[0047] after real-time data x.sub.t of the nuclear power plant is acquired, the first data of the sliding window is deleted, and the new data x.sub.t is added to the end of the sliding window at the same time, so as to form the training data X′:

X′={x.sub.2,x.sub.3, . . . ,x.sub.Tx.sub.t};

[0048] (4) state monitoring is performed by an isolation forest method, the training data X′ is input into the isolation forest model for abnormal detection training, and an abnormal score of real-time data is calculated, so as to obtain an online monitoring result;

[0049] the realization process of the isolation forest state monitoring method is as follows:

[0050] 1) random sampling is performed on the training data, an isolated tree model is constructed by using the data obtained by random sampling each time, a maximum depth of the isolated tree model is set as l, and all isolated trees are integrated into an isolation forest model:

l=ceiling [log.sub.2(φ)],

[0051] where φ represents a number of subsamples, and y=ceiling(x) represents a round-up function, that is, the smallest integer greater than or equal to x is taken;

[0052] 2) an average path length c(n) of each isolated tree is calculated, a path length h(x) of the real-time data x.sub.t of the nuclear power plant in the isolated tree is calculated, and an abnormal score s of the real-time data is calculated:

[00003]$c\left(n\right)=\{\begin{array}{cc}2H\left(n-1\right)-\frac{2\left(n-1\right)}{n}& n>1\\ 0& \mathrm{otherwise}\end{array},s\left(x,n\right)=2^{-\frac{E\left[h\left(x\right)\right]}{c\left(n\right)}},$

[0053] where n represents a sample size contained in a root node of an isolated tree, H(n) represents a harmonic function H(n)=ln(n)+ε, ε=0.5772156649 represents an Euler's constant, h(x) represents a path length of the real-time data x.sub.t of the nuclear power plant in the isolated tree, and E[h(x)] represents an expected value of the path length of the real-time data x.sub.t in all isolated trees in the isolation forest;

[0054] 3) if the abnormal score is greater than 0.5, it is determined that the state of the nuclear power plant system is abnormal, and if the abnormal score is less than or equal to 0.5, the system is normal; and

[0055] (5) the real-time operation data of the nuclear power plant at the next moment is acquired, and steps (3), (4) and (5) are repeated.

[0056] The above description is merely preferred embodiments of the present disclosure and is not intended to limit the present disclosure, and various changes and modifications of the present disclosure may be made by those skilled in the art. Any modifications, equivalent substitutions, improvements, and the like made within the spirit and principle of the present disclosure should be included within the protection scope of the present disclosure.