JP2000259223A - Plant monitoring device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a plant monitoring device which monitors the abnormality according to a characteristic function that is necessary for the desirable estimation and then properly displays the information for the plant monitoring. SOLUTION: This monitoring device includes a plant data input device 1 which inputs the plant data received from a plant, a data preserving device 2 which preserves the plant data, a standard production device 3 which produces a monitoring standard on the basis of the plant data on a normal mode, a monitoring processor 4 which reads the plant data and decides a normal or abnormal state according to the given monitoring standard and a display device 5 which displays the deciding result of the processor 4 in addition to the plant data and the monitoring standard. Then the processor 4 compares the plant data to be monitored with the monitoring standard produced by the device 3. The abnormality is decided when the plant data exceed the fixed threshold value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plant monitoring system for monitoring an operating state of a large-scale plant, such as a nuclear power plant, a thermal power plant, or a chemical plant, based on a process signal, and detecting and diagnosing the occurrence of an abnormality at an early stage. Related to the device.

[0002]

2. Description of the Related Art In a large-scale plant such as a nuclear power plant, a large number of process signals are measured for the purpose of monitoring the performance of the plant and the health of various systems and devices constituting the plant. Because it is difficult for plant operators to constantly monitor all of the large numbers of signals,
Many plants are provided with a monitoring system that captures process signals using a computer and detects abnormal changes in the plant.

[0003] In these monitoring systems, a threshold determination is often used in which a signal to be monitored exceeds a certain threshold value and is regarded as abnormal. Further, when the signal to be monitored is y, this signal is obtained from a known characteristic function f from y itself or other process quantities x1, x2,.
A prediction error when the prediction is performed using (x1, x2,...)
That is, by using δy = y−f (x1, x2,...) As monitoring indices and comparing them with the abnormality detection threshold,
A method of detecting an abnormality is also used.

On the other hand, it is necessary to display the results of these determinations for notifying the operator. However, due to the limitation of displaying them as a two-dimensional graph on the display device, the monitoring target signal y and its predicted value f (x1, x2 , ...) and the prediction error δy
Are plotted and displayed as a function of time, or one of the signals xi used for input is plotted on the horizontal axis and the monitored signal y is plotted on the vertical axis.

[0005]

The following problems have been pointed out in the above-mentioned prior art plant monitoring apparatus.
That is, when a physical model such as a mass energy balance model is used as a characteristic function used for predicting a monitoring target signal, prediction accuracy is insufficient, and a desired result cannot be obtained. In addition, a characteristic function obtained by fitting a linear or non-linear regression model from normal data may be used. However, when a signal used as an input is inappropriate, the prediction accuracy is insufficient, which is not preferable. In particular, if the process volume varies in a similar manner throughout the plant, it becomes difficult to select which signal to use for the input.

[0006] Further, the prediction function of the characteristic function used for prediction may be deteriorated when the fluctuation range of the input signal is expanded to an unexpected range. If the abnormality determination is performed ignoring the fluctuation range of the input signal, occurrence of an erroneous determination cannot be avoided.

On the other hand, since the relationship between the signal used for input and the signal to be monitored is conventionally viewed in the range of a two-dimensional plot, if the signal to be monitored changes depending on a plurality of input signals, it is simply Unpredictable behavior can occur in a plot with one input signal.

Further, in a plant, disturbance occurs even in a normal state due to instrument calibration and surveillance, so that the disturbance is regarded as abnormal, and an erroneous determination occurs.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a plant monitoring apparatus capable of performing abnormality monitoring based on a characteristic function for preferable prediction and displaying information for monitoring appropriately.

[0010]

According to a first aspect of the present invention, there is provided a plant data input device for inputting plant data from a plant, a data storage device for storing plant data provided from the plant data input device, and a data storage device for storing the plant data. A reference creation device that creates a monitoring standard based on normal plant data stored in the storage device, and reads the plant data saved in the data storage device and determines normal / abnormal according to the monitoring reference given by the reference creation device And a display device for displaying the plant data and the monitoring criteria given to the monitoring processing device together with the determination result.

In the above plant monitoring device, the reference creation device creates a monitoring standard from the plant data in a normal state, and compares the monitoring target plant data provided from the data storage device with the obtained monitoring standard in the monitoring processing device. I do. If the comparison result exceeds, for example, a threshold value, it is determined to be abnormal, and the determination result is displayed on the display device together with the monitoring target plant data and the monitoring standard. In this manner, various abnormalities occurring in the system, equipment, and the like in the plant can be found at an early stage.

For this monitoring criterion, for example, a linear or non-linear regression model created from normal plant data is used, and its prediction error is compared with a threshold value. With this method, it is possible to improve the accuracy of the abnormality monitoring, and it is possible to perform monitoring with less erroneous determination.

According to a second aspect of the present invention, the monitoring / processing apparatus determines whether the plant data to be determined is normal or abnormal by inputting a plurality of other plant data and generating the linear or non-linear regression from the normal plant data. This is performed using the difference between the predicted value predicted from the model and the plant data to be determined.

In the above-described plant monitoring apparatus, a prediction is made by a linear or non-linear regression model created from plant data in a normal state, and the prediction error is compared with a certain threshold value or is determined by performing a statistical evaluation. Becomes possible.

Further, according to the third aspect of the present invention, a supervisory processor extracts a small number of principal component signals from a plurality of plant data provided, and executes regression analysis of plant data to be determined using the obtained principal component signals. Is what you do.

In the above plant monitoring apparatus, it is possible to extract a small number of principal component signals without using all of the plurality of signals, and perform regression analysis of target plant data narrowed down to the principal component signals obtained at this time. . Thereby, the monitoring accuracy can be further improved.

According to a fourth aspect of the present invention, the display device selects a signal to be determined from a plurality of input signals used for the determination.
As a function of two signals, whether it is normal on the three-dimensional display graph,
The color is displayed differently depending on whether or not the image is displayed, and is displayed based on the curved surface of the linear or nonlinear regression model used for the determination.

In the above plant monitoring apparatus, two signals are selected from the input signals, taken on both axes on a plane, and the behavior of the monitored signal is appropriately displayed in a three-dimensional graph with the predicted signal on the vertical axis. be able to. The predicted value of the regression model used for the determination is displayed as a curved surface graph in the three-dimensional graph, and the observed signal to be determined is displayed as a point in the graph, so that the normal or abnormal plant behavior can be intuitively determined. You can judge. In addition, the points that became abnormal are displayed in different colors,
It is possible to easily determine the presence or absence of an abnormality even in a three-dimensional graph.

Further, according to a fifth aspect of the present invention, the display device sets the variation range of the input signal used for creating the reference regression model to two.
Displayed on a two-dimensional graph, and displayed on the same two-dimensional graph with different colors depending on whether or not the input signal used for semi-fixed time input is within a certain error range in the fluctuation range of the reference regression model It is something to do.

In the above plant monitoring apparatus, the fluctuation range of the input signal is displayed on a two-dimensional graph in order to determine that there is no error in the judgment according to the reference regression model obtained based on the plant data in the normal state. This makes it possible to prevent erroneous determinations from occurring.

The present invention according to claim 6 is characterized in that the monitoring processor is normal / abnormal only when the input signal at the time of determination deviates from the fluctuation range of the input signal used for creating the reference regression model within a certain error range. Is not obtained.

In the above plant monitoring apparatus, the judgment itself is forgotten only when the input signal deviates from the behavior range of the input signal used to create the reference regression model within a certain error range. This makes it possible to prevent erroneous determinations from occurring.

Further, the invention according to claim 7 is characterized in that the monitoring processor is normal / abnormal only when the input signal at the time of determination deviates from the fluctuation range of the input signal used for creating the reference regression model within a certain error range. Is obtained, and the reference regression model is re-learned using the plant data.

In the above plant monitoring apparatus, if the judgment itself is forgotten and, at the same time, plant data exceeding the behavior range of the input signal used for learning the reference regression model is obtained, learning is performed again using the plant data. . This makes it possible to gradually expand the application range of the reference regression model.

[0025]

FIG. 1 shows an overall configuration of an embodiment of a plant monitoring apparatus according to the present invention. The plant data input apparatus 1 inputs a plurality of time-synchronized plant data from a plant. A data storage device 2 for storing plant data given from the plant data input device 1, a reference creation device 3 for creating a monitoring standard based on normal plant data, and a plant data stored in the data storage device 2. Is read and the reference creation device 3
And a display device 5 for displaying the plant data and the monitoring standard on a display device such as a CRT.

In the above configuration, the signal obtained from the plant is converted into time-series data by the plant data input device 1, transmitted to the data storage device 2 and stored therein. The data storage device 2 constantly and continuously stores plant data for a certain period. Here, a plurality of plant data transmitted from the plant data input device 1 are continuously stored cyclically in a form as shown in FIG. 2 with data at a plurality of different sampling intervals. It is created so that data up to the point is stored.

Further, the reference creation device 3 identifies a linear or non-linear regression model using data at an arbitrary sampling interval stored in the data storage device 2 and uses the data to create a reference as shown in FIG. The plant data, the obtained reference regression coefficient, the prediction error at the time of identification, the criterion at the time of abnormality determination, and the additional plant data at the time of learning are stored.
The prediction error at the time of identification is used by the abnormality determination processing device together with a determination criterion at the time of abnormality determination that is input in advance.

Further, the monitoring processing device 4 performs an abnormality determination process using the data at an arbitrary sampling interval of the data storage device 2 and the monitoring criterion given from the criterion creating device 3. Further, the display device 5 displays an abnormality determination result, a reference regression model, and the like.

Hereinafter, embodiments of each device will be described in detail. The reference creation device 3 creates a prediction model of a signal to be monitored from plant data in a normal state. The monitoring target signal targets each signal taken into the device. Assuming that a specific signal to be monitored is y (t), this signal is obtained from m input signals x1 (t)... Xm (t) by y (t) = f (x1 (t),. ))… (1) In the case of the linear regression model, t is omitted, and y = a + b1 * x1 +... + Bm * xm (2). In the case of the quadratic expression, y = a + b1 * x1 + ... + bm * xm + c1 * x1 * x1 + ... cm * xm * xm + d12 * x1 * x2 + ... + dm-1, m * xm-1 * xm (3) It takes shape.

As shown in FIG. 4, from m inputs, a linear,
It is possible to predict a signal to be monitored by a non-linear or neural network. The coefficient here is based on the normal plant data.

(Equation 1) That is, it can be determined to minimize the root mean square of the error. The model to be used is determined so as to minimize the weighted average of the prediction error I (a, b1,..., Bm) of each model and the complexity (number of coefficients) of the model.

As another method of this prediction method, FIG.
The principal component analysis as shown in FIG. Here, two or three small signals (Ui, i = 1, 2,...) Are extracted from the m input signals (Xi, i = 1, 2,... M) by principal component analysis. Using the extracted time-series signal (U), the monitoring target signal T is predicted by Ta = f (U) (5). Here, (Ti, i = 1,2, ...) is the monitoring target signal (yi, i
= 1, 2,... K) obtained by performing principal component analysis. When there is one signal to be monitored, in the present embodiment, T =
It becomes Y. As the prediction function f (U), any of the linear, nonlinear, and neural networks described above can be applied. The abnormality determination is performed based on whether or not the deviation between the observed principal component signal T and its predicted value Ta exceeds a certain threshold. Such extraction of the principal component signal is particularly useful in monitoring a large-scale plant for the following reasons.

In a large plant, the monitored signal of interest fluctuates due to a number of other process signals. However, it is not affected independently of all process signals and is usually affected by a small number of state variables. For example, in a nuclear power plant, normal operation is controlled by output and flow rate, and although many observed signals behave in a complex manner, when extracting principal components, they are represented by a small number of state variables corresponding to output and flow rate. be able to. By predicting the plant state with such a small number of signals, it is possible to create a reference regression model that can be universally applied even in various operation cycles.

Furthermore, if the plant follows different output change curves in different operation cycles, the observation itself will behave differently between the cycles, but by extracting and evaluating the principal components, the cycle can be evaluated. It can be determined that the states between are equivalent.

Further, in the present embodiment, a plurality of signals to be monitored can be simultaneously monitored as a group. This produces a small number of monitored signals T by applying principal component analysis to k monitored signals, as shown in FIG. The monitoring target signal T can be monitored by performing least square prediction using the prediction function f (U).

In preparing the above-mentioned reference model, the signals to be used are previously set to have an average of zero and a fluctuation width of -1.
It is assumed that the stability in numerical calculation is ensured by normalizing to the range from 1 to 1.

The reference creation device 3 stores the coefficients of the reference regression model, the reference data used for identification, the magnitude of the prediction error at the time of identification (standard deviation σ), and the normalization constant at the time of identification as a database. , To the monitoring processor 4.

Further, in the monitoring processing device 4, the determination target signal y (t) at the monitoring time provided from the data storage device 2 is provided.
And m input signals xm (t), calculate a prediction error δy by the following equation, and compare it with a threshold value.

Δy = y−f (x1,..., Xm) (6) The threshold value is k times the error σ at the time of creating the reference model or a fixed value given in advance.

In this embodiment, it is possible to make an abnormality judgment using a hypothesis testing method such as a sequential probability ratio test, utilizing the fact that the prediction error δy is normally zero on average and given by the variance σ. is there.

Since the prediction error of the regression model is used in the judgment by the present apparatus, it is important to check the applicable range of the model in order to avoid erroneous judgment. In order to confirm the application range of the model, as shown in FIG. 6, a variation range of the input data used for identifying the reference model is displayed with a certain tolerance, and the input data at the time of monitoring is displayed in the range. It is determined whether it is included or not. If it is not within the allowable range, the monitoring processor 4 forgoes obtaining the determination result.

In addition, instead of the above method, when the plant data is not included in the allowable range, the plant data can be added to the reference data and the reference regression model can be learned again. As a result, it is possible to increase the monitoring range as the monitoring time increases.

Next, the display device 5 will be described in detail. In a conventional display device, a time series trend of a monitoring target signal is shown by comparing it with a predicted value thereof, and a time trend of a prediction error is displayed. These display methods are effective in determining at which point the abnormality has occurred, but it is not possible to know from what input value this predicted value was predicted,
Prevents intuitive understanding of operators.

A two-dimensional correlation display is also performed in which one of the input signals is plotted on the horizontal axis and the signal to be monitored is plotted on the vertical axis. This is because the signal to be monitored is affected by a plurality of input signals. Is often not enough to understand the behavior of

On the other hand, in a large-scale plant, by taking advantage of the fact that the behavior of the monitored signal is not independently affected by all of the input signals that can be affected, a relatively small number of components can be obtained by taking the main components of the input signal. The behavior of the monitoring target signal can be predicted from the principal component signal. By taking two of the principal component signals on both axes on a plane and displaying the behavior of the monitoring signal as a three-dimensional graph on the vertical axis, the correlation between input and output can be intuitively grasped. .

In the present embodiment, as shown in FIG. 7, data to be monitored is displayed as points together with a reference curve calculated from a reference regression model on a three-dimensional graph. Also,
Since it is difficult to intuitively grasp the deviation from the reference phase on the three-dimensional screen, abnormal data in the determination, that is, points that are separated from the reference curved surface by a certain distance or more are displayed in different colors. It is also important to determine whether or not the data at the time of this monitoring is within the allowable error range of the input data at the time when the reference regression model was created. This is based on the two-dimensional graph shown in FIG. And whether or not it is included in the allowable range can be determined by different colors.

FIG. 8 shows the plant behavior in different operation cycles on a three-dimensional graph. Although different paths are taken on the three-dimensional graph, there is a difference only by looking at the time change of the monitoring signal as shown in FIG. 9, but as shown in FIG. 8, the same reference surface is taken on the three-dimensional graph. And it can be determined as normal.

On the other hand, FIG. 10 shows a case where all the monitoring signals of the plant are input, and after performing principal component analysis, individual monitoring signals y are predicted by the above-mentioned linear / nonlinear regression model, and the deviation from the observed value is calculated. The example which performed abnormality detection is shown.
In the case of a sensor failure that often occurs in a plant, the output deviation increases by the corresponding signal, and the sensor failure can be specified and detected. Here, the monitored signal y
Is characterized by the fact that the same signal is also used as an input when predicting, and the number of inputs and outputs is the same, and the extracted principal components are common.

FIG. 11 shows a method different from the above method. It is an example of the same sensor failure diagnosis, but in order to improve the disadvantage of using the same signal as the output for the input, the same input signal corresponding to each monitored signal is omitted, and the principal component analysis is performed only on the remaining signals, Create a regression model and make predictions. This can compensate for the drawback of the method shown in FIG. 10 in that when a sensor failure occurs, the signal is also used as an input signal, leading to a reduction in prediction accuracy.

[0049]

According to the present invention, the long-term behavior from the start of the plant to the stop of the plant is monitored with higher accuracy compared to the reference model obtained from the linear or non-linear regression model created by the reference creation device. be able to.

Therefore, according to the present invention, various abnormalities occurring in the system, equipment, etc. in a large-scale plant can be discovered at an early stage, and the soundness of the large-scale plant in operation is prevented from being impaired. Can be prevented.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing an embodiment of a plant monitoring device according to the present invention.

FIG. 2 is a diagram showing a data storage mode of the data storage device of the present invention.

FIG. 3 is a diagram showing a data storage form of the reference creation device of the present invention.

FIG. 4 is a diagram showing a multivariable regression model using a linear, non-linear, and neural network for predicting a monitoring signal according to the present invention.

FIG. 5 is a diagram showing a regression prediction model of a main component of an input signal and a main component of a monitor signal according to the present invention.

FIG. 6 is a diagram showing a fluctuation range of an input signal used for creating a model of the present invention and a fluctuation range of an input signal during monitoring.

FIG. 7 is a diagram showing a display example of a monitoring result by a three-dimensional graph in which an input signal of the present invention is taken on two horizontal axes and a monitoring signal is taken on a vertical axis.

FIG. 8 is a diagram showing a display example in which operation histories of different operation cycles according to the present invention are compared on a three-dimensional graph.

FIG. 9 is a view showing a display example in which the operation histories of different operation cycles according to the present invention are compared by time trend.

FIG. 10 is a diagram showing a configuration example of a sensor diagnosis for predicting m signals by a regression model of m input signals according to the present invention.

FIG. 11 is a diagram showing a configuration example of sensor diagnosis for performing principal component analysis for each prediction target signal according to the present invention.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 plant data input device 2 data storage device 3 reference creation device 4 monitoring and processing device 5 display device

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Mitsuhiro Enomoto 8 Shinsugita-cho, Isogo-ku, Yokohama-shi, Kanagawa Prefecture Inside the Toshiba Yokohama Office (72) Inventor Masahiro Okawa 8 Shinsugita-cho, Isogo-ku, Yokohama-shi, Kanagawa (72) Inventor Yu Fujinami 8th Shin-Sugitacho, Isogo-ku, Yokohama-shi, Kanagawa Prefecture F-term (reference) 5E501 AA01 AC02 AC32 CA03 EA32 FA27 FA46 FB28 5H004 GA28 GA40 GB01 JB07 JB18 KC26 KC28 KD42 MA48 MA50 5H223 AA01 EE02 FF05

Claims (7)

[Claims] 1. A plant data input device for inputting plant data from a plant, a data storage device for storing plant data given from the plant data input device, and a normal plant data stored in the data storage device A reference creation device that creates a monitoring reference based on the data, a monitoring processing device that reads plant data stored in the data storage device, and determines normal / abnormal according to the monitoring reference given by the reference creation device; A display device for displaying the plant data and the monitoring standard given to the processing device together with the determination result. 2. The system according to claim 1, wherein the monitoring processing apparatus determines whether the plant data to be determined is normal or abnormal by inputting a plurality of other plant data and predicting the linear or non-linear regression model created from the normal plant data. The plant monitoring device according to claim 1, wherein the determination is performed using a difference between the value and the plant data to be determined. 3. The method according to claim 2, wherein the monitoring processing device extracts a small number of principal component signals from the plurality of plant data provided, and performs regression analysis of the plant data to be determined using the obtained principal component signals. 3. The plant monitoring device according to claim 2, wherein: 4. A function of two signals selected from a plurality of input signals used for determination by the display device as a determination target signal, and different colors on a three-dimensional display graph depending on whether the display is normal or not. The plant monitoring apparatus according to claim 3, wherein the display is performed based on a curved surface of the linear or non-linear regression model used for the determination. 5. A two-dimensional graph showing a fluctuation range of an input signal used by the display device to create a reference regression model,
The color is changed depending on whether the input signal used for the input at the time of the judgment is included in a certain error range in the fluctuation range of the reference regression model, and is displayed on the same two-dimensional graph. The plant monitoring device according to claim 4, wherein 6. A normal / abnormal judgment result is obtained only when the input signal at the judgment point deviates from a fluctuation range of the input signal used for creating the reference regression model by the monitoring processing device within a certain error range. 4. The plant monitoring device according to claim 2, wherein the plant monitoring device is set off. 7. A normal / abnormal judgment result is obtained only when the input signal at the judgment point deviates from a fluctuation range of the input signal used for creating the reference regression model by the monitoring processing device within a certain error range. 7. The plant monitoring apparatus according to claim 6, wherein the reference regression model is relearned using the plant data.