JPH08328644A - Method for analyzing fluctuation in process - Google Patents

Abstract

PURPOSE: To analyze significant fluctuation in a process systematically. CONSTITUTION: In a process having lots of process variables in which causality relations are related to each other in terms of a tree structure, the process variables are traced sequentially along a path from the resulting point toward the cause point, secular measurement data of two adjacent process variables are subjected to FFT processing at need to obtain the correlation coefficient and a weight is multiplied at need by the plural correlation coefficients along the path to obtain a geometric mean and the degree of the effect of the process variables of the cause side onto the fluctuation in the process variables of the result side is analyzed depending on the magnitude of the geometric mean obtained from all possible paths. Then the maximum geometric mean among those obtained from all possible paths is obtained and a path indicating the maximum geometric mean is extracted in this analysis.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for analyzing fluctuations in various processes.

[0002]

2. Description of the Related Art As a conventional method for analyzing the variation of a process having a large number of process values that are causally related to each other in a tree structure, a threshold value is set individually for each process value and meaning is set. There is a method in which the fluctuations are detected individually, and then the detected fluctuations are systematically analyzed using IF-THEN rules corresponding to preset meanings.

For example, in the calorific value adjustment process of city gas, the following reasoning flow can be given as an example. That is, it is detected that the calorific value exceeds the deviation → The change in the flow rate of the line with the mixer is detected, and if there is no conditional branching → If there is no, pay attention to the slope of the calorific value signal and check the IF corresponding to this signal. -Determine the cause by THEN rule (ie, (a) large slope of IF calorific value THEN loop abnormality,
(b) IF calorific value is declining and the THEN instrument is abnormal, and (c) IF calorific value signal is interrupted, THEN instrument power is shut off.

[0004]

SUMMARY OF THE INVENTION In such a method,
The purpose of the IF-THEN rule is to specify the behavior of temporal data for each process value with individual and fixed threshold values without any systematic norm corresponding to the process state etc. In some cases, the conditional clause may not hold and the analysis may not be as intended. The present invention aims to solve such problems.

[0005]

In order to solve the above-mentioned problems, according to the present invention, first, in a process having a large number of process values that are causally related to each other in a tree structure, one route from the result side to the cause side is used. By sequentially tracing the process values along with, the time-dependent measurement data for each two adjacent process values are processed as they are, or all are FFT processed, or if there is a non-negligible time delay between them. FFT processing is performed to obtain those cross-correlation coefficients, a geometric mean of a plurality of cross-correlation coefficients along the path is obtained, and the result side is determined by the magnitude of the geometric mean obtained for all possible paths. We propose a method to analyze the fluctuation in the process, which analyzes the degree of the influence of the process value on the causative side to the fluctuation of the process value.

Further, according to the present invention, in a process having a large number of process values which are causally related to each other in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and two adjacent processes are processed. Each time-dependent measurement data for each value is subjected to FFT processing as it is, or if there is a non-negligible time delay between them, FFT processing them to obtain their cross-correlation coefficient, A plurality of cross-correlation coefficients along the path are multiplied by a predetermined weight to obtain a geometric mean thereof, and the magnitude of the geometric mean obtained for all possible paths is used for the fluctuation of the process value on the result side. We propose a method for analyzing fluctuations in the process of analyzing the degree of influence of process values on the causative side. In this method, the weight can be set by, for example, the flow rate ratio of the fluid flowing through the path.

In the above-described method of the present invention, the process value in which the variation is detected by comparison with the threshold value is used as the starting process value, and the process value can be traced from this process value to the cause side for analysis.

Further, in the above-mentioned method of the present invention, as one of the analysis of the fluctuation in the process, the operation of obtaining the maximum value of the geometric mean obtained for all possible routes and extracting the route showing this maximum value is proposed. To do.

[0009]

In the two process values having a causal relationship, the fluctuation of the process value on the cause side due to the significant fluctuation of the process also appears as the fluctuation on the process value of the result side. Therefore, when the cross-correlation coefficient between the measurement data of these process values over time is obtained, the value increases according to the strength of the causal relationship. On the other hand, when the time-series data of the process values on the cause side and the result side are subjected to FFT processing, the common variation appears as a component of a specific frequency on the frequency axis, and these are not affected by the time delay of the time-series data. Therefore, when these cross-correlation coefficients are obtained, the value increases in accordance with the strength of the causal relationship, as described above, and serves as an index indicating the strength of the causal relationship. That is, both the raw cross-correlation coefficient of the process data of the cause side and the result side and the cross-correlation coefficient of the FFT value of the time data are indicators of the strength of the causal relationship, and the latter is the time series. This is useful when there is a non-negligible time delay between the data. In this way, when there are multiple causal processes for the process value on the result side, the cross-correlation coefficient of each causes the causal relationship between the process value on the result side and each process value on the causal side. You can ask for strength.

By deriving such a causal relationship strength from the result side to the cause side in order, the strength of the causal relationship along one path across a plurality of layers can be obtained. By obtaining the geometric mean of the cross-correlation coefficients sequentially obtained as described above, it is possible to normalize the difference in the number of route layers and obtain the strength of the causal relationship at the same level.

In this way, if the geometric mean is calculated for all possible above-mentioned paths in a process having a large number of process values that are causally related to each other in a tree structure, the magnitudes of the geometric mean along the respective paths are determined according to their magnitudes. It is possible to analyze the degree of influence of the process value on the causative side on the fluctuation of the process value on the result side.

The degree of fluctuation on the result side due to the fluctuation on the cause side also changes depending on conditions such as the flow rate of the fluid in the process. In this case, these flow rate ratios are used as weights and the cross-correlation coefficient obtained as described above is multiplied, and thereafter the geometric mean is calculated.

By performing the above-mentioned calculation at predetermined time intervals with respect to time-dependent data in a predetermined time range, it is possible to detect and analyze fluctuations in real time.

[0014]

The present invention will be described below with reference to the embodiments shown in the drawings. First, FIG. 1 schematically shows a configuration of an example of a process to which the present invention is applied. In this process, the flow rate adjusting paths A, B and C merge with the pressure adjusting path D and the flow rate adjusting paths B and C are combined. Shows a configuration in which the fluid supply paths E and F merge through the orifices 1b and 1c. In each of the routes A, B, C and D, F is a flow rate measuring unit and P is a pressure measuring unit. In the routes A, B and C, the measured value of the flow rate measuring unit F is input to the controller R to regulate the flow rate. Valves 2a, 2b, 2
c is operated to adjust the flow rate, and in path D, pressure measuring unit P
The measured value is input to the controller R to operate the pressure control valve 3 to control the pressure. Further, a heat quantity measuring section Q is provided on the path D, the measured value of the heat quantity measuring section Q is inputted to the heat quantity controller M, and the output of the heat quantity controller M is inputted to the controller R of the path A. The flow rate setting value is changed. Further, vaporizers V are provided on the paths E and F, respectively.

FIG. 2 relates to the process shown in FIG. 1. The upper left stage is the flow rate value F (A) of the route A, the middle left stage is the flow rate value F (B) of the route B, and the lower left stage is the pressure of the route D. Value P
It shows the time transition of (D). Flow rate value F (A) or F
(B) and the pressure value P (D) have a causal relationship in which the former is the causative side, and there is a time delay between the measurement points of the former and the latter. That is, as shown in the figure, the time-dependent data of the flow rate value F (A) has a fluctuation at the time point t ₁ , and this fluctuation, after a certain period of time, passes through the pressure value P (B) at the time point t _2.
It appears as a variation in the time-course data of.

When the time delay between the two process values is large, the range of shifting the data in the calculation of the cross-correlation coefficient becomes large, the time interval of the temporal data required for the calculation becomes large, and the required calculation amount becomes large. In this case, the chronological measurement data of the two process values is
After T processing and conversion into frequency axis data, a cross-correlation coefficient between the respective conversion values is calculated. That is, the waveform on the right side of the figure is obtained by performing FFT processing on the respective temporal data on the left side, and by performing this conversion, the frequency component of the fluctuation included in the temporal data is extracted. That is, in the FFT processing data of the flow rate value F (A) of the route A and the flow rate value F (B) of the route B, a remarkable peak appears around the frequency f ₁ , whereas the flow rate value of the route B No remarkable peak appears in the F (B) FFT processed data.

The result of calculating the cross-correlation coefficient between the data on the frequency axis obtained as described above is shown in FIG. 3. The upper waveform in the figure shows the flow rate value F (A) and the pressure value P.
The correlation coefficient of the frequency axis data of (D), and the lower waveform in the figure shows the correlation coefficient of the frequency axis data of the flow rate value F (B) and the pressure value P (D). In this figure, a remarkable peak appears around the frequency f ₁ in the upper correlation coefficient, whereas
No peak appears in the lower correlation coefficient. From this,
Even if there is a process time delay between measurement points of two process values, the strength of the causal relationship between the process value on the cause side and the process value on the cause side with respect to the fluctuation of the process value on the result side is obtained by the cross-correlation coefficient for the FFT-processed frequency axis data. You can see that

Based on the above, according to the present invention, as described above, in a process having a large number of process values that are causally related to each other in a tree structure, the process is sequentially performed from the result side to the cause side along one path. By tracing the values, the time-dependent measurement data for each two adjacent process values is subjected to FFT processing if necessary to obtain their cross-correlation coefficient, and a plurality of cross-phases along the above-mentioned path are obtained. The number of relations is multiplied by a weight if necessary to obtain a geometric mean of them, and according to the magnitude of the geometric mean obtained for all possible routes, the process value on the causative side with respect to the fluctuation of the process value on the result side is calculated. It analyzes the degree of the impact of. Then, this analysis includes obtaining the maximum value of the geometric mean obtained for all possible routes and extracting the route showing this maximum value.

That is, when the above method of the present invention is expressed by a general formula, it can be expressed by the following formula. Ra = (γa ₁ · fa ₁ × γa ₂ · fa ₂ × ... × γan · fan) ^ (1 / n) Rb = (γb ₁ · fb ₁ × γb ₂ · fb ₂ × ... × γbm · fbm) ^ (1 / m) …………………………………………………………………… Rz = (γz ₁ · fz ₁ × γz ₂ · fz ₂ × …… × γzy Fzy) ^ (1 / y) R = MAX (Ra, Rb, ..., Rz) where γa ₁ , ...: individual cross-correlation coefficient, fa ₁ : individual weight Ra, ...: individual route Geometric mean of cross-correlation coefficients (including weighted) along, n, ...: Number of layers in each path (depth of analysis). R: Largest geometric mean (geometric mean of strongest causal route) MAX: Operator for finding the maximum value

Next, the above method of the present invention will be described with reference to FIGS. First, FIG. 4 schematically shows a causal relationship between a large number of process values in a process, and these causal relationships have a tree structure. Such a tree structure corresponds to the route in the process route, and the left side in the figure is the cause side and the right side is the result side. That is, for example, the process value Pm on the most result side in the figure fluctuates due to fluctuations in the process values Pi, Pj, Pl, and the process value Pi fluctuates due to fluctuations in the process values Pa, Pb. Therefore, the process value Pm is the process value P
It fluctuates due to fluctuations in a, Pb, ....

FIG. 5 shows the operation of the present invention in the case where a change in the process value Pm is detected in the configuration of FIG. 4, and the change in the process value Pm is detected by a detection method using a threshold value or the like. It is detected using an appropriate method. When a change in the process value Pm is detected, the process value Pm is used as the starting process value, and all process values related to the causal relationship are traced from the process value Pm to the cause side. Analysis. That is, there are eight routes to be followed: a route a from Pm to Pi to Pa, a route b from Pm to Pi to Pb, ..., A route from Pm to Pl, and a route h from Pk to Ph. It is a route.

As shown in FIG. 5, in the route a, the time-dependent measurement data of each process value is used as it is, or the time-dependent measurement data is subjected to the FFT processing as necessary, and
The cross-correlation coefficient γmi between Pm and Pi and the cross-correlation coefficient γia between Pi and Pa are calculated, and the geometric mean Ra of these is calculated. (Note that the weights are ignored in this explanation.) Similarly, in the route b, the cross-correlation coefficient γmi between Pm and Pi.
, And the cross-correlation coefficient γib between Pi and Pb is obtained, and the geometric mean Rb of these is obtained. In this way, the geometric mean of each route is obtained, and in route h, the cross-correlation coefficient γml between Pm and Pl, the cross-correlation coefficient γlk between Pl and Pk, and Pk
Then, a cross-correlation coefficient γkh between P and Ph is obtained, and a geometric mean Rh of these is obtained.

The geometric mean Ra, Rb,
.., Rh are not affected by the number of hierarchies, and each represents the strength of the causal relationship with the process value Pm at the same level. It is possible to extract the route that causes the fluctuation of Pm. For example, if there is only one route with the largest geometric mean value, it can be extracted as the cause of the fluctuation, and the value of each cross-correlation coefficient along this route can be examined. It is also possible to extract the process value that causes it. Further, even when there are a plurality of routes having a large geometric mean value, it is possible to extract the process value that causes the cause by examining the value of the cross-correlation function along each route.

In the present invention described above, the process value on the cause side and the process value on the result side for obtaining the cross-correlation coefficient are
Not only the same type, but also different types may be used.

[0025]

As described above, the present invention has the following effects. Significant process variability can be associated with a large number of causal related process values and normalized for all possible paths of the tree structure to obtain a systematic analysis result. Since the analysis results are normalized, it is possible to easily compare the strengths of the causal relationships among the paths even if the depth of analysis or the number of layers in the tree structure is different. It is possible to flexibly deal with the difference in strength and weakness of the causal relationship due to the configuration of the process by using the weight multiplied by the cross-correlation coefficient. Since the analysis results described above can be expressed as a systematic formula, conventional individual knowledge expressions can be aggregated as a knowledge group, for example, about 1/3 of the conventional IF-THEN rule.
It becomes possible to express in degree.

[Brief description of drawings]

FIG. 1 is a system diagram schematically showing a configuration of a process to which the present invention is applied.

FIG. 2 is an explanatory diagram showing a part of an example of the operation of the present invention.

FIG. 3 is an explanatory diagram showing a part of the rest of the example of the operation of the present invention.

FIG. 4 is an explanatory diagram schematically showing a causal relationship between a large number of process values in a process to which the present invention is applied.

FIG. 5 is an explanatory diagram schematically showing a part of the arithmetic operation of the present invention.

[Explanation of symbols]

 A, B, C, D, E, F Path F Flow rate measurement part P Pressure measurement part R Controller M Heat quantity controller 1b, 1c Orifice 2a, 2b, 2c Flow rate control valve

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G06F 17/15 G06F 15/336

Claims (9)

[Claims] 1. In a process having a large number of process values that are causally related to each other in a tree structure, the process values are sequentially traced along one path from the result side to the cause side and every two adjacent process values. The cross-correlation coefficient of each time-dependent measurement data is obtained, and the geometric mean of the cross-correlation coefficients along the above route is obtained. Analysis method for process fluctuations, characterized by analyzing the extent of the influence of the process value on the causative side on the fluctuation of the process value 2. In a process having a large number of process values that are causally related to each other in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and for every two adjacent process values. FF for each time measurement data
T-process to obtain those cross-correlation coefficients, and also obtain a geometric mean of a plurality of cross-correlation coefficients along the path,
A method for analyzing fluctuations in a process, characterized by analyzing the extent of the influence of the process value on the causal side on the fluctuation of the process value on the result side by the magnitude of the geometric mean obtained for all possible routes. 3. A process having a large number of process values which are causally related in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and every two adjacent process values are obtained. FFT of each measured data over time if there is a non-negligible time delay between them
The cross-correlation coefficient is calculated and the geometric mean of a plurality of cross-correlation coefficients along the path is calculated. A method for analyzing fluctuations in a process, characterized by analyzing the degree of influence of a process value on the causative side with respect to fluctuations in the process value 4. In a process having a large number of process values that are causally related in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and every two adjacent process values. The cross-correlation coefficient of each time-dependent measurement data is obtained, and a plurality of cross-correlation coefficients along the above route are multiplied by a predetermined weight to obtain a geometric mean thereof, and the above is obtained for all possible routes. A method for analyzing fluctuations in a process, characterized by analyzing the degree of influence of the process value on the causal side on the fluctuation of the process value on the result side according to the size of the geometric mean. 5. In a process having a large number of process values that are causally related in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and for every two adjacent process values. FF for each time measurement data
The cross-correlation coefficient is calculated by T processing, and a plurality of cross-correlation coefficients along the path are multiplied by a predetermined weight to calculate a geometric mean of the cross-correlation coefficients. A method for analyzing fluctuations in a process, characterized by analyzing the degree of influence of the process value on the causal side on the fluctuation of the process value on the result side according to the size of the geometric mean. 6. In a process having a large number of process values that are causally related in a tree structure, the process values are sequentially traced along one path from the result side to the cause side, and for every two adjacent process values. FFT of each measured data over time if there is a non-negligible time delay between them
The cross-correlation coefficient is processed to obtain a geometric mean of the cross-correlation coefficients obtained by multiplying a plurality of cross-correlation coefficients along the route by a predetermined weight, and the above-mentioned synergistic value obtained for all possible routes. A method for analyzing fluctuations in a process, characterized by analyzing the degree of the influence of the process value on the causal side on the fluctuation of the process value on the result side according to the average size. 7. A process value in which a variation is detected by comparison with a threshold value is set as a starting process value, and the process value is traced from this process value to the cause side for analysis. To analyze fluctuations in the process of 8. The method for analyzing fluctuations in the process according to claim 1, wherein the maximum value of the geometric mean obtained for all possible routes is determined, and the route exhibiting this maximum value is extracted. 9. The method for analyzing fluctuations in the process according to claim 4, wherein the weight is set by a flow rate ratio of the fluid flowing through the path.