JPH11161327A - Method and device for diagnosing abnormality of process - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily decide whether or not a process is abnormal by calculating a predicted value by using state quantities except that of the process itself and then calculating a normalized output value, and comparing the mean value of output values found as to state quantities with a decision value. SOLUTION: A predicted value arithmetic means 12 calculates the predicted value from detected values of other sensors except itself by using a neural network model as to the detected values of the state quantities of the process and outputs it. An error normalizing arithmetic means 13 performs normalization by using the predicted value and detected values so that the absolute differences of the standard deviation of the predicted error of the model that the predicted value arithmetic means 12 has and the predicted values corresponding to the state quantities based on past time-series data are distributed in an arbitrary range. An error averaging arithmetic means 14 calculates the averaged state quantity, and a process abnormality decision means 15 compares a process state index with the decision value and sends an abnormality signal to a display means 16 when the decision value is exceeded.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method for determining a process abnormality and an apparatus for determining a process abnormality.

[0002]

2. Description of the Related Art Chemical plants are equipped with sensors for measuring various state quantities such as flow rate, temperature, pressure, water level, and composition of chemical substances. If these sensors fail, there is a possibility that the quality of the product will be degraded or the operation will be stopped. Conventionally, it has been an issue to introduce a process monitoring system that can constantly monitor the state of the plant and quickly catch the situation in the event of an abnormality.
In this system, how to deal with sensor failure is an important issue.

[0003] As a technique for monitoring an abnormality, conventionally, a state quantity detected by a sensor is used to predict another state quantity, and this is compared with a detected value of the state quantity. A method for detecting whether the value is normal or abnormal has been proposed. Many mathematical models for predicting state quantities that use physical models such as mass balance and heat balance have been proposed. These models include process nonlinearities, multivariable structures, and labor required for modeling. In view of the above, it is often difficult to deploy in a large-scale plant. However, in recent years, the use of neural networks has made it relatively easy to build mathematical models using past process data, and this mathematical model has reduced the time required to calculate predicted values. Therefore, by comparing the predicted value and the detected value, the abnormality of the sensor can be immediately judged online.

[0004]

However, an abnormality in the detected value of the sensor can be determined by comparing the detected value with the predicted value using the mathematical model as described above, but the abnormality in the detected value of the sensor is a failure of the sensor. It is difficult to determine whether the error is due to an abnormality in the process itself where the sensor is being used. SUMMARY OF THE INVENTION It is an object of the present invention to provide a method and an apparatus for determining an abnormality in a process irrespective of an abnormality in a detection value of each sensor.

[0005]

According to the present invention, there is provided an abnormality diagnosis method for detecting a plurality of state quantities of a process and diagnosing the presence or absence of an abnormality in the process based on the detection result. For each of the detected state quantities, a predicted value is calculated using the remaining detected state quantities excluding itself, and the absolute value of the difference from the corresponding predicted value for each detected state quantity is calculated as a standard of the error having the predicted value. By calculating an output value normalized so as to be distributed in the same range by dividing by the deviation, calculating an average value of the output values for the plurality of state quantities, and comparing this with a predetermined determination value. The presence or absence of a process abnormality can be easily determined. Further, the determination method is an abnormality diagnosis device for diagnosing the presence or absence of a process abnormality based on a plurality of state quantities detected by a sensor, and for each of the plurality of state quantities, the remaining detected quantity excluding itself is detected. Predictive value calculating means for calculating a predicted value based on the detected state quantity, and dividing an absolute value of a difference between each detected state quantity and a corresponding predicted value by a standard deviation of an error of the predicted value. Error normalizing operation means for calculating an output value normalized so as to be distributed in the same range, error averaging operation means for calculating an average value of the output values for the plurality of state quantities, and And a determination means for determining whether or not the process is abnormal by comparing the value with a predetermined determination value. Kill.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An ordinary chemical plant is composed of a plurality of processes. A number of devices are incorporated in each process, and from each device, a number of detectors (sensors) detect the state quantities of the process, such as temperature, pressure, flow rate, liquid level, and composition of the process substance. The state quantity of the detected process is distributed instrumentation system (DCS)
Is displayed to the operator via the like.

In the method of the present invention, a plurality of state quantities of a process are detected by a large number of sensors, and the presence or absence of a process abnormality is diagnosed based on the detected state quantities. FIG.
5 is a flowchart showing an algorithm used for abnormality diagnosis of the method of the present invention.

In step 1, for each of the plurality of state quantities detected from the equipment of the process, a predicted value is calculated using the remaining plurality of detected state quantities excluding itself. Means for calculating the predicted value is not particularly limited.For example, based on time series data of the state quantity of the past process, a statistical model such as a regression equation, a model constructed by a neural network, or a process mass balance or Although a physical model constructed based on heat balance may be used, it is preferable to use a model using a neural network from the viewpoint of reducing the time required for modeling and speeding up the operation.
When a plant is composed of a plurality of processes, a mathematical model corresponding to each process may be constructed and operated.

In step 2, the absolute value between the detected value of each state quantity and the difference between its corresponding predicted value is normalized. This is based on, for example, the standard deviation of the error of the predicted value determined depending on the past time series data used to construct the mathematical model for calculating the predicted value, and the past time series data of the state quantity. Is performed by dividing the absolute value of the difference between the detected value of the state quantity and the corresponding predicted value by a coefficient determined so as to be distributed in the same range for a plurality of state quantities.

The method of calculating the coefficient will be described more specifically. The error of the predicted value is distributed according to a certain existence probability distribution, and this existence probability distribution can be obtained based on past time-series data. When the error of the predicted value is plotted on a rectangular coordinate system with the horizontal axis representing the error of the predicted value and the vertical axis representing the distribution of the existence probability, the probability that the error of the predicted value exists in a certain value width is calculated by integrating the existence probability distribution in the width. Given as a value.
Generally, the width of the value at which the error existence probability is 99.5% is set to be three times the standard deviation, and the absolute value of the difference between the detected value and the corresponding predicted value is within the range. Is considered to be within the normal error range. When actual detection value data is used, it may not follow a general existence probability distribution.Therefore, to obtain the above coefficient for a plurality of state quantities, the product of this coefficient and the standard deviation of the error of the predicted value is used. May be determined such that the existence probability of the error of the predicted value, which exists within the range of, is constant for a plurality of state quantities. As a result, the state quantities of processes having different dimensions can be handled in a unified manner.

In step 3, an output value (hereinafter, referred to as an average value) obtained by averaging the absolute value of the difference between the detected value and the predicted value of the state quantity normalized as described above for a plurality of state quantities of the process. "Process state index"). In averaging, the sum of the values normalized in step 2 may be simply divided by the number of state quantities, or important state quantities may be distinguished from non-important state quantities. For this purpose, values obtained by multiplying the values normalized as described above by coefficients corresponding to the respective weights may be added together and divided by the number of state quantities.

In step 4, an abnormality is determined by comparing the process state index calculated as described above with a determination value previously determined based on past time-series data. If the process state index is larger than the determination value, the process is diagnosed as having an abnormality. If the process state index is smaller than the determination value, the process is diagnosed as having no abnormality. I do. When it is determined that there is no abnormality in the process, if the difference between the detected value and the predicted value of any sensor is large and it is determined that the process is abnormal, it can be determined that the sensor itself is abnormal.

If it is determined that there is an abnormality in the process, the process proceeds to step 5 to output an abnormality occurrence output, for example, a display means such as a CRT display, an alarm output means, a printer means for printing on a paper medium by a printer, and the like. The abnormality of the process is transmitted to the operator via the transmission means. In this case, even if the detection value of any of the sensors is determined to be abnormal, it cannot be determined whether or not the sensor itself is abnormal. By performing the above steps periodically, abnormality diagnosis of the process of the plant is realized.

The above-described algorithm can be executed by an apparatus such as a computer in which a program for executing the algorithm is installed. For example, a program including a predicted value calculating means for executing step 1 described above, an error normalizing means for executing step 2, a process state index calculating means for executing step 3, and a judging means for executing step 4 may be used as a workstation, for example. The system may be implemented on a simple computer, and a system may be constructed so that the detection value of the sensor of the state quantity of the process can be collected via the distributed instrumentation system / process computer.

The process state index calculated and output as described above reflects the difference between the detected value and the predicted value of the state quantity detected from the process as a whole. That is, even if the difference between the detected value of a single state quantity and the corresponding predicted value is larger than the allowable range, if the difference between the detected value of other state quantities and the predicted value is sufficiently small, the process Since the state index does not become a value larger than the allowable range, it is not determined that the process is abnormal, and it is determined that the sensor is abnormal.
Even if the difference between the detected value and the predicted value of each state quantity is within the allowable range, if many of them show relatively large values, the process state index becomes a value larger than the allowable range, and the process Is determined to be abnormal. As described above, by comparing the process state index with the predetermined determination value from the time series data of the state quantity of the past process, it is possible to correctly determine whether or not the process is abnormal. Hereinafter, the present invention will be described more specifically with reference to FIG.

The plant 6 to be monitored for abnormality is a plant for producing vinyl chloride monomer from ethylene and chlorine, and is a relatively large plant in which the detection values of the sensors exceed 2000 points at all points. This plant is 2
The process was divided into blocks of process 0, and a neural network model was constructed using only the detected values of the sensors having correlation in each block. The number of input points of the state quantity of the process of the model was 100 or less for each model. These sensors detect a plurality of state variables of a process in the plant. The detection signal of the state quantity is provided to the process data collection unit 9 in the process computer 8 via the distributed instrumentation system (DCS) 7. The detection signal is stored in the process database 10 as a detection value of the state quantity of the process, and is given to the predicted value calculation means 12 in the workstation 11.

The predicted value calculation means 12 has a neural network model of the divided process,
For each of the detection values of the state quantities of the process given from the process data collection unit 9, a prediction value is calculated from the detection values of the other sensors excluding itself and output using the corresponding neural network model, and the error is calculated. It is provided to the normalizing means 13.

The error normalizing operation means 13 uses the predicted value output from the predicted value operation means 12 and the detected value provided from the process data collection unit to calculate the detected value and predicted value for each state quantity. Is determined such that the difference between the state quantity and the corresponding predicted value is distributed in an arbitrary range based on the standard deviation of the prediction error of the model of the predicted value calculating means 12 and the past time-series data. The result is normalized by being divided by the coefficient, and is provided to the error averaging operation means 14.

The error averaging calculating means 14 multiplies the output value of the error normalizing means 13 by a predetermined coefficient for each process block constituting the plant to be monitored, and calculates the state quantity of the process block. Then, the averaged state quantity is calculated by dividing by the number of pieces, and is provided to the process abnormality determining means 15.

The process abnormality judging means 15 compares the process state index provided from the error averaging operation means 14 with a judgment value given in advance for each process. If the process state index exceeds the judgment value, an abnormal signal is sent to the display means 16.

On the CRT display of the display means 16, a block diagram showing the process flow of the monitoring plant is displayed. By changing the display color of the block indicating the process to which the abnormal signal has been sent, the plant operator can be notified. You will be notified of the presence of the anomaly. The operator can see the screen on which the output value of each sensor can be checked by expanding the portion of the process flow diagram where the display color has been changed. Since the predicted value of the process state quantity, the output value of the error normalizing operation means, and the process state index are stored together with the detected value in the process database, these values can be referred to as a trend in a time series. I have.

FIG. 3 (a) is a trend graph of a process state index immediately after start-up of a distillation process of hydrochloric acid produced as a by-product in the vinyl chloride monomer production process of the above plant, and FIG. 3 (b) is a graph showing the same process. It is a trend graph of a process state index at the time of steady operation. . FIG. 3 (a)
Since the process state index immediately after start-up is unstable, the process state index shows a large value as a whole as compared with the process state index shown in FIG. 3B, and the process becomes stable over time. It can be seen that the value gradually decreases as it goes down. The process state index shown in FIG. 3B is stable reflecting the stable state of the plant, and its value is small.

[0023]

According to the present invention, a plurality of state quantities of a process are detected, and for each of the plurality of state quantities, a predicted value is calculated using a plurality of remaining state quantities excluding itself, and each state quantity is calculated. The absolute value of the difference between the detected value and the corresponding predicted value is divided by the standard deviation of the error of the predicted value to calculate a normalized output value so that the output value is distributed in the same range. By comparing the average value of each process for a plurality of state quantities of the output value with a predetermined determination value,
The presence / absence of a process abnormality can be determined by distinguishing from a device abnormality such as a sensor failure.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an algorithm used for abnormality diagnosis of the method of the present invention.

FIG. 2 is a block diagram showing an example of an abnormality diagnosis device according to the present invention.

FIG. 3 (a) is a non-stationary trend graph of a process state index as an example to which the abnormality diagnosis method of the present invention is applied,
(B) is a steady state trend graph of the process state index as an example to which the abnormality diagnosis method of the present invention is applied.

Claims (4)

[Claims] 1. An abnormality diagnosis method for detecting a plurality of state quantities of a process and diagnosing the presence or absence of an abnormality in the process based on a result of the detection. A predicted value is calculated using the estimated state quantity, and an absolute value of a difference between the detected state quantity and the corresponding predicted value of the state quantity is divided by a standard deviation of an error having the predicted value. An output value normalized to be distributed in the same range is calculated, an average value of the output values is obtained for the plurality of state quantities, and this is compared with a predetermined determination value to determine whether there is a process abnormality. A method for diagnosing an abnormality of a process, characterized by determining. 2. The process abnormality diagnosis method according to claim 1, wherein the means for calculating the predicted value of the state quantity is a calculation means using a neural network model. 3. An abnormality diagnosis device for diagnosing the presence or absence of a process abnormality based on a plurality of state quantities detected by a sensor, wherein each of the plurality of state quantities excludes the remaining detected quantity excluding itself. A predicted value calculating means for calculating a predicted value based on the state quantity, and dividing an absolute value of a difference between each detected state quantity and a corresponding predicted value by a standard deviation of an error having the predicted value. Error normalizing operation means for calculating an output value normalized so as to be distributed in the same range, error averaging operation means for calculating an average value of the output values for the plurality of state quantities, and calculating the average value in advance. A process abnormality diagnosis device comprising: a determination unit configured to determine whether there is a process abnormality by comparing the determined value with a predetermined determination value. 4. The process abnormality diagnosis apparatus according to claim 3, wherein the predicted value calculating means for calculating the predicted value of the state quantity is a means based on a neural network model.