JP2018036726A - Data analysis device, control device, control method and control program of data analysis device, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a desired analysis result.SOLUTION: A data analysis device comprises: a data input unit (40) for obtaining time-series data; a cluster processing unit (11) for, by using the similarity between intermediate clusters, which are obtained by dividing the time-series data along a time series, found by a determination value calculated from a plurality of kinds of indexes indicating characteristics of the intermediate clusters, integrating the intermediate clusters and thereby dividing the time-series data into a predetermined number of final clusters; and an output unit (50) for outputting a result of division.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a data analysis apparatus that divides time-series data into a plurality of clusters along a time series.

  Conventionally, in order to confirm the processing status in a plant, a waste incineration facility, or the like, time-series data taken out from a sensor provided in the facility or the like has been analyzed.

  For example, in a waste incineration facility, steam is produced from waste heat generated by burning garbage, and electricity is produced by sending this steam to a steam turbine generator. Although it is desirable to continue supplying these electricity at a stable output, in reality, since the amount of heat and the amount of steam fluctuate due to uneven combustion of dust, it is difficult to output stably as it is. Therefore, it is important to appropriately control the combustion of garbage.

  Currently, the combustion control of waste is performed by automatic combustion control (ACC), but the operator visually checks the amount of steam, etc., and if necessary, the amount of dust supplied (the amount of dust supplied to the furnace), combustion air The amount is manually operated. In the automatic combustion control, threshold values for upper and lower limits of numerical values of the dust supply amount and the combustion air amount are provided, and control is performed to recover the target value in the event of an abnormality.

  However, the operation by the operator varies depending on the skill of the operator, and fine control cannot be performed by the control with the threshold value uniformly determined.

  Therefore, in order to realize fine control, it is necessary to appropriately analyze the time series data of the detected steam amount.

  As a method for analyzing time series data, for example, there are techniques described in Patent Documents 1 and 2. Patent Document 1 describes a diagnostic apparatus that normalizes time-series data and performs analysis such as principal component analysis and principal component regression on the normalized data.

  Patent Document 2 discloses a method of displaying a trend graph by adding an arbitrary offset value for each data in displaying a trend graph.

JP 2012-141712 A (published July 28, 2012) Japanese Patent Laying-Open No. 2005-351832 (released on December 22, 2005)

Motohide Uno, Takahiro Iwasa, Katsumi Takahashi, “Expression of Time-Series Data Based on Words Based on Global Trends and Local Features—Determining Fuzzy Sets of Periods Based on Data Trends”, 30th Fuzzy System Symposium, Lecture Proceedings pp808-813, 2014.9.1 Motohide Uno, Kazuhisa Seta, Katsumi Takahashi, “Expression of Time Series Data Based on Words Based on Global Tendency and Local Features: Determination of Fuzzy Sets of Periods”, 25th Fuzzy System Symposium, Proceedings, 2009.7. 14 Motohide Uno, Katsumi Takahashi, Kazuhisa Seta, “Expression of Time Series Data Based on Words Based on Global Tendency and Local Features-Determination of Fuzzy Sets of Periods”, 25th Fuzzy System Symposium, Proceedings, 2009.7. 14 Motohide Uno, Mitsuhiro Okamura, Kazuhisa Seta, "Expression of time-series data in words based on global trends and local features-Representation of vibrations by expanding the number of changes", 24th Fuzzy System Symposium, Proceedings pp774 -779, 2008.9.3

  However, conventional analysis methods such as principal component analysis and principal component regression are generally premised on static phenomena and are not suitable for complex data that considers temporal factors. Further, even in time series analysis considering time factors, there is a possibility that the model will not fit with the passage of time, and the analysis result desired by the user may not be obtained.

  The data analysis method described in Patent Document 1 is the same as the conventional analysis method described above, and cannot solve the above-described problems. The technique described in Patent Document 2 is a trend graph display method, and does not mention a data analysis method.

  The present invention has been made in view of the above-described problems, and an object thereof is to realize a data analysis apparatus and the like that can obtain an analysis result desired by a user.

  In order to solve the above problems, a data analysis apparatus according to an aspect of the present invention is a data analysis apparatus that analyzes time-series data used for process control, and a data acquisition unit that acquires the time-series data; , By integrating the intermediate clusters using the similarity between the intermediate clusters obtained from the determination values calculated from a plurality of types of indicators indicating the characteristics of the intermediate cluster obtained by dividing the time series data along the time series, A cluster dividing unit that divides the time-series data into a predetermined number of final clusters and an output unit that outputs a result of the division are provided.

  According to the above configuration, the determination value is calculated using a plurality of types of indexes indicating the characteristics of the intermediate cluster. Therefore, the similarity can be determined according to the characteristics of the intermediate clusters, and the intermediate clusters can be integrated. Thereby, similar intermediate clusters can be integrated and divided into time-series data final clusters.

  Moreover, since similar intermediate clusters are integrated, clustering close to human sense can be realized.

  Further, since clustering can be performed using the determination value, stable clustering can be realized as compared with the case where human clustering is performed.

  In the data analysis apparatus according to an aspect of the present invention, the cluster dividing unit may calculate the determination value by weighting the plurality of types of indices.

  According to the above-described configuration, the determination value is calculated by weighting a plurality of types of indices, so that the determination of the similarity of the intermediate cluster can be made closer to the human sense by bringing the weight closer to the human sense.

  In the data analysis device according to one aspect of the present invention, the cluster dividing unit may calculate the determination value by changing the weighting ratio according to the length of the intermediate cluster in the time series direction. Good.

  When a human determines that clusters are similar, what is considered important differs depending on the length of the cluster. And according to the said structure, the weighting of several types of parameter | index is varied with the length of the time series direction of an intermediate cluster. Therefore, it is possible to determine the similarity between the intermediate clusters closer to a human sense.

  In the data analysis device according to one aspect of the present invention, the index includes a representative value of the time series data in the intermediate cluster, a change state index indicating a change tendency of the time series data in the intermediate cluster, and the intermediate cluster. May be at least one of vibration state indicators indicating the vibration tendency of the time-series data.

  According to the above configuration, at least one of a representative value of data in the intermediate cluster, a change state index indicating a change tendency of time series data in the intermediate cluster, and a vibration state index indicating a tendency of vibration of the time series data in the intermediate cluster The determination value is calculated using Thereby, it is possible to calculate a determination value that appropriately reflects the characteristics of the intermediate cluster.

  An example of the representative value is an average value of time series data included in the intermediate cluster. In addition, as the change state index, there is an inclination with respect to the time series direction of the line segment connecting the first and last data in the time series direction in the intermediate cluster. The vibration state index includes a standard deviation of time series data included in the intermediate cluster.

  In the data analysis device according to one aspect of the present invention, the cluster dividing unit may assign the weight of the change state index to the representative among the indices when the length of the intermediate cluster in the time series direction is longer than a predetermined length. When the determination value is calculated to be larger than a value weight, and the length of the intermediate cluster in the time series direction is shorter than a predetermined length, the weight of the representative value among the indices is the weight of the change state index The determination value may be calculated with a larger value.

  The human sense is that it is easy to recognize similarity in the case of data that is long in the time series direction, whether the overall slope is similar or not, and in the case of data that is short in the time series direction, the value is similar. It is easy to recognize similarity from the point of whether or not. According to the above configuration, in the weighting of the index for calculating the determination value, when the length of the intermediate cluster in the time series direction is longer than the predetermined length, the weight of the change state index is larger than the weight of the representative value. When the length of the intermediate cluster in the time series direction is shorter than the predetermined length, the weight of the representative value is set larger than the weight of the change state index. Therefore, the similarity using the determination value can be made closer to a human sense.

  In order to solve the above-described problem, the control according to an aspect of the present invention is a control device that controls a control target based on time-series data collected from an external device, and the time-series data is subjected to the data analysis. A transmission unit that transmits to a device; a feedback control unit that performs feedback control using the characteristics of the shape of the time-series data in the final cluster that is a result of division of the time-series data output from the data analysis device; It is characterized by having.

  According to the above configuration, it is possible to perform feedback control using the characteristics of the shape of the time-series data in the final cluster divided by the data analysis device described above.

  Further, if the time series data of the final cluster is languageized, feedback control can be performed using the language.

  In order to solve the above-described problem, a control method for a data analysis apparatus according to an aspect of the present invention is a control method for a data analysis apparatus that analyzes time-series data used for process control, the time-series data being The intermediate cluster using the similarity between the intermediate clusters obtained from the data acquisition step to be acquired, and determination values calculated from a plurality of types of indices indicating the characteristics of the intermediate cluster obtained by dividing the time series data along the time series By integrating the time series data into a predetermined number of final clusters, and an output step for outputting the result of the division.

  According to the above method, the same effect as described above can be obtained.

  The data analysis apparatus according to each aspect of the present invention may be realized by a computer. In this case, the data analysis apparatus is operated on each computer by causing the computer to operate as each unit (software element) included in the data analysis apparatus. The control program for the data analysis apparatus realized by the above and the computer-readable recording medium on which the control program is recorded fall within the scope of the present invention.

  According to one aspect of the present invention, the determination value is calculated using a plurality of types of indices indicating the characteristics of the intermediate cluster. Therefore, the similarity can be determined according to the characteristics of the intermediate clusters, and the intermediate clusters can be integrated. As a result, similar intermediate clusters can be integrated and divided into time-series data final clusters. Moreover, since similar intermediate clusters are integrated, there is an effect that clustering close to a human sense can be realized. Furthermore, since clustering can be performed using the determination value, there is an effect that stable clustering can be realized as compared with the case where human clustering is performed.

It is a block diagram which shows the principal part structure of the analyzer which concerns on this embodiment. It is a block diagram which shows the principal part structure of the PID control apparatus which concerns on this embodiment. It is a flowchart figure which shows the flow of the process in the said analysis apparatus. It is a flowchart figure which shows the flow of control by the said PID control apparatus. It is a figure for demonstrating the method of calculating a judgment value, (a) is a figure for demonstrating the value of a cluster, (b) is a figure for demonstrating the variation | change_quantity of a cluster, (c ) Is a diagram for explaining how the cluster vibrates. It is a figure for demonstrating the method to change the ratio of weighting, (a) is a figure for demonstrating the case where a cluster width is large, (b), (c) demonstrates the case where a cluster width is small. FIG. It is a figure which shows the example of a cluster with a small cluster width. It is a figure which shows a cluster example. It is a figure for demonstrating the problem of clustering. It is a figure for demonstrating the clustering which concerns on this embodiment. It is a figure for demonstrating the example used for the PID control apparatus. It is a figure which shows the example which offset generate | occur | produces. (A)-(c) is a figure which shows the example with which the result of the verbalization of the clustered time series data mentioned above and the ratio of PID were matched. (A), (b) is a figure which shows the other example of control. (A), (b) is a figure which shows the example of the clustered time series data to which this embodiment is applied, and the example of the result of verbalization.

  Hereinafter, embodiments of the present invention will be described in detail. An analysis apparatus (data analysis apparatus) 1 according to the present embodiment is time-series data used for process control, and divides input time-series data into a plurality of clusters along the time-series direction. The analysis device 1 according to the first embodiment can be applied to various devices that output time-series data. For example, it can be used for analysis of time-series data taken out from a sensor provided in a garbage incineration facility.

  In the present embodiment, the initial cluster is obtained by dividing the time series data first, the final cluster is a cluster whose features are output as a result of the division, and the initial cluster is integrated into the final cluster. These clusters are called intermediate clusters.

[Configuration of analysis equipment]
The analysis device 1 will be described with reference to FIG. FIG. 1 is a block diagram showing a main configuration of the analysis apparatus 1. As shown in FIG. 1, the analysis apparatus 1 according to the present embodiment includes a control unit 10, a storage unit 20, a verbalization parameter setting unit 30, a data input unit (data acquisition unit) 40, and an output unit 50.

  The control unit 10 executes various processes in the analysis apparatus 1 and performs clustering of time-series data and verbalization of clustered time-series data. A cluster processing unit (cluster dividing unit) 11, feature extraction Part 12 and feature expression part 13.

  The cluster processing unit 11 performs clustering of time-series data. The initial cluster creation unit 113, the cluster value calculation unit 114, the weighting unit 115, the similarity determination unit 116, the cluster integration unit 117, and the cluster number determination Part 118.

  The initial cluster creation unit 113 creates an initial cluster by dividing the time series data input via the data input unit 40. The initial cluster is a cluster obtained by dividing time series data for each data.

  The cluster value calculation unit 114 calculates the cluster value (judgment value) of the initial cluster created by the initial cluster creation unit 113. Further, the cluster value calculation unit 114 calculates the cluster value of the intermediate cluster created by the cluster integration unit 117 described later integrating the clusters. The intermediate cluster is a cluster created by integrating the initial cluster, a cluster created by integrating the intermediate cluster, and not the final cluster. The cluster value is a value indicating the characteristics of the cluster, the average value (representative value) of the data value of the cluster, and the time series direction of the line segment connecting the first and last data in the time series direction of the cluster It is calculated from the index indicating the characteristics of the cluster, which is the inclination (change state index) with respect to and the degree of dispersion of data values (vibration state index) in the cluster. A specific calculation method will be described later.

  The weighting unit 115 weights three values used when the cluster value calculation unit 114 calculates a cluster value. A specific weighting method will be described later. The three values are the average value of the data values of the cluster, the slope of the segment connecting the first and last data in the time series direction with respect to the time series direction, and the degree of dispersion of the data values in the cluster. .

  The similarity determination unit 116 determines the similarity between clusters using the cluster value calculated by the cluster value calculation unit 114.

  The cluster integration unit 117 integrates clusters using the similarity determined by the similarity determination unit 116.

  The cluster number determination unit 118 determines whether or not the number of clusters as a result of cluster integration by the cluster integration unit 117 is equal to or less than the designated number of clusters. The designated number of clusters may be a number designated by the user, or may be a number obtained by analysis time ÷ 60.

  The feature extraction unit 12 extracts features of the time series data (final cluster) clustered by the cluster processing unit 11. Features include the overall trend of the final cluster, local time, local value, position, early vibration, middle vibration, late vibration, and the like. Details of the features will be described later.

  The feature expression unit 13 is a verbalization unit that verbalizes clustered time-series data (final cluster) using the features extracted by the feature extraction unit 12 and verbalization data 21 described later. Since the verbalization process can be performed by a known technique, the description thereof is omitted.

  The storage unit 20 is a storage unit that stores various data used for processing of the analysis apparatus 1, and includes verbalized data 21. The verbalized data 21 is used to execute a verbalization process in the feature expression unit 13.

  The verbalization parameter setting unit 30 sets various parameters used for verbalization.

  The data input unit 40 receives input of time series data to be analyzed by the analysis device 1 and transmits it to the cluster processing unit 11.

  The output unit 50 outputs the result of clustering by the cluster processing unit 11, the features extracted by the feature extraction unit 12, and the result of verbalization by the feature expression unit 13.

[Configuration of PID control device]
The output result of the analysis apparatus 1 according to the present embodiment can be used for feedback control. An example in which the output result of the analysis apparatus 1 is used for PID (Proportional-Integral-Differential) control is shown in FIG. FIG. 2 is a block diagram showing a main configuration of the PID control device (control device) 100. The PID control device 100 is a device that performs PID control by changing a PID parameter using the output result of the analysis device 1 and outputs the control result to a control target.

  As shown in FIG. 2, the PID control device 100 includes an analysis device 1, a data acquisition unit (transmission unit) 110, a storage unit 120, and a PID parameter change unit (feedback control unit) 130. The analysis device 1 is not necessarily included in the PID control device 100, and may be provided outside the PID control device 100.

  The data acquisition unit 110 acquires time-series data from an external device and transmits it to the analysis device 1.

  The storage unit 120 is a storage unit that stores various data used for processing in the PID control device 100, and includes adjustment data 121. The adjustment data 121 is used to change the PID parameter from the time series data analysis result by the analysis apparatus 1.

  The PID parameter changing unit 130 changes the PID parameter using the adjustment data 121 from the analysis result of the time series data by the analysis device 1.

  In the present embodiment, by converting the trend of the time series data into a language, for example, when the time series data is inclined, the degree can be controlled in a plurality of stages. In the case of verbalization, for example, “large increase”, “slight increase”, etc. Then, by predetermining the PID control value according to a plurality of stages, the PID control value can be obtained according to the ratio corresponding to each stage.

[Processing flow in the analyzer]
Next, with reference to FIG. 3, the flow of processing in the analysis apparatus 1 will be described. FIG. 3 is a flowchart showing the flow of processing in the analysis apparatus 1.

  As shown in FIG. 3, the analysis apparatus 1 first reads a parameter file storing various parameters used for clustering time-series data (S101). When the parameter file is successfully read, time-series data is read (S102, data acquisition step). When the time series data is successfully read, the initial cluster creation unit 113 creates the initial cluster by dividing the time series data (S103).

  Next, the cluster number determination unit 118 determines whether or not the current number of clusters is larger than the specified number of clusters (current number of clusters ≧ specified number of clusters) (S104). Further, it is determined whether there is a cluster whose length in the time series direction (hereinafter also referred to as cluster width) is shorter than a reference value (S104). If the current number of clusters is greater than the specified number of clusters, or if the cluster width is shorter than the reference value (YES in S104), the similarity determination unit 116 determines whether the clusters between adjacent clusters in the time series direction. The similarity is calculated (S105). Next, the similarity determination unit 116 calculates the similarity between the cluster and the cluster further adjacent to the cluster adjacent in the time series direction (S106). In other words, the similarity between clusters skipped by one in the time series direction is calculated.

  Next, the cluster integration unit 117 integrates the clusters having the highest similarity among the similarities calculated in steps S105 and S106 (S107). If the similarity calculated in step S106 is the highest, in addition to the clusters for which the similarity is determined, the clusters sandwiched between the clusters are also integrated. Steps S103 to S107 are cluster division steps.

  Then, it returns to step S104 and it is determined whether the number of clusters after integration is larger than the designated number of clusters (S104). Further, it is determined whether there is a cluster width shorter than the reference value (S104).

  In step S104, if the number of clusters after integration is equal to or less than the specified number of clusters and no cluster width is shorter than the reference value (NO in S104), the feature extraction unit 12 is the cluster after integration. The feature of the time-series data of the final cluster is extracted, and the feature expression unit 13 verbalizes the extracted feature (S108).

  Thereafter, the analysis apparatus 1 outputs the characteristics of the final cluster, which are analysis results, and the verbalization results from the output unit 50 (S109, output step). The output may be a graph display or a character output. A graph or the like can be displayed, for example, with GIF (Graphic Interchange Format) data. Moreover, the output by a character can be performed by the data of CSV (Comma Separated Value) format, for example.

  When the above-described control by the PID control device 100 is performed, the final cluster feature extraction and verbalization results in step S108 may be output to the PID control device 100 (see FIG. 3A).

[Process flow in PID control device]
With reference to FIG. 4, the flow of control by the PID control device 100 will be described. FIG. 4 is a flowchart showing a flow of control by the PID control device 100.

  As illustrated in FIG. 4, the PID control apparatus 100 acquires the time-series data characteristics, verbalization results, and time-series data of the final cluster processed in step S <b> 108 of FIG. 3. The PID parameter changing unit 130 determines whether or not there is a change in tendency from the acquired time series data (S201). If it is determined that the trend has changed (YES in S201), the PID parameter changing unit 130 adjusts the PID parameter using the adjustment data 121 from the characteristics of the time-series data after the trend change and the result of verbalization. (S202). An example of adjusting the PID parameter is to change the PID ratio.

[Details of cluster processing section]
Next, details of the processing in the cluster processing unit 11 will be described with reference to FIGS. First, a method for calculating a determination value for determining the similarity between clusters will be described with reference to FIG. The determination value is calculated using three elements: a cluster value, a cluster change amount, and a cluster vibration level. FIG. 5 is a diagram for explaining a method for calculating a determination value, (a) is a diagram for explaining a cluster value, and (b) is a diagram for explaining a change amount of the cluster. And (c) is a diagram for explaining the vibration state of the cluster.

[Cluster value]
As shown in FIG. 5A, consider a case where time-series data is divided into intermediate clusters C501, C502, and C503. When the value of the time series data in the intermediate cluster C501 is (60, 61, 53, 70), the average value is (60 + 61 + 53 + 70) / 4 = 61. In this embodiment, the value of the intermediate cluster C501 in this case is “61”, which is an average value. By calculating in the same manner, the value of the intermediate cluster C502 becomes “80” and the value of the intermediate cluster C503 becomes “72”.

  Here, when the similarity between clusters is determined using only the value of the intermediate cluster, attention is paid to the difference between the values of the intermediate clusters C501, C502, and C503. The difference between the values of the intermediate clusters C501 and C502 is | 80−61 | = 19. Further, the difference between the values of the intermediate clusters C502 and C503 is | 71-80 | = 9. Therefore, the intermediate clusters C502 and C503 are more similar than the intermediate clusters C501 and C502.

[Change in cluster]
The change amount of the cluster is calculated from the inclination θ (c) of the time series data in the cluster. Specifically, when the first value in the time series direction in the cluster is (t _i , x (t _i )) and the last value is (t _k , x (t _k )), the following formula ( Obtained in 1).
θ (c) = arctan ((x (t _k ) −x (t _i )) / (t _k −t _i )) (1)
Specifically, when applied to the intermediate clusters C501, C502, and C503, the change amounts are “52”, “−20”, and “−14”, respectively ((b) of FIG. 5).

  Here, when similarity is determined using the amount of change in the intermediate cluster, attention is paid to the difference in the amount of change between the intermediate clusters C501, C502, and C503. The difference in the change amount between the intermediate clusters C501 and C502 is | 52 − (− 20) | = 72. Further, the difference in change amount between the intermediate clusters C502 and C503 is | (−20) − (− 14) | = 6. Therefore, the intermediate clusters C502 and C503 are more similar than the intermediate clusters C501 and C502.

[How the cluster vibrates]
The state of vibration of the cluster is the degree of dispersion of time-series data values in the cluster. Specifically, the calculation is performed as follows. The following formula is calculated using the standard deviation of values.

First, when there are two pieces of time-series data existing in a cluster and the respective values are x _i and x _{i + 1} , the vibration condition σ (c) of the cluster is obtained by the following calculation formula (2).

Here, m = (x _i + x _{i + 1} ) / 2.

Further, when there are three or more time-series data existing in the cluster, the first value in the time-series direction is x _i , and the last value is x _{i + k} , the vibration condition σ (c) of the cluster is Obtained by calculation formula (3).

here,

It is.

Then, the similarity d _σ (c, c ′) between the clusters of the intermediate clusters c and c ′ is obtained by the following calculation formula.
d _σ (c, c ′) = | σ (c) −σ (c ′) |
This is because the degree of similarity is calculated using the absolute value of the difference in the degree of vibration of each cluster in the same manner as the method for calculating the degree of similarity between the clusters from the cluster value and the amount of change in the cluster. Become.

  Note that the standard deviation is used in the calculation method described above, but the standard condition may not be used, and the vibration state of the cluster may be calculated using the sum of differences from the center of vibration.

  This will be specifically described with reference to FIG. In FIG. 5C, an object whose vibration condition is calculated is an intermediate cluster Ct. The intermediate cluster Ct includes seven values as time series data.

  When the time series data for which the difference from the center of vibration is to be obtained is P511, first, the value of the center of vibration in P511 is obtained (AV511 in FIG. 5C). Then, a difference d511 between the center value AV511 of the vibration and the value of the time series data P511 is obtained. The same calculation is performed for all the time-series data included in the intermediate cluster Ct, and the sum of the differences d is used as the vibration condition of the intermediate cluster Ct. In FIG. 5C, a line segment connecting the center values of vibrations is indicated by a broken line.

  In addition, you may perform the calculation of a vibration condition using the increase / decrease frequency. Specifically, “Increase from increase” and “Decrease from decrease” are 0 times, “Increase from decrease”, and “Increase from decrease” are once, “Increase from decrease or constant”, and “Increase from constant” Or “decrease” may be 0.5 times, and the sum of the number of times may be obtained, and this value may be used as a vibration condition. Further, instead of simply obtaining the sum, the sum may be obtained by correcting the number of times according to the angle formed by the adjacent time-series data.

[Weighting method]
Next, a method for obtaining a determination value by weighting a difference in values between clusters, a difference in change amount, and a difference in vibration level will be described. In the present embodiment, the determination value is obtained using the following formula using the weights W ₁ , W ₂ , and W ₃ . Note that W ₁ + W ₂ + W ₃ = 1.
Determination value = W ₁ × value difference + W ₂ × change amount difference + W ₃ × vibration difference The smaller the determination value, the higher the similarity determination unit 116 determines that the similarity between the clusters is higher. To do.

Further, the weighting unit 15 changes the ratio of the weights W ₁ , W ₂ , and W ₃ according to the length of the cluster in the time series direction (cluster width). This is because the importance of “value difference” and “difference in change” varies depending on whether the cluster width is small or large.

  A method for changing the weighting ratio will be described with reference to FIG. FIG. 6 is a diagram for explaining a method of changing the weighting ratio, (a) is a diagram for explaining a case where the cluster width is large, and (b) and (c) are small cluster widths. It is a figure for demonstrating a case.

  As shown in FIG. 6A, when the cluster width is large, weighting is performed so that the change amount is more important than the cluster value.

  In addition, when the difference in the amount of change between adjacent clusters is close to each other, the value tends to be different between adjacent clusters because there are many cases of decreasing or increasing. Therefore, in this case, the weight of the value difference is reduced.

  On the other hand, as shown in FIG. 6B, when the cluster width is small, weighting is performed with more emphasis on the value than the amount of change. However, in the example shown in FIG. 6C, the following problem occurs.

  The portion that vibrates from t = 5 to t = 10 in (c) of FIG. 6 can be integrated by increasing the similarity by weighting the cluster value as described above. However, in the above-described method, the clusters cannot be integrated when the difference in the amount of change in the clusters is large and the difference in the values is small, such as t = 0 to t = 5.

  Therefore, in the example of time-series data as shown in FIG. 6C, the cluster values are set so that both t = 0 to t = 5 and t = 5 to t = 10 can be integrated. Of the differences and the difference between the change amounts, the smaller one is weighted. In the case of the same, weighting is made with emphasis on the value difference. That is, weighting is performed by an ordered weighted average (OWA).

[Processing a cluster with a small cluster width]
When cluster integration is advanced by the above-described method, a cluster having a small cluster width as shown in FIG. 7 may be created. In the present embodiment, in order to facilitate integration of clusters having a small cluster width, the determination value is corrected by the following calculation formula (4).
d _ALL (c _i , c _{i + 1} ) = d _all (c _i , c _{i + 1} ) × (1 + k × l / (t _n −t ₁ )) (4)
Here, k is a parameter and is a positive value. T _n -t ₁ is the length of the entire period of the time series data. Further, l is the width of the smaller of the cluster _{c i} width l _{(c i)} and the cluster _{c i + 1} of the width l of _{(c i + 1).} Note that the larger width L may be used instead of l.

According to the above formula (4), when the cluster width is 0 (l = 0),
d _All (c _i , c _{i + 1} ) = d _all (c _i , c _{i + 1} )
It becomes.

When the cluster width is the largest (l = t _n −t ₁ ),
d _All (c _i , c _{i + 1} ) = d _all (c _i , c _{i + 1} ) × (1 + k)
It becomes. As a result, a cluster with a large cluster width has a larger correction determination value d _All than a cluster with a small cluster, and therefore, a cluster with a small cluster width is easily grouped. When the value of k is increased, clusters having a small width are more easily collected. An example of k is k = 5.

[Processing non-adjacent clusters]
Further, as the cluster integration proceeds, a cluster as shown in FIG. 8 may be created. FIG. 8 is a diagram illustrating an example of a cluster. In the cluster shown in FIG. 8, the adjacent cluster C801 and the cluster C802 are not similar, but the cluster C801 and the cluster C803 sandwiching the cluster C802 are similar. In such a case, it is desired to integrate the cluster C801 to the cluster C803 as one cluster.

Therefore, in this embodiment, the similarity between clusters that exceed one adjacent cluster is also determined. Specifically, the determination value is calculated by the following calculation formula (5).
d _ALL (c _i , c _{i + 2} ) = d _ALL (c _i , c _{i + 2} ) × (1 + l (c _{i + 1} ) / min (l (c _i ), l (c _{i + 2} ))) (5)
Note that when the cluster c _{i + 1} sandwiched is small compared to the cluster c _i and the cluster c _i, so should integrate three clusters, in the equation (5), the cluster c _{i + 1} of the width l Correction is performed using (c _{i + 1} ).

Further, min (minimum value) in the above calculation formula (5) may be max (maximum value) or ave (average value). The width of the cluster c _{i + 1} may be anywhere between width of the cluster c _i of the cluster c _i.

[Verbalization / normalization]
The time series data clustered by the above-described method is verbalized by the feature expression unit 13. The clustered time-series data that is verbalized in this embodiment remains in a clustered aspect ratio. This point will be described with reference to FIGS. FIG. 9 is a diagram for explaining a problem of clustering. FIG. 10 is a diagram for explaining clustering according to the present embodiment.

  When time-series data is clustered, clusters with various cluster widths are generated. If the time-series data of these clusters is made to correspond to the aspect ratio of the graph, there is a possibility that time-series data having a shape different from the desired shape will be created.

  In the time-series data shown in FIG. 9, when the time-series data of the cluster C901 portion is extracted and the aspect ratio is made to correspond to the aspect ratio of the graph, that is, when vertical: horizontal = a: A, it is stretched horizontally. Time series data will be created. In such time-series data, since the shape is different from the desired one, appropriate verbalization cannot be performed. In particular, in order to verbalize clustered time-series data with a tendency for human judgment, it is necessary to fully consider how the time-series data looks.

  In the present embodiment, as shown in FIG. 10, the clustered time-series data is verbalized with the aspect ratio being clustered. That is, the aspect ratio of the time series data of the cluster C1001 is b: a, the aspect ratio of the time series data of the cluster C10021 is c: a, the aspect ratio of the time series data of the cluster C1003 is d: a, and so on. Cluster time-series data as follows. Thereby, the clustered time-series data can be appropriately verbalized.

Specifically, in the present embodiment, the absolute value of the horizontal axis width is X, the horizontal axis scale width (for example, s: second) is X ′, the vertical axis width absolute value is Y, and the scale width (for example, , Cm) is Y ′. Also, the angle between the time series data and the vertical axis is A ′, and the slope of A ′ time series data is A, and the slope A is calculated using the following calculation formula (6).
A = ((Y / Y ′) / (X / X ′)) × A ′ (6)
For example, in the example of the time series data shown in FIG. 10, the slope Ab of the time series data in the cluster C1001, the slope Ac of the cluster C1002, and so on are similarly calculated. Then, by calculating the ratio between the horizontal axis and the vertical axis from the slopes Ab to Ah, verbalization is performed while maintaining the aspect ratio of the time series data.

[Application to PID control]
Next, an example in which the analysis result of the analysis device 1 is used in the PID control device 100 will be described with reference to FIGS. In PID control, a time point (trend change point) to be controlled is found using the characteristics of clustered time-series data. Then, when it is determined that the tendency of the pv (Process Variable) value (measured value) has changed, the preset PID control is adjusted. Note that the tendency may be determined using the result of verbalizing the clustered time-series data, and the PID control may be adjusted.

  FIG. 11 is a diagram for explaining an example used in the PID control device. In the example shown in FIG. 11, pv value, sv (Set Variable) value (target value), and mv (Manipulative Variable) value (operation amount) are shown.

  In the conventional PID control, the output of the PID is adjusted so as to be symmetrical with the graph of the pv value and the x axis. Specifically, PID control is performed in such a direction that the difference between the sv value and the pv value becomes smaller. In this case, for example, as shown in FIG. 11A, when the pv value decreases with a constant slope from the sv value, if the conventional adjustment as described above is performed, the pv value is excessive (see FIG. 11). (A) overshoot of the pv value immediately after the 1101 portion) occurs.

  Therefore, for example, when PID control is used for combustion control of a waste incinerator in a waste incineration facility, for example, an evaporation amount pv value of waste heat steam (generated by a steam turbine) generated by burning the waste is an evaporation amount sv value. Combustion control is performed when it has decreased. Specifically, in the control device, when it is confirmed by the in-furnace camera that the burnout point of the garbage is in front of the inside of the furnace, the speed of the dust supply device is reduced. In addition, when it is confirmed that there is no dust, the speed of the dust feeder is increased. By performing such control, the evaporation amount (dust combustion) can be stabilized.

  That is, according to the above configuration, it is possible to accurately grasp the trend of the time series data in real time by the feature of the clustered time series data of the steam amount, and execute the appropriate control as described above. it can. As a result, fine control can be performed without human intervention.

  Further, when the mv value as time series data is diverging as in the example shown in FIG. 11B, the conventional PID control has been adjusted so as to cancel the vibration. However, when adjustment is performed by such a conventional method, there is a possibility that a phase shift occurs and adjustment cannot be performed properly. Furthermore, the deviation may call for further deviation, resulting in more intense vibration. This is because PID control has a demerit that control delay occurs.

  As described in the present embodiment, the above-described problems can be prevented by performing adjustment using clustered time-series data. For example, by storing the past waveform shapes of the time-series data, searching for waveforms having similar shapes in the measurement data, and starting adjustment, control without delay can be performed. Therefore, it is possible to suppress inappropriate control that causes the operation amount to diverge more than necessary.

  Further, when PID control is performed, an offset as shown in FIG. 12 may occur. In this case, conventionally, an alarm such as “large standard deviation” is output to notify the worker and adjust manually.

  As described in the present embodiment, the above-described problems can be prevented by performing adjustment using clustered time-series data. For example, the offset shape can be stored by searching for a similar shape in the measurement data and adjusting (amplifying the integral gain (I)).

  Further, according to the present embodiment, it is possible to determine “not offset”. Automatic adjustment is possible by setting a reference in advance.

  Generally, when an offset occurs, it is dealt with by increasing the integral gain (I). However, if it is too large, vibration may occur. In the conventional technique, even if the offset is found, the integral gain (I) is adjusted manually. In addition, humans also check the adjustment results.

  According to the present embodiment, even if there is a problem with the adjustment value of the integral gain (I) that has been set in advance, it is possible to obtain an optimum output by performing automatic adjustment.

[Automatic adjustment of PID control]
FIG. 13 is a diagram illustrating an example in which the verbalization result of the clustered time-series data described above is associated with the PID ratio. In this way, automatic adjustment of PID control is possible by preparing a table in which the verbalization result and the ratio of PID as adjustment contents are associated in advance. In the following, P ratio, I ratio, and D ratio are ratios of proportional operation (P), integration operation (I), and differentiation operation (D) in PID control.

  FIG. 13A shows an example of the relationship between the verbalized slope and the PID ratio. For example, if the result of verbalization is “slight decrease” 60% and “medium decrease” 40%, the P ratio of PID is A × 60% + D × 40%, and the I ratio is B × 60% + E × 40%, and D ratio is C × 60% + F × 40%.

  FIG. 13B shows an example including the relationship with the operation amount in addition to the relationship between the verbalized inclination and the ratio of the PID. For example, as a result of verbalization, if “slight decrease” is 60% and “medium decrease” is 40%, the operation amount I ′ is 60% and II ′ is 40%.

  That is, in this embodiment, the degree of similarity with respect to each classification of verbalization is calculated for certain time-series data, in other words, membership values (contribution degrees) are calculated for a plurality of classifications (languages). This is the verbalized output value.

  Note that it may be determined by comparing the pv value with the sv value that the evaporation amount pv value has deviated from the sv value, or the time series data may be clustered with the pv value alone. Further, not only the shape of the pv value but also the accuracy of adjustment can be improved by using a plurality of judgment criteria such as looking at the in-furnace camera. FIG. 13C shows an example in which the shape (inclination) of the pv value, the in-furnace image, and the operation content (classification content) are associated with each other. In the example shown in (c) of FIG. 13, when the shape of the pv value is “medium decrease 80%, almost constant 20%” and the in-furnace image is “the burning point is near”, “the speed of the dust supply device” Adjust “lower PV”.

  The following adjustment methods can also be considered. As shown in FIG. 14A, before the control adjustment, when the amplitude of the time series data exceeds a predetermined value, automatic adjustment is performed to change the PID parameter or the operation amount. By repeating this, it is possible to always set an appropriate PID parameter or operation amount.

  Furthermore, as shown in FIG. 14B, when the wavelength is long, automatic adjustment is performed, and the wavelength can be shortened by changing the PID parameter or the operation amount.

  Adjustment may be performed automatically and evaluation may be performed automatically. The evaluation method is performed based on how far the time series data measured after the control is from the set value (or how much the operation amount converges). If the measured value is farther than the set value (this may also be determined fuzzically), it is determined that the PID ratio is not appropriate, and thereafter the PID control may not be executed in the same case. .

  Moreover, the structure which performs a learning process may be sufficient. That is, when the control result is determined to be appropriate, the data may be accumulated, and when the control result is determined to be inappropriate, the control content may be deleted.

[Parameter setting example]
The parameter can be specified and set by the user. For example, the user can set the following values as parameters.
Maximum value of the vertical axis of the graph = 50.0
Minimum value on the vertical axis of the graph = 0.0
graoph_r: Aspect ratio of aspect ratio graph when displaying time series data (fixed) = 4.89361702
Fuzzy set parameter representing cluster size a = 0.1
Fuzzy set parameter b = 0.2 representing cluster size
Parameter to apply correction with cluster size k = 100.0
w1 (when the average is similar and the cluster is small) = 1.00 w1 (when the average is similar and the cluster is large) = 0.2
w2 (when the average is similar and the cluster is small) = 0.00 w2 (when the average is similar and the cluster is large) = 0.5
w3 (when the average is similar and the cluster is small) = 0.00 w3 (when the average is similar and the cluster is large) = 0.3
w1 (when the angle is similar and the cluster is small) = 0.00 w1 (when the angle is similar and the cluster is large) = 0.2
w2 (when the angle is similar and the cluster is small) = 1.00 w2 (when the angle is similar and the cluster is large) = 0.5
w3 (when the angle is similar and the cluster is small) = 0.00 w3 (when the angle is similar and the cluster is large) = 0.3
Final minimum cluster width = 60
[Application example]
FIG. 15 is a diagram illustrating an example of clustered time-series data to which the present embodiment is applied and an example of verbalization results.

  As shown in FIG. 15A, in this embodiment, time-series data is automatically clustered at portions having similar trends. In the example shown in FIG. 15A, the cluster is formed into four clusters, that is, cluster C001, cluster C002, cluster C003, and cluster C004.

FIG. 15B shows the result of verbalization of clustered time-series data and the result of extracted features. In addition, the meaning of each term shown in (b) of FIG. 15 is as follows.
Overall trend: The overall trend of time-series data (graphs) of clustered sections.
Local time: Each cluster is divided into three in a fuzzy manner to indicate the position in the section where the point is farthest from the overall trend graph.
Local value: Shows the overall trend of the value at the local time.
Position: Indicates where the displayed graph is located.
First period of vibration: The section is simply divided into three parts to indicate how much vibration is generated in the first part.
Middle period of vibration: The section is simply divided into three parts to show how much the medium part vibrates.
Late vibration: The section is simply divided into three parts to indicate how much vibration is occurring in the latter part.

  Here, for example, for the cluster C001, the overall trend is “almost constant” (0.873), “small decrease (slight decrease)” (0.127), and the local time is “previous period” (0.884). “Medium period” (0.116), local value is “small decrease (small decrease)” (0.698), “medium decrease (medium decrease)” (0.302), (High) ”(0.993),“ low (slightly low) ”(0.007), the first half of vibration is“ small and low (standard deviation is small, the number of vibrations is small) ”(0.654),“ No small (standard deviation is small, no vibration) "(0.188)," None is small (no standard deviation, few vibrations) "(0.123)," None without (standard deviation) No vibration (no vibration) ”(0.035) ”(0.692)”, “small and small (small standard deviation, few vibrations)” (0.303), “none without (no standard deviation, no vibration)” (0 0.003), “Nothing small (standard deviation is small, not oscillating)” (0.002), Late vibration is “Non-small (no standard deviation, few vibrations)” (0.447) ), “None without (no standard deviation and no vibration)” (0.138), “None without (standard deviation is small and no vibration)” (0.098). The numbers in parentheses indicate ratios. Hereinafter, the same applies to other clusters.

  As described above, according to the present embodiment, time series data is divided into a plurality of clusters based on the similarity of trends, and the overall tendency, local characteristics, vibration, graph position, etc. for each cluster are verbalized. Is output.

[Specific Effects According to this Embodiment]
Since time series data obtained from sensors such as garbage incineration facilities is very complex, at present, time series data is often analyzed visually. However, there is a possibility that the visual expression of time series data and the criteria for judging where the tendency has changed will change every time it is confirmed. This is because it is difficult for humans to maintain the standards, and there is a possibility that both correct judgment standards and wrong judgment standards may be forgotten.

  According to the present embodiment, time series data is clustered according to a predetermined standard, so that the same time series data is always clustered in the same way and verbalized in the same way. Moreover, since it is possible to automate the work currently performed by human eyes, it is possible to reduce processing time and labor cost.

  Conventionally, as the set value for PID control, a value along which the measured value is aligned has been used. Therefore, when the measured value deviates greatly from the set value, it cannot be adjusted properly, and human intervention and adjustment are performed manually.

According to this embodiment, since it is possible to grasp the trend (trend) of time series data,
The part where the tendency of the time-series data has changed can be determined and used for PID control. Thereby, it can adjust appropriately by PID control after the part from which the tendency changed.

  Further, when detailed control is necessary, time series data of measured values and set values is expressed by mathematical formulas, and conditions are added to DCS (Distributed Control System). However, this method is not easy because it is necessary to accurately derive calculation formulas for various time series data. Furthermore, since this method makes judgments based on threshold values, it is difficult to respond flexibly.

  According to this embodiment, since the tendency of the time series data is verbalized, it can be controlled in the same way as a human sense. Therefore, it is not necessary to accurately derive the calculation formula. Further, since the shape of the time series data itself is identified, it is possible to search for past similar patterns and grasp the current state of the time series data.

  Further, in the conventional technique, the time series data of clusters divided by parts having similar features is languageized with the aspect ratio of the graph. Therefore, the shape of the clustered time-series data does not always coincide with human senses, and it may not be possible to make a language suitable for human senses. Further, there is a method for normalizing data (for example, kernel method), but there is no method for normalizing clustered time-series data.

  In this embodiment, by normalizing the clustered time-series data, the aspect ratio of the clustered time-series data can be maintained, and languageization suitable for human senses can be performed.

  One of the techniques often used in image processing is pattern matching. This technique calculates the sum or square of the absolute value of the difference between the coordinates of each point, and determines that the larger the difference, the less similar. The disadvantage of pattern matching is that vibration and amplitude shift (when the shape is the same and the phase is shifted even at the wavelength) cannot be determined. Depending on the target data, the time width of the time series data may be shifted, and it is difficult to properly analyze the time series data by pattern matching.

  According to the present embodiment, since the tendency of the time series data is determined, it can be appropriately determined even if the period of vibration or amplitude is shifted.

  Further, there are “step response method” and “limit sensitivity method” in general automatic control adjustment (auto tuning). These methods actually input data and look at the output to adjust the PID parameters. Therefore, there is a disadvantage in that adjustment cannot be performed unless it is actually operated. In addition, auto-tuning is appropriate for simple adjustment such as temperature adjustment of an air conditioner. However, in the case of complicated data, the PID parameter is always adjusted, and control may become unstable. In particular, in a waste incineration facility, the quality of fuel is constantly changing, so time-series data tends to be complicated.

In this embodiment, since the tendency of the time series data and the control rule (PID parameter, operation amount) are stored in advance, the PID parameter and the operation amount can be adjusted from the tendency of the time series data. it can.
Further, by performing the control adjustment only when the measured value is more than a certain distance from the set value, there is no possibility that the control becomes unstable. Furthermore, even complex time-series data can be expected to be controlled close to the control considered by humans because the shape tendency of the time-series data and the control rules are stored in association with each other.

  Further, in the conventional automatic control adjustment (auto tuning), it is necessary for a human to perform auto tuning by determining that the pattern of time series data has changed.

  In this embodiment, since it can be determined that the tendency of the time-series data has been changed by the apparatus, appropriate adjustment can be automatically performed.

  One embodiment of the present invention can also be expressed as follows.

  In the data analysis device according to an aspect of the present invention, the cluster dividing unit may perform the integration by using a similarity between the intermediate clusters adjacent in the time series direction.

  According to the above configuration, adjacent intermediate clusters can be integrated.

  In the data analysis device according to one aspect of the present invention, the cluster dividing unit uses the similarity between the target intermediate cluster and the intermediate cluster further adjacent to the intermediate cluster adjacent in the time series direction, and The three intermediate clusters may be integrated.

  According to the above configuration, it is possible to integrate the intermediate clusters having the intermediate clusters sandwiched therebetween, including the intermediate clusters sandwiched therebetween. Thereby, intermediate clusters that are not adjacent but are similar can be appropriately integrated.

  In the data analysis device according to an aspect of the present invention, the cluster dividing unit may integrate the intermediate clusters using the similarity according to the length of the intermediate clusters in the time series direction.

  According to the above configuration, the degree of similarity is in accordance with the length in the time series direction. Thereby, it is possible to easily integrate intermediate clusters having a short length in the time series direction.

  The data analysis apparatus according to an aspect of the present invention includes a feature expression unit that applies the classification content indicating the feature of the shape to the shape of the time-series data in the final cluster while maintaining the aspect ratio at the time of division. It may be.

  According to the above configuration, since the classification content indicating the characteristics of the shape of the time series data in the final cluster while maintaining the aspect ratio at the time of division is applied, the aspect ratio of the graph showing the time series data as in the past Compared with the case where the classification content indicating the feature of the shape is applied, it can be applied to the content indicating the feature of the shape more appropriately.

  In the data analysis device according to one aspect of the present invention, the feature expression unit may apply the shape of the time series data to the plurality of classification contents and calculate a contribution degree to each classification content. .

  According to the above configuration, the contribution degree (membership value) can be calculated from the shape of the time series data.

  In the data analysis apparatus according to an aspect of the present invention, the feature expression unit may be a language that applies a language indicating the feature of the shape to the shape of the time-series data.

  According to the above configuration, the shape of the time series data can be verbalized.

  In the control device according to one aspect of the present invention, the control device performs PID (Proportional-Integral-Differential) control, and the feedback control unit includes a proportional operation (P) and an integral operation (PID control). I), the ratio of the differential operation (D), and the operation amount may be controlled.

  According to the above configuration, the proportional operation (P), integration operation (I), and differentiation operation (D) in the PID control using the characteristics of the shape of the time-series data in the final cluster divided by the data analysis device described above. The ratio and the operation amount can be controlled.

[Example of software implementation]
Control blocks of the analysis apparatus 1 and the PID control apparatus 100 (particularly the cluster processing unit 11 (initial cluster creation unit 113, cluster value calculation unit 114, weighting unit 115, similarity determination unit 116, cluster integration unit 117, cluster number determination unit 118) ), The feature extraction unit 12, the feature expression unit 13, and the PID parameter change unit 130) may be realized by a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like, or a CPU (Central Processing). Unit) and may be realized by software.

  In the latter case, the analysis apparatus 1 and the PID control apparatus 100 include a CPU that executes instructions of a program that is software that realizes each function, and a ROM (in which the program and various data are recorded so as to be readable by a computer (or CPU)). A Read Only Memory) or a storage device (these are referred to as “recording media”), a RAM (Random Access Memory) for expanding the program, and the like are provided. And the objective of this invention is achieved when a computer (or CPU) reads the said program from the said recording medium and runs it. As the recording medium, a “non-temporary tangible medium” such as a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. The program may be supplied to the computer via an arbitrary transmission medium (such as a communication network or a broadcast wave) that can transmit the program. The present invention can also be realized in the form of a data signal embedded in a carrier wave in which the program is embodied by electronic transmission.

  The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope shown in the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments. Is also included in the technical scope of the present invention.

1 Analysis device (data analysis device)
10 control unit 11 cluster processing unit (cluster division unit)
12 feature extraction unit 13 feature expression unit 20, 120 storage unit 21 verbalized data 30 verbalization parameter setting unit 40 data input unit (data acquisition unit)
50 output unit 100 PID control device (control device)
110 Data acquisition unit (transmission unit)
113 Initial cluster creation unit 114 Cluster value calculation unit 116 Similarity determination unit 117 Cluster integration unit 118 Cluster number determination unit 130 PID parameter change unit (feedback control unit)

Claims (9)

A data analysis device for analyzing time series data used for process control,
A data acquisition unit for acquiring the time series data;
By integrating the intermediate clusters by using the similarity between the intermediate clusters obtained from the determination values calculated from a plurality of types of indicators indicating the characteristics of the intermediate clusters divided along the time series of the time series data, A cluster divider for dividing the time series data into a predetermined number of final clusters;
And an output unit that outputs a result of the division.   The data analysis apparatus according to claim 1, wherein the cluster dividing unit calculates the determination value by weighting the plurality of types of indices.   The data analysis apparatus according to claim 2, wherein the cluster dividing unit calculates the determination value by changing the weighting ratio according to the length of the intermediate cluster in the time series direction.   The index includes a representative value of the time series data in the intermediate cluster, a change state index indicating a change tendency of the time series data in the intermediate cluster, and a vibration indicating a vibration tendency of the time series data in the intermediate cluster. The data analysis apparatus according to claim 1, wherein the data analysis apparatus is at least one of status indicators. The cluster dividing unit includes:
When the length of the intermediate cluster in the time series direction is longer than a predetermined length, the determination value is calculated by setting the weight of the change state index among the indices to be larger than the weight of the representative value,
When the length of the intermediate cluster in the time series direction is shorter than a predetermined length, the determination value is calculated by making the weight of the representative value out of the indices larger than the weight of the change state index. The data analysis apparatus according to claim 4. A control device that controls a control object based on time-series data collected from an external device,
A transmission unit that transmits the time-series data to the data analysis device according to any one of claims 1 to 5;
A feedback control unit that performs feedback control using a feature of the shape of the time-series data in the final cluster that is a result of division of the time-series data output from the data analysis device; Control device. A method for controlling a data analysis device for analyzing time series data used for process control,
A data acquisition step of acquiring the time series data;
By integrating the intermediate clusters by using the similarity between the intermediate clusters obtained from the determination values calculated from a plurality of types of indicators indicating the characteristics of the intermediate clusters divided along the time series of the time series data, A cluster dividing step of dividing the time series data into a predetermined number of final clusters;
An output step for outputting the divided result; and a control method for the data analysis apparatus.   A control program for causing a computer to function as the data analysis apparatus according to claim 1, wherein the control program causes the computer to function as the cluster dividing unit.   A computer-readable recording medium on which the control program according to claim 8 is recorded.

Images