JP5868216B2 - Clustering apparatus and clustering program - Google Patents

Description

  The present invention relates to a plant abnormality detection apparatus for detecting a sign of abnormality such as a failure of a device constituting a plant or performance deterioration.

  In power plants such as thermal power, hydropower, and nuclear power, chemical plants, steel plants, and water and sewage plants, instrumentation systems for controlling plant processes are introduced. In these plant instrumentation systems, various time-series data acquired by sensors attached to the apparatus are accumulated. There is a need to use this time-series data for plant monitoring and maintenance.

  For example, Patent Document 1 shown below describes a method of detecting anomalies by outputting the degree of detachment of observation data by calculating the similarity between sensor signal data of the past plant and observation data. Has been. However, in the plant, there are large variations in the sensor signal data of the past plant due to operation modes such as start, steady, and stop, differences in heat generation efficiency due to fuel components, and deterioration of equipment. There is a problem that it is necessary to collect signal data for each operation condition, and this load is large.

  As a method for solving the above-described problem, in Patent Document 2 shown below, based on temporal changes, time-series data is divided into a plurality of trajectory segments, and each trajectory segment is divided. A method of constantly evaluating the performance deterioration state of the equipment constituting the plant by modeling the object is described.

JP 2004-531815, “Diagnostic system and method for predictive state monitoring” JP 2010-92355 “Abnormality Detection Method and System”

  In the method of Patent Document 2, the target data is handled as another cluster if the distance between the data exceeds a predetermined threshold along the time, and is treated as the same cluster if the threshold is not exceeded. It is divided. For this reason, since the relationship between the several sensor signals resulting from operation conditions, such as starting of a plant, a steady state, and a stop, can be caught, the operation | work which collects past data for every operation condition is made unnecessary. However, when determining abnormality, it is determined by whether there are similar divisions in the past, so it is not possible to determine statistical bias such as how rare the abnormality of the division is. There is. Therefore, if the collected data contains abnormal data due to a sensor failure or data immediately before an abnormality occurs, the accuracy of abnormality detection may be reduced due to the data.

  In addition, abnormality detection based on statistical processing does not have a detection accuracy of 100%. Therefore, when plant maintenance personnel and supervisors take measures, there is a function that easily explains the basis for determining an abnormality. Desired. However, in the prior art (Patent Document 2), since the plant system is modeled by the subspace method based on the segmented trajectory, it is an abnormality determination for the result of mathematical conversion, and the relationship between the sensor signals and the abnormality There is a problem that it is difficult to explain the grounds of judgment.

  This invention improves the accuracy of abnormality detection of an apparatus for detecting a sign of abnormality such as a failure of a facility or equipment constituting a plant or performance deterioration by utilizing time series data accumulated by an instrumentation system. With the goal.

The clustering device of this invention
Time series data belonging to the time range is extracted for each N time ranges (N is an integer of 2 or more) from the first time range to the Nth time range from a plurality of different time series data. A local time-series data extraction unit that generates N local time-series data including a plurality of sets of time-series data in a range;
N local time-series data extracted by the local time-series data extraction unit is divided into initial clusters having a predetermined number of initial clusters according to an initial cluster dividing rule set in advance as an initial cluster dividing rule, and divided. N is generated by generating representative information indicating the characteristics of the initial cluster for each initial cluster, and distributing N local time-series data according to the re-clustering rule set in advance as a re-clustering rule for each generated representative information. Re-clustering is performed to divide the local time-series data into clusters, representative information is regenerated for each re-clustered cluster, and N local time-series data extraction units extracted for each re-generated representative information Re-cluster local time series data,
Similarly,
Repeat the re-clustering of N local time-series data and the regeneration of the representative information, and every time the representative information is regenerated, is the representative information generated this time changed from the representative information generated immediately before? When there is a change, the next representative information regeneration process is continued, and when there is no change, the next representative information regeneration process is continued without re-clustering and representing the N local time-series data. And a local time system data clustering unit that terminates the process of regenerating information.

  According to the present invention, detection accuracy can be improved in an apparatus for detecting a sign of abnormality such as a failure or performance deterioration of equipment or equipment constituting a plant.

1 is a block diagram showing a configuration of a plant abnormality detection device 100 according to Embodiment 1. FIG. FIG. 4 illustrates time-series data according to Embodiment 1. Explanatory drawing of local time series data extraction in Embodiment 1. FIG. Explanatory drawing which shows an example (continuous change) of the change of the correlation between the sensor signals in Embodiment 1. FIG. Explanatory drawing which shows an example (discontinuous change) of the change of the correlation between the sensor signals in Embodiment 1. FIG. Explanatory drawing which shows an example (dependency to the area of a value) of the change of the correlation between the sensor signals in Embodiment 1. FIG. Explanatory drawing which shows an example of the local time series data clustering in Embodiment 1. FIG. FIG. 3 is an explanatory diagram illustrating an example of estimation of a global time series data model in the first embodiment. 5 is a flow for explaining the overall flow of processing of the plant abnormality detection device 100 according to the first embodiment. 5 is a flow for explaining a processing flow of the local time-series data clustering unit 104 according to the first embodiment. The figure explaining the local time series data in the case of classifying by the range of the data value in the first embodiment. FIG. 3 is a diagram conceptually showing processing of a local time series data model estimation unit 103 in the first embodiment. FIG. 3 is a diagram conceptually showing “N × K” local structures S (L _ki ) in the first embodiment. FIG. 6 shows the meaning of (Formula 5) in Embodiment 1; The figure which shows notionally the process of S1004 of FIG. The figure which conceptualized the process of S1005 of FIG. The figure explaining the case where it returns to S1004 from the first S1006 of FIG. The conceptual diagram which shows after execution of k = 2 of FIG. 10, k = 3. The conceptual diagram which shows after execution of k = 1-K of FIG. The figure which shows the whole process outline | summary of FIG. FIG. 4 is a flowchart for explaining a processing flow of a global time series data model estimation unit 105 according to the first embodiment. The figure which shows an example of the external appearance of the plant abnormality detection apparatus 100 in Embodiment 2. FIG. The figure which shows the hardware structural example of the plant abnormality detection apparatus 100 in Embodiment 2. FIG.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a configuration of an example of a plant abnormality detection apparatus 100 according to the first embodiment. The functional outline of each component will be described. In the following description of the embodiments, t represents time.
(1) The plant time-series database 101 is a database that stores a plurality of time-series data obtained by sequentially observing plant equipment and the like that are targets of abnormality detection as time passes.
(2) The local time series data extraction unit 102 receives multidimensional time series data in the plant time series database 101 as input. For example, as shown in FIG. 2 to be described later, a set of a calorific value y (t) for the input fuel, a fuel input amount x1 (t), and a temperature x2 (t), (y (t), x1 (t), x2 (t )) As input. The local time series data extraction unit 102 extracts the divided time series data by classifying the input data by “time” or “data value” in accordance with the temporal change of the input data. This divided time series data is referred to as “local time series data 301”.
(3) The local time series data model estimation unit 103 estimates the model of the local time series data 301 extracted by the local time series data extraction unit 102 by multivariate analysis or a time series analysis method. Model estimation is, for example, processing for obtaining a regression equation for each “local time series data 301”.
(4) The local time series data clustering unit 104 divides the “model set” of the local time series data 301 estimated by the local time series data model estimation unit 103 into clusters, and “represents a cluster for each cluster”. Estimate (calculate) "local parameters".
(5) The global time series data model estimation unit 105 estimates a global representative time series data model by connecting the models estimated by the local time series data clustering unit 104.
(6) The outlier detection unit 106 is a set of representative local time-series data models obtained by the local time-series data clustering unit 104 or the global time-series data model estimation unit 105 with respect to separately provided segment data. A large outlier is detected as an abnormality.

(Example of multidimensional time series data)
FIG. 2 is an explanatory diagram of time-series data input by the local time-series data extraction unit 102. The time series data can be expressed as a function x (t) that associates the sensor signal value at time t with respect to time t. Time t may be continuous or may be recorded at regular intervals by sampling. In the present specification, as described above, a plurality of sets of time series data (multidimensional time series data) are handled.
FIG.
(A) Time series data y (t),
(B) Time series data x ₁ (t),
(C) Time series data x ₂ (t),
An example of three sets of time series data is shown. A set of a plurality of time series data can be regarded as a vector value function (y (t), x ₁ (t), x ₂ (t)) at time t.

(Example of time-series data classification: Part 1)
FIG. 3 is a diagram illustrating an example of time-series data segmentation by the local time-series data extraction unit 102 and an explanatory diagram illustrating an example of the operation of the local time-series data extraction unit 102.
(A) is an example of time-series data x (t) that is original data stored in the plant time-series database 101.
(B) is an example of “time-series data after smoothing processing” obtained by the local time-series data extraction unit 102 smoothing the time-series data (a).
(C) is an example of time-series data having a value obtained by taking a time difference (x (t _{i + 1} ) −x (t _i )) with respect to time-series data (b). That is, (c) indicates Δx / Δt.
(D) is an example of the local time series data 301 obtained by classifying the original signal data (a) at a time when the absolute value of the value of (c) is equal to or greater than a threshold value. In (d), the original signal data (a) is divided into eight sections. That is, (d) shows a case where the time series data x (t) of (a) is divided into eight local time series data 301.

(Example of time-series data classification: Part 2)
FIG. 4 is a diagram illustrating another example of time-series data segmentation by the local time-series data extraction unit 102, and is an explanatory diagram illustrating an example (continuous change) of a change in correlation between sensor signals. (A) and (b) are examples of time-series data. Let (a) be y (t) and (b) be x (t). (C) is an example of time-series data indicating the correlation between y (t) and x (t). In the example of (c), the correlation is regarded as a coefficient when regression analysis is performed on the time series data y (t) and x (t), and the value gradually decreases with time. When the efficiency of the equipment is gradually decreasing due to deterioration of the equipment, the value continuously changes in this way. For example, if y (t) is the amount of heat generated and x (t) is the amount of fuel input, it indicates that the efficiency of the facility is gradually decreasing due to the deterioration of the facility. The local time-series data extraction unit 102 can set the five sections in FIG. 4C as local time-series data 301, respectively.

(Example of time-series data classification: Part 3)
FIG. 5 is a diagram illustrating another example of time-series data classification by the local time-series data extraction unit 102, and is an explanatory diagram illustrating an example (discontinuous change) of a change in correlation between sensor signals. (A) and (b) are examples of time-series data. Let y (t) and x (t) respectively. (C) is an example of time-series data indicating the correlation between y (t) and x (t). In the example of (c), the correlation is regarded as a coefficient when regression analysis is performed on the time series data y (t) and x (t), and the values take two values discontinuously over time. Yes. Section 1, section 2, and section 5 have high values, and section 3, section 4, and section 6 have low values. For example, assuming that y (t) is a heat generation amount and x (t) is a fuel input amount, such a correlation is obtained when the efficiency of the facility differs depending on the type of fuel. The types of fuel are different between the group of section 1, section 2, and section 5 and the group of section 3, section 4, and section 6, and the former group has a heat generation efficiency with the same fuel amount compared to the latter group. It represents a good thing. In (c), the local time-series data extraction unit 102 includes two local time-series data, that is, the local time-series data 301 including sections 1, 2, and 5 and the local time-series data 301 including sections 3, 4, and 6. Can be classified.

(Example of dividing time-series data divided by time into a range of time-series data values)
FIG. 6 is a diagram illustrating another example of time-series data segmentation by the local time-series data extraction unit 102, and is an explanatory diagram illustrating an example of change in correlation between sensor signals (dependence of values on a section). It is. FIG. 6 illustrates a case where the time-series data divided by time is further divided by the value range of the time-series data. (A) and (b) are examples of time-series data. Let y (t) and x (t) respectively. (C) is an example of time-series data indicating the correlation between y (t) and x (t). In the example of (c), the correlation is regarded as a coefficient when regression analysis is performed on the time series data y (t) and x (t), and the change tendency of the correlation value changes with the passage of time. ing. In section 1 and section 3, the coefficient has a constant value, and in section 2, the coefficient continuously changes between the constant values. This change in correlation is not dependent on time but can be interpreted as representing dependence on the value of the time-series data x (t). For example, when y (t) is a calorific value and x (t) is a fuel input amount, even if a large amount of fuel is input by the equipment control system, it is controlled so as not to exceed a certain calorific value. If this is the case, this behavior is shown. In order to grasp the relationship between sensor signals exhibiting such behavior, it is effective to classify time series data (y (t) or x (t)) according to a range of values. (D) shows local signal data when a point where the value of the correlation value fluctuates is extracted as a category of the value y. In (d), the time-series data y is divided into value categories A, B, and C. In (d), sections 1, 2, and 3 become local time series data 301, respectively. At this time, the local time-series data extraction unit 102 further uses the time-series data value (the value of y (t) in (a) is adopted in this example) for the obtained local time-series data 301 divided by time. With respect to the local time series data 301 in section 2, the local time series data 301 is further divided into data of section 2-1 and section 2-2.

(Local and global clusters)
FIG. 7 is an explanatory diagram showing an example of local time-series data clustering. FIG. 7 is a diagram conceptually illustrating processing results of the local time series data model estimation unit 103, the local time series data clustering unit 104, and the global time series data model estimation unit 105. The graph of FIG. 7 is a scatter diagram in which the vertical axis indicates the heat generation amount y and the horizontal axis indicates the fuel input amount x.
The “scatter diagram” is a set of time series data x (t) and y (t) at a certain time mapped as a point <x (t), y (t)> on a two-dimensional graph. When there is a relationship between signals as shown in FIG. 5, a set of time series data on the scatter diagram can be classified into two clusters, a cluster 701 and a cluster 702. Hereinafter, FIG. 7 will be described.

  In FIG. 7, small clusters 703 and 704 and large clusters 701 and 702 are shown as clusters. Small clusters 703 and 704 are called local clusters (also referred to as local time series clusters), and large clusters 701 and 702 are called global clusters. The global cluster 701 corresponds to a set of x and y values when an efficient fuel is used. The global cluster 702 corresponds to a set of x and y values when using inefficient fuel. The global cluster 701 has peaked, but the global cluster 702 has a straight line. This shows an example of the relationship between signals as shown in FIG.

  The ranges (frame lines) of the local clusters 703 and 704 are examples of clusters obtained as a result of clustering the local time series data 301 by the local time series data clustering unit 104. This corresponds to the local time series data 301 divided by the time section and the value range section. In the local time series data 301, regression equations (representative local parameters) estimated by regression analysis, an AR model, or the like are regression equations 705 and 706, respectively. The abnormality determination target data 707 and 708 are examples of time series data to be determined as abnormality. In the abnormality detection, an object whose distance from the nearest regression equation is equal to or greater than a threshold is determined to be abnormal. Since 707 is close to the closest regression equation 704, it is determined to be normal. 708 is determined to be abnormal because it is more than a certain distance from the closest regression equation 704.

  FIG. 8 is an explanatory diagram showing an example of estimation of a global time series data model. Reference numerals 801, 802, 803, and 804 of (a) are representative regression equations obtained by clustering a set of local time series data 301. 809 in (b) is a global representative regression equation obtained by connecting 801, 802, 803, and 804 in (a). Similarly, 810 in (b) is a global representative regression equation obtained by connecting 805, 806, 807, and 808 in (a).

(Description of operation)
Hereinafter, the operation of the first embodiment will be described with reference to the flowcharts of FIGS. FIG. 9 is a flowchart for explaining the overall processing flow of the plant abnormality detection apparatus. FIG. 10 is a flowchart illustrating the processing flow of the local time series data clustering unit 104. FIGS. 11 to 20 are diagrams for supplementing the processing description of FIG. 10.
FIG. 21 is a flowchart for explaining the processing flow of the global time-series data model estimation unit 105.

(S901, S902: processing of local time series data extraction unit 102)
S <b> 901 is a local time series data extraction process executed by the local time series data extraction unit 102. In step S901, the input data is classified according to how the input data changes with time, using a plurality of sets of time-series data as input. The procedure for obtaining the input classification (predetermined local time-series data generation rule) is, for example, that one of a plurality of input data is an objective variable and the other is an explanatory variable.
(1) "Categorical regression analysis" described in references (Kawaguchi Zhisho, Multivariate Analysis 2pp. 60-64, Morikita Publishing)
(2) The local stationary AR model of the reference (Genjiro Kitagawa, Introduction to Time Series Analysis, pp113-124, Iwanami Shoten) is used.
(3) Alternatively, as shown in FIG. 3, the time series data may be subjected to smoothing processing and time difference processing and then the sections may be extracted with threshold values. S902 is also processing executed by the local time-series data extraction unit 102.

を入力として、区分として注目する変数ｙｉの区間の分割

により分割された時系列データの集合

を抽出する。
但し、（Ｔ_ｉｓ，，Ｔ_ｉｅ］は、Ｓ９０１で得られた区間区分とする。
以下では、Ｌ_ｋｉを局所時系列データ３０１として、局所時系列データＬ_ｋｉと呼ぶ。なお、以下では、Ｌ_ｋｉを局所時系列データと呼んで説明するが、データ区分を含まない時間範囲のみで分割したものも局所時系列データ（広義）である。 In S902, the local time-series data extraction unit 102 is a process of extracting new time-series data by further classifying the local segment time-series data obtained in S901 with a range of data values. Specifically, a set of time-series data divided by time

Divide the variable yi section of interest as a category

A set of time series data divided by

To extract.
_{However, (T is,, T ie} ] shall be obtained interval division in S901.
In the _following, the _{L ki} as a local time-series data 301, referred to as a local time-series data _{L ki.} In the following description, L _{ki is referred} to as local time-series data, but local time-series data (in a broad sense) is also divided only by a time range that does not include data sections.

With reference to FIG. 11, the meanings of the above (formula 1) to (formula 3) will be described in detail. FIG. 11 is an example of three cases of time series data y (t), time series data x ₁ (t), and time series data x ₂ (t). Hereinafter, the time series data y (t) may be described simply as y (t). Since x ₁ (t) and x ₂ (t) are two, “M = 2” in (Expression 1). (Expression 1) represents a set of N time-series data of 1 to N related to i indicating a time section, and the time-series data indicated by (Expression 1) in the case of “i = 1” is T in FIG. _1s, <t ≦ T _2e included in the range y (t), x ₁ (t), x ₂ (t)
It is.
Further, the division in the range of the data value of (Expression 2) means the division such as Y1 to Y2 and Y2 to Y3 with respect to the vertical axis of y (t) in FIG. The range of Y1 to Y2 of y (t) corresponds to the case of k = 1 in (Expression 3). That is, in (Equation 3),
Y1 <y _{i = 1} ≦ Y2
This is the case.
In the case of k = 1 and i = 1 in (Equation 3), that is, an example of L ₁₁ will be described below. When i = 1, the graph portions of y (t), x ₁ (t), and x ₂ (t) included in the time zone (vertical zone) of T _1s, <t ≦ T _{2e in} FIG. 11 correspond. To do. For k = 1 under i = 1 In addition, of the graph of y (t), the graph of y (t) of the portion belonging to the data range of Y1~Y2 is part belonging to _{L 11.} This is shown as a thick line portion of y (t) in FIG. 11 (Y left side and Y right side on both sides of the time range). The _x 1 belonging to _{L 11 (t), x 2} (t) _is determined y (t) belonging to _{L 11,} respectively 11,
x ₁ left side, x ₁ right side, x ₂ left side, x ₂ right side.
Similarly, _{L 21} in the case of k = 2, i = 1 is a graph portion shown by oblique lines shown in FIG. 11.

のすべてに対して、多変量解析、または、時系列データによりモデル推定する。
以下では、多変量解析の例として線形回帰分析を用いて説明するが、因子分析、特異値分解、ＡＲモデル、状態空間モデルなどでもよい。時系列データＬ_ｋｉ（ｔ）に対して回帰分析を実施すると、
回帰式ｙ（ｔ）＝Ｆ_ｋｉ（ｘ_１，ｘ_２，…，ｘ_Ｍ）と、残差の平方和Ｅ_ｋｉを得る。
以下では、（ｘ_１，ｘ_２，…，ｘ_Ｍ）をベクトルｘとして記載し、
Ｆ_ｋｉ（ｘ_１，ｘ_２，…，ｘ_Ｍ）をＦ_ｋｉ（ｘ）と記載する。
以下では、
時系列データＬ_ｋｉ（ｔ）、
ｙの区間（Ｙｋ，Ｙｋ＋１］、
回帰式Ｆ_ｋｉ（ｘ）、
残差の平方和Ｅ_ｋｉの
四つ組（Ｌ_ｋｉ（ｔ），（Ｙｋ，Ｙｋ＋１］，Ｆ_ｋｉ（ｘ），Ｅ_ｋｉ）を、
局所時系列データ３０１の集合Ｌ_ｋｉに対する局所構造Ｓ（Ｌ_ｋｉ）と呼ぶ。 (S903: Processing of local time series data model estimation unit 103)
In step S903, the local time series data model estimation unit 103 (an example of a local time series data regression equation generation unit) estimates the local time series data L _ki using a multivariate analysis or a time series analysis method. “Model estimation” is, for example, processing for obtaining a regression equation. For example, time-series data that is an element of the local time-series data 301 that exists for each section k divided in the range of the value of the sensor signal (target variable y) in S902

All models are estimated by multivariate analysis or time series data.
In the following, linear regression analysis is used as an example of multivariate analysis, but factor analysis, singular value decomposition, AR model, state space model, and the like may be used. When regression analysis is performed on the time series data L _ki (t),
The regression equation y (t) = F _ki (x ₁ , x ₂ ,..., X _M ) and the residual sum of squares E _ki are obtained.
In the following, (x ₁ , x ₂ ,..., X _M ) is described as a vector x,
F _ki (x ₁ , x ₂ ,..., X _M ) is _denoted as F _ki (x).
Below,
Time series data L _ki (t),
y section (Yk, Yk + 1],
Regression equation F _ki (x),
A quadruple of residual sums of squares E _ki (L _ki (t), (Yk, Yk + 1], F _ki (x), E _ki ),
This is called a local structure S (L _ki ) for the set L _ki of the local time series data 301.

(Local structure)
That means
Local structure S (L _ki ) = {L _ki (t), (Yk, Yk + 1], F _ki (x), E _ki }
It is.

FIG. 12 conceptually shows the processing of the local time series data model estimation unit 103 described above. The local time series data model estimation unit 103 associates S (L _ki ) with one local time series data L _ki extracted by the local time series data extraction unit 102. In this case, “i” indicating the number of time segments is N from 1 to N as shown in (Formula 1). Further, k indicating the number of data sections for the target variable y (t) (designated time series data) is 1 to K (corresponding to m = 1 to K in (Expression 2)).
That is, i = 1 to N, k = 1 to K.
Therefore, “N × K” pieces of local time series data L _ki can be _generated .
Therefore, “N × K” local structures S (L _ki ) can be formed. FIG. 13 conceptually shows “N × K” local structures S (L _ki ). In FIG. 13, the horizontal axis represents the number of time divisions “i”, and the vertical axis represents the number of data divisions “k”. In that case, one cell corresponds to a certain S (L _ki ).

(S904: Operation of Local Time Series Data Clustering Unit 104)
S904 is a local data clustering process executed by the local time-series data clustering unit 104. In S904, the set of local time-series data models estimated by the local time-series data model estimation unit 103 (that is, N · K S (L _ki )) is divided into clusters, and a representative local representing the cluster for each cluster. Estimate the parameters.
FIG. 10 is a flowchart showing details of the processing flow of S904 executed by the local time-series data clustering unit 104. The subject of the operation in FIG. 10 is the local time series data clustering unit 104, but it will be omitted because it becomes complicated. S904 is the local structure S (L _i ) = (L _i (t), (Yk, Yk + 1], F _i (x), E _i ) obtained in S903.
As a set. However, since it is executed for each k (each data section), the subscript k of L, F, and E is omitted for the sake of brevity. Further, the capital letter N is the number of local time series data (L _i ) (that is, the number of time ranges), and the number of data sections “i” takes values from 1 to N as described above.

If this is explained with reference to FIG. 13, the local time-series data clustering unit 104 executes S (L _i ) for each k, for example, in the case of “k = 2”, the data in the shaded part of FIG. , S (L ₁ ) to S (L _N ).

(S1001)
In S1001, from among the _{S i,} look for the _{S i} with the minimum of _{E i.} Let S _i with the smallest E _{i be} S ₃ . If k = 2, the local time-series data clustering unit 104 in FIG. 13 searches for S _i having the minimum E _i from S _{1 to} S _N.
Next, F _i is substituted into candidate variable m ₁ of the representative local parameter. In this case, F ₃ (regression equation) belonging to S ₃ having the smallest E _i is substituted into the candidate variable m ₁ of the representative local parameter.
in this case,
m ₁ = F ₃
It is.
Next, 1 is substituted into the variable c.
That is, c = 1
It is.
Note that the set number of local clusters that appear in S1003, which will be described later, is “C *” to distinguish it from the variable c.

ここで、ｄｉｓｔ_ｌｍ（Ｓ（Ｌ_ｉ），｛ｍ_１，・・ｍ_Ｃ＊｝）はＳ（Ｌ_ｉ）と｛ｍ_１，・・ｍ_Ｃ＊｝との距離を示し、また回帰式間の距離ｄｉｓｔ_ｒは、回帰式の係数をベクトルとみなした場合のベクトル間の距離とする。距離ｄｉｓｔ（Ｆ_ｉ（ｘ），ｍ_１）はＦ_ｉとｍ_１との距離であるが、予め設定された計算式に基づき算出する。なお回帰式Ｆ_ｉ（ｘ）は簡略化してＦ_ｉ（ｘ）とも表記する。図１４の「ｃ＝１」は、Ｓ１００１でｃ＝１となった場合の（式５）の意味を示している。
変数ｃ＝１の場合、
ｄｉｓｔ（Ｆ_１，ｍ_１）〜ｄｉｓｔ（Ｆ_Ｎ，ｍ_１）のＮ個の距離のなかから、最大の距離を探す。例えば、ｄｉｓｔ（Ｆ_５，ｍ_１）が最大とする（ｉ＝５）。
つまり、
ｍａｘ＝ｄｉｓｔ（Ｆ_５，ｍ_１） 次に、ｍ_ｃ＋１に、Ｆ_ｉ（ｘ）を代入する。次に、ｃに、ｃ＋１を代入する。
この設例では、
ｍ_１＋１＝ｍ_２＝Ｆ_５，
ｃ＝１＋１＝２
となる。 (S1002)

_{Here, dist lm (S (L i} ), {m 1, ·· m C *}) represents the distance between the _{S (L i) {m 1} , ·· m C *}, also among the regression equation distance dist _r of is the distance between vectors in the case where the coefficients of the regression equation was considered vector. The distance dist (F _i (x), m ₁ ) is the distance between F _i and m ₁ and is calculated based on a preset calculation formula. Note that the regression equation F _i (x) is simplified and expressed as F _i (x). “C = 1” in FIG. 14 indicates the meaning of (Formula 5) when c = 1 in S1001.
If variable c = 1,
The maximum distance is searched from among N distances of dist (F ₁ , m ₁ ) to dist (F _N , m ₁ ). For example, dist (F ₅ , m ₁ ) is the maximum (i = 5).
That means
max = dist (F ₅ , m ₁ ) Next, F _i (x) is substituted into m _{c + 1} . Next, c + 1 is substituted for c.
In this example,
m _{1 + 1} = m ₂ = F ₅ ,
c = 1 + 1 = 2
It becomes.

(S1003)
In S1003, it is determined whether or not the variable c is equal to the constant C *.
However, as described above, the constant C * is assumed to be given as a parameter in advance as a number indicating the number of local clusters. If equal, the process proceeds to S1004. If they are not equal, the process returns to S1002.

In this example, it is assumed that the process returns to S1002 because c = 2 at present.
In the state returned to S1002,
c = 2,
m ₂ = F ₅ ,
It is.
From S1001,
m ₁ = F ₃ ,
It is.
Then, similarly to the first S1002 (when c = 1), the maximum distance is searched based on (Expression 5).
“C = 2” in FIG. 14 indicates the meaning of (Formula 5) in the case where c = 2 in S1002.
If variable c = 2,
The maximum distance is searched from N distances “dist (F ₁ , m ₁ ) + (F ₁ , m ₂ )” to “dist (F _N , m ₁ ) + (F _N , m ₂ )”. .
The subsequent operation is the same as the previous S1002.
And
When c = C *, the process proceeds to S1004.
In this example, c = 20 (the number of local clusters).

The above processing of S1002 and S1003 determines C * number of m.
mj, j = 1, 2,..., C *,
It is said.

Hereinafter, D _j , j = 1, 2,..., C *, which will be described later, are C * clusters obtained by clustering the local structure.

(S1004)
In S1004, carried initializes the cluster _{D j.} For example, for N L _i (S _i has L _i as an element) in FIG. 13 (k = 2), C * dist _lm (L _i , m _j ) (predetermined distance defining formula) Find m _j that minimizes.
Next, L _i is substituted into cluster D _j .
FIG. 15 is a diagram conceptually showing this processing.
For example, consider the _{L 1.} The distance dist _lm (L _i , m _j ) between L ₁ and each of m _{1 to} m _{c *} is calculated, and m _j that minimizes the distance is searched for. The equation for obtaining the distance may be a method for determining the distance between vectors when the coefficient of the regression equation is regarded as a vector as used in S1002, or may be another equation.
The distance dist _lm (L _i , m _j ) in this case is
It means the regression equation F _i belonging to the same local structure S _i as the time series data L _i . This is the same as in the case of the distance formula of (Formula 5).
That means
dist _lm (L _i , m _j ) = dist _lm (F _i , m _j )
Since both F _i and m _j are regression equations, the distance between the regression equations can be obtained.
However, it is expressed using the L _i because the clustering of the target is a time-series data.
For example, if L ₁ (F ₁ ) has the shortest distance from m ₁ , the time series data L ₁ belongs to the cluster D ₁ .
Similarly, if L ₂ (F ₂ ) has the shortest distance from m ₁ as shown in FIG. 15, the time-series data L ₂ also belongs to cluster D ₁ .
Similarly, if L ₃ (F ₃ ) has the smallest distance from m ₂ , the time series data L ₃ belongs to the cluster D ₂ .
_Hereinafter, the same to L 4 ~L _N.
Through the processing in S1004, L _{1 to} L _N belong to any cluster D _j of D _{1 to} D _{c *} .

(S1005)
After S1005 (the loop of S1005, S1006, and S1004) is a process of reclustering the local time series data Li (i = 1 to N) based on the initial cluster set in S1003 → S1004. In S1005, with respect to C * number of _{D j,} with respect to the union ∪L _jk ∈D _j of the local time-series data _{L jk} belonging to the cluster _{D j} is some j (in this example k = 2), the regression analyse. By this regression analysis, a regression formula F _j (x) of the cluster D _j is obtained.
Next, the obtained regression equation F _j (x) is substituted into the representative regression equation candidate m _j of the cluster D _j .
FIG. 16 is a diagram conceptualizing the processing of S1005, and shows a reclustering rule.
At the stage when the processing of the first S1004 is finished,
Cluster D ₁ includes local time series data L ₁ and L ₂ ,
A member of the local time-series data _L 3 ~L ₅ to cluster _{D 2,}
... and so on.
In that case,
For cluster D ₁ , find regression equation F _{j = 1} for union L ₁ ∪L ₂ ,
For the cluster D ₂ , the regression equation F _{j = 2} is obtained for the union L ₁ ∪L ₂ ∪L ₃ . The same applies to other clusters.
By this processing, a regression equation is determined for each of C * clusters of the clusters D _{1 to} D _{C *} .
The C * regression equations are set as new m _{1 to} m _{C *} with respect to m _{1 to} m _{C *} obtained in FIG. 15 (S1002, S1003).

In S <b> 1006, it is determined whether or not “C * m _i are all unchanged”.
In the first S1006, the previous “m _{1 to} m _{C *} ” is a so-called initial value created in the loop of S1002 and S1003. Therefore, there is usually a change between “m _{1 to} m _{C *} ” obtained in the first S1005.
If there is a change, the process returns to S1004.
If there is no change, exit.
The C * clusters of D ₁ , D ₂ ,..., D _{C *} when finished are local time series clusters. Ends _"m 1 _{~m C *"} at the time with the capital letter _{"M 1, M 2, ...,} M C * " will be described as these local time series cluster _"D 1 _{to D C *"} in Each representative local parameter.

(Multiple S1004)
A case where the process returns from S1006 to S1004 will be described. FIG. 17 is a conceptual diagram showing the second processing (the same applies to the third and subsequent times). The second time is different from the first S1004 only in that “m _{1 to} m _{C *} ” is “new m _{1 to} m _{C *} ” obtained in S1005. That is, in S1004 for a plurality of times, “new m _{1 to} m _{C *} ” obtained in the immediately preceding S 1005 is used, and clustering of “L _{1 to} L _N ” is performed again. That is, new “new m _{1 to} m _{C *} ” is used, and “L _{1 to} L _N ” is re-clustered.

Since FIG. 10 shows processing for each k, when executed in the order of k = 2 and k = 3, S _i (i: 1 to N) is processed for k = 2 as shown in FIG. Next, S _i (i: 1 to N) is processed for k = 3. Thus, as shown in FIG. 18, for k = 2 the local cluster _D 1 _{to D C *} determines the local cluster _D 1 _{to D C *} is determined for k = 3. Therefore, if it is executed for 1 to K, as shown in FIG. 19, local clusters D ₁ to DC _* are determined for ₁ to K, respectively. A representative local parameter is determined for each of the local clusters D _{1 to} D _{C *} . This is illustrated in FIG. 7, and clusters 703, 704, etc. indicate local clusters. Regression equations 705 and 706 are representative local parameters of each local cluster. In the case of the k difference in FIG. 19, it is displayed as a local cluster and a representative local parameter for each k in FIG. 7, but the k difference is not expressed in FIG.

(Outlier detection unit 106)
The outlier detection unit 106 determines whether or not separately provided segment data corresponds to an outlier for the local cluster and the representative local parameter shown in FIG. That is, the outlier detection unit 106 is the evaluation target data separately given as the evaluation target based on the representative local parameter determined by the local time series data clustering unit 104, and is the type that is the source of the generation of the local time series data. The value defined as the distance for the evaluation target data consisting of a set of multiple time-series data of the same type as multiple time-series data of different values is a threshold value with any of the representative local parameters Detect whether or not. When the threshold value is exceeded, the outlier detection unit 106 determines the evaluation target data (abnormality determination target data 708 determined to be abnormal in FIG. 7) as an outlier.

  In the above description of local clustering (FIG. 10), the variable of interest is assumed to be y and the data partition is considered. However, it is not essential to reflect the data partition. The case where the data classification is not reflected (broadly defined local time-series data) corresponds to the case where only k = 1 in FIGS.

  Note that S1001, S1002, and S1003 show an example of a method (initial cluster division rule) for setting an initial cluster for local time-series data clustering. This initial cluster selection method may be replaced with a known selection method of clustering such as selecting at random.

The case of selecting at random with reference to the flowcharts of FIGS. 20 and 10 will be described.
The initial cluster is denoted as D _j ⁽⁰⁾ (j = 1 to C *).
In FIG. 10, k = 2.
For simplification of explanation, the local time series data L _i is assumed to be 10 pieces,
The set number C * of local clusters is 3.
When selecting at random, the local time series data clustering unit 104
The local time series data L _{1 to} L ₁₀ are divided into initial clusters as follows, for example (S01, S02).
D ₁ ⁽⁰⁾ = L _{1 to} L ₃ ,
D ₂ ⁽⁰⁾ = L _{4 to} L ₆ ,
^{_{D 3 (0) = L 7}} ~L 10.
This is a state in which the processing of S1004 (first time) in FIG. 10 has been completed.
Next, a regression equation of D ₁ ^{(0) to} D ₃ ⁽⁰⁾ is obtained in S1005 (first time), and this is obtained as “m ₁ ^{(1) to} m ₃ ⁽¹⁾ ” (first regression equation that is representative information ^). (S02).
Next S1004 distances between the _{L i} in Run 2 identifies the _m ^{j (1)} having the minimum based on the defining equation of the S1004 (predetermined distance defining equation) (S03).
Then, in S1005 (second time), the local that becomes the basis of the “time series data regression equation (F (x) belonging to S (L _ki ))” that uses the identified m _j ⁽¹⁾ (first regression equation). First clusters D ₁ ^{(1) to} D ₃ ⁽¹⁾ , which are clusters composed of time-series data, are generated in correspondence with m ₁ ^{(1) to} m ₃ ⁽¹⁾ .
D ₁ ⁽¹⁾ = L _{2 to} L ₄ ,
D ₂ ⁽¹⁾ = L _{5 to} L ₇ ,
D ₃ ⁽¹⁾ = L _{8 to} L _10, L ₁ ,
And
The first cluster ^{_{D 1 (1) ~D 3 (}} 1) first cluster _D ¹ by performing a regression analysis on ⁽¹⁾ _~D ³ (1) each to _m ^{1 (2)} ~m ₃ ⁽²⁾ Generate (second regression equation as representative information) (S04).
In step S < ^b > 1006, it is determined whether m ₁ ^{(2) to} m ₃ ⁽²⁾ generated this time is changed from m ₁ ^{(1) to} m ₃ ⁽¹⁾ generated last time. If there is no change, the process ends. If there is a change, the process proceeds to S1004 (third time).
In S1004 (third time), for each L _i , m _j ⁽²⁾ is specified that minimizes the distance from m ₁ ^{(2) to} m ₃ ⁽²⁾ based on the definition formula of S1004 (S05).
Then, in S1005 (third time), the local that is the basis of “time series data regression equation (F (x) belonging to S (L _ki ))” that uses the identified m _j ⁽²⁾ (first regression equation). Second clusters D ₁ ^{(2) to} D ₃ ⁽²⁾ , which are clusters composed of time-series data, are generated in correspondence with m ₁ ^{(2) to} m ₃ ⁽²⁾ .
D ₁ ⁽²⁾ = L _{3 to} L ₅ ,
D ₂ ⁽²⁾ = L _{6 to} L ₈ ,
D ₃ ⁽²⁾ = L _{9 to} L ₁₀ , L _{1 to} L _2,
And
And
The second cluster _D ¹ (2) _~D ³ second cluster _D ¹ (2) by performing a regression analysis on ^{_{(2) ~D 3 (2)}} m 1 (3) every _~m ^{3 (3 )} (Third regression equation as representative information) is generated (S06).
In step S < ^b > 1006, it is determined whether m ₁ ^{(3) to} m ₃ ⁽³⁾ generated this time has a change from m ₁ ^{(2) to} m ₃ ⁽²⁾ generated last time. If there is no change, the process ends.
If there is a change, the process proceeds to S1004 (fourth time).
Through S1004 (the fourth) (S07), the S1005 (fourth), in the same manner as described above, the current _m ¹ (4) _~m ^{3 (4)} is generated (S08).
In S1006, if m ₁ ^{(4) to} m ₃ ⁽⁴⁾ generated this time is not changed from m ₁ ^{(3) to} m ₃ ⁽³⁾ generated last time, the process ends, but m ₁ ^{( 4) to} m ₃ ⁽⁴⁾ are assumed to be unchanged from the previously generated m ₁ ^{(3) to} m ₃ ⁽³⁾ . In this case, the process ends.
In this case, the third clusters D ₁ ^{(3) to} D ₃ ⁽³⁾ at the end of the process are local clusters,
m ₁ ^{(4) to} m ₃ ⁽⁴⁾ are representative local parameters (local cluster representative information) representing each local cluster.

  In FIG. 10, S1001, S1002, S1003, S1004, and S1005 are regression equation generation processes for generating a regression equation, and S1006 indicates whether or not there is a change from the previously generated regression equation each time a new regression equation is generated. When there is a change, the regression formula generation process of the next new regression formula is continued, and when there is no change, the regression formula generation process is terminated without continuing the regression formula generation process of the next new regression formula This is a determination process.

Also, the definition of dist _lm (S (L _i ), {m ₁ ,..., M _C } in S1002 shows an example.In this distance, the error square sum criterion used in the clustering field is used. Dispersion, scattering criteria, trace criteria, determinant criteria, invariant criteria, etc. may be used (reference: Richard O. Duda et al., Translated by Morio Onoe, Pattern Identification, pp. 543-548, New Technology Communication Co., Ltd.) ).

(S905: Operation of Global Time Series Data Model Estimation Unit 105)
S905 is global data model estimation executed by the global time-series data model estimation unit 105. In S905, a global representative time series data model is estimated by connecting the models estimated by the local time series data clustering unit 104.
FIG. 21 is a flowchart showing details of the processing flow of S905.

In S1101, an initial set G of global time-series data candidates is created, and a final global data estimation model is obtained while merging elements of the set in S1102 and thereafter. In step S1101, clusters Di obtained as a result of local time-series data clustering are sequentially extracted to create an initial set G of global time-series data candidates. The initial set G is a set {S (D ₁ ), S (D ₂ ),..., S (D _N )} of the local structure S (Di) of the cluster Di obtained by the processing of S904.
And
There are C _l local time-series data clusters for each interval l of Y,
There are N = ΣC _{l in} total. In the following, S (Di) can represent the local structure after cluster merging,
Objective variable interval (y _is , y _ie ],
A set of cluster representative regression equations Fik (x),
Residual sum of squares Eil,
Local time series data Li,
A set of five time intervals Til in which Li is defined ((y _is , y _ie ], {Fik (x)}, {Eik}, Li, {Til}).
Here, the interval (y _is , y _ie ) of the objective variable is
y _is <y ≦ y _ie .

  In S1102, it is determined whether or not to terminate the connection process in the global time series estimation process. The connection process is executed when the objective variable classifications are adjacent to each other and the condition that the time interval of the local time series data 301 that is an element of the cluster and the representative regression function are included is included.

但し、｜｜ｘ｜｜は、ユークリッド距離とする。Ｓ（Ｄｉ）には、複数の局所構造をもつので、代表回帰式Ｆｉｋは複数存在するので、Ｓ（Ｄｉ）とＳ（Ｄｊ）では、すべてｉｋとｊｌの組を比較した際の最小値をとるように定義する。 For example, for a set of all elements Di, Dj of the set G,
Condition Dist (S (Di), S (Dj)) <δ
Judge whether to satisfy. If the condition is satisfied, the process proceeds to S1105. If the condition is not satisfied, the process proceeds to S1103.
here,
Dist (S (Di), S (Dj)) is defined below, for example.

However, || x || is the Euclidean distance. Since S (Di) has a plurality of local structures, there are a plurality of representative regression equations Fik. Therefore, in S (Di) and S (Dj), the minimum value when all pairs of ik and jl are compared is set. Define to take.

  In S1103, the merge process in the global time series estimation process is executed. For example, Di and Dj that minimize Dist (S (Di), S (Dj)) are obtained for all Di and Dj pairs in the set G.

Ｄｉｓｔ（Ｓ（Ｄｉ），Ｓ（Ｄｊ））がｙ_ｉｅ＝ｙ_ｊｓ
の場合にしか定義されないので、併合後のｙの区間は連続した一つの区間
（ｙ_ｉｓ，ｙ_ｊｅ］になる。 In S1104, the global time-series data candidate set G is updated. Delete S (Di) and S (Dj) from set G and add S (Di + Dj). However, S (Di + Dj) is defined below, for example.

Dist (S (Di), S (Dj)) is y _ie = y _js
Therefore, the y section after merging becomes one continuous section (y _is , y _je ]).

  In S1105, piecewise regression analysis is performed on all Di in the set G. The number of sections at this time may be freely selected, or may be the number of clusters at the time of G initialization that is a component of the cluster Di (that is, equal to the number of representative regression equations included in S (Di)). Good. The piecewise regression equation obtained in S1105 is the estimated global time series data model.

(S906: Operation of Outlier Detection Unit 106)
S906 is an outlier detection process executed by the outlier detection unit 106. With respect to separately provided segment data, a large outlier in a set of representative local time series data models obtained by the global time series data model estimation unit 105 is detected as an anomaly.

As described above, in the plant abnormality detection device 100 according to the first embodiment, local time-series data with less frequency is represented by the processing of S904 for clustering a set of local time-series data divided by time and sensor signal values. Since local parameters are not so much affected, even if the collected data contains data variability due to equipment deterioration or data immediately before abnormalities are mixed, the frequency is low. The effect of preventing a decrease in the accuracy of abnormality detection can be obtained.
Further, it is possible to obtain a global relational expression between sensor signals by generating a global representative time series data model by connecting the models estimated by the local time series data clustering unit 104 obtained in S904. become able to. Therefore, by indicating to the user that the abnormality is determined based on the deviation from the global relationship graph, the explanation of the basis of the abnormality determination can be explained in an easy-to-understand manner.
As shown in FIG. 8, the processing for obtaining this global relational expression connects locally expressed relationships even when the relationship between signals is a non-linear relationship that is unknown in advance. It also has the effect that a global relational expression can be obtained.

(1) The plant abnormality detection apparatus according to the present embodiment extracts a representative trajectory segment in the sense that the trajectory segment has a high frequency by clustering the trajectory segments obtained by dividing into time segments. . Thereby, the accuracy of abnormality detection can be improved by reducing the influence of the trajectory segment that occurs rarely.
(2) Further, the global time series data model estimation unit 105 generates a global representative trajectory by connecting the representative trajectory segments. Therefore, it is possible to obtain a global relationship graph between sensor signals and indicate to the user that the abnormality has been determined based on the deviation from the global relationship graph. can do.

Embodiment 2. FIG.
The fourth embodiment will be described with reference to FIGS. Embodiment 2 demonstrates the hardware constitutions of the plant abnormality detection apparatus 100 which is a computer. FIG. 22 is a diagram illustrating an example of the appearance of the plant abnormality detection device 100 that is a computer. FIG. 23 is a diagram illustrating an example of hardware resources of the CPU allocation time management apparatus 1000 described in the first embodiment.

  In FIG. 22 showing the appearance, the plant abnormality detection device 100 includes a system unit 830, a display device 813 having a CRT (Cathode / Ray / Tube) or LCD (liquid crystal) display screen, a keyboard 814 (Key / Board: K / B). ), A hardware resource such as a mouse 815 and a compact disk device 818 (CDD: Compact Disk Drive), which are connected by a cable or a signal line. The system unit 830 is connected to the network.

  In FIG. 23 showing hardware resources, the plant abnormality detection apparatus 100 includes a CPU 810 (Central Processing Unit) that executes a program. The CPU 810 is connected to a ROM (Read Only Memory) 811, a RAM (Random Access Memory) 812, a display device 813, a keyboard 814, a mouse 815, a communication board 816, a CDD 818, and a magnetic disk device 820 via a bus 825. Control hardware devices. Instead of the magnetic disk device 820, a storage device such as an optical disk device or a flash memory may be used.

  The RAM 812 is an example of a volatile memory. Storage media such as the ROM 811, the CDD 818, and the magnetic disk device 820 are examples of nonvolatile memories. These are examples of a “storage device” or a storage unit, a storage unit, and a buffer. The communication board 816, the keyboard 814, and the like are examples of an input unit and an input device. The communication board 816, the display device 813, and the like are examples of an output unit and an output device. The communication board 816 is connected to the network.

  The magnetic disk device 820 stores an operating system 821 (OS), a window system 822, a program group 823, and a file group 824. The programs in the program group 823 are executed by the CPU 810, the operating system 821, and the window system 822.

  The OS 821 and the program group 823 store programs that execute the functions described as “˜units” in the description of the above embodiments. The program is read and executed by the CPU 810.

  In the description of the above embodiment, the file group 824 includes “to determination result”, “to calculation result”, “to extraction result”, “to generation result”, and “to processing result”. The described information, data, signal values, variable values, parameters, and the like are stored as items of “˜file” and “˜database”. The “˜file” and “˜database” (for example, the plant time series database 101) are stored in a recording medium such as a disk or a memory. Information, data, signal values, variable values, and parameters stored in a storage medium such as a disk or memory are read out to the main memory or cache memory by the CPU 810 via a read / write circuit, and extracted, searched, referenced, compared, and calculated. Used for CPU operations such as calculation, processing, output, printing, and display. Information, data, signal values, variable values, and parameters are temporarily stored in the main memory, cache memory, and buffer memory during the CPU operations of extraction, search, reference, comparison, operation, calculation, processing, output, printing, and display. Is remembered.

  In the description of the embodiment described above, the data and signal values are the memory of the RAM 812, the compact disk of the CDD 818, the magnetic disk of the magnetic disk device 820, other optical disks, mini disks, and DVDs (Digital Versatile Disk). Or the like. Data and signals are transmitted on-line via the bus 825, signal lines, cables, and other transmission media.

  In the above description of the embodiment, what has been described as “to part” may be “to means”, and “to step”, “to procedure”, and “to processing”. May be. That is, what has been described as “˜unit” may be implemented by software alone, a combination of software and hardware, or a combination of firmware. Firmware and software are stored as programs in a recording medium such as a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, and a DVD. The program is read by the CPU 810 and executed by the CPU 810. That is, the program causes the computer to function as the “˜unit” described above. Alternatively, the computer executes the procedure and method of “to part” described above.

  In the above embodiment, although the plant abnormality detection apparatus 100 was demonstrated, it is natural from the above description that operation | movement of the plant abnormality detection apparatus 100 can also be grasped | ascertained as a program for making a computer perform. Moreover, it is also possible to grasp the operation of the plant abnormality detection device 100 described in the above embodiment as a detection method and a detection method performed by each unit of the plant abnormality detection device 100.

In the above embodiment,
When a set of time-series data obtained by observing sequentially over time is used as input data, and the input data is divided by time according to how the input data changes over time. A local time series data extraction unit for extracting series data;
A local time series data model estimation unit for estimating the modeled data by multivariate analysis or a time series analysis method;
A local time series data clustering unit that divides a set of models of local time series data estimated as described above into clusters, and estimates representative local parameters representing the clusters for each cluster;
A plant abnormality detection apparatus having an outlier detection unit that detects, as an abnormality, a large outlier in the set of representative local time series data models obtained as described above with respect to separately provided segment data has been described.

In the above embodiment,
A plant abnormality detection apparatus having a local time-series data extraction unit, characterized by extracting time-series data obtained by further classifying time-series data divided by time into a range of values of time-series data Explained.

In the above embodiment,
This plant abnormality detection apparatus provided with the global time series data model estimation part which produces | generates the candidate of the global representative time series data obtained by connecting the estimation model of representative local time series data was demonstrated.

  DESCRIPTION OF SYMBOLS 100 Plant abnormality detection apparatus, 101 Plant time series database, 102 Local time series data extraction part, 103 Local time series data model estimation part, 104 Local time series data clustering part, 105 Global time series data model estimation part, 106 Outlier Detection unit, 301 local time series data, 701 cluster, 702 cluster, 703 cluster, 704 cluster, 705 regression equation (representative local parameter), 706 regression equation (representative local parameter), 707, 708 abnormality determination target data, 901, 902 Corresponding range of local time series data.

Claims (4)

Time series data belonging to the time range is extracted for each N time ranges (N is an integer of 2 or more) from the first time range to the Nth time range from a plurality of different time series data. A local time-series data extraction unit that generates N local time-series data including a plurality of sets of time-series data in a range;
N local time-series data extracted by the local time-series data extraction unit is divided into initial clusters having a predetermined number of initial clusters according to an initial cluster dividing rule set in advance as an initial cluster dividing rule, and divided. N is generated by generating representative information indicating the characteristics of the initial cluster for each initial cluster, and distributing N local time-series data according to the re-clustering rule set in advance as a re-clustering rule for each generated representative information. Re-clustering is performed to divide the local time-series data into clusters, representative information is regenerated for each re-clustered cluster, and N local time-series data extraction units extracted for each re-generated representative information Re-cluster local time series data,
Similarly,
Repeat the re-clustering of N local time-series data and the regeneration of the representative information, and every time the representative information is regenerated, is the representative information generated this time changed from the representative information generated immediately before? When there is a change, the next representative information regeneration process is continued, and when there is no change, the next representative information regeneration process is continued without re-clustering and representing the N local time-series data. the local time-series data clustering unit to end the process with regeneration of the information,
A local time series data regression equation generating unit that generates a local time series data regression equation corresponding to the N local time series data generated by the local time series data extracting unit for each of the N local time series data. When
With
The local time series data clustering unit
By performing regression analysis on the union of local time series data belonging to each divided initial cluster, a first regression equation is generated as representative information for each initial cluster, and each of the N local time series data is generated. The first regression equation that has the shortest distance calculated according to a predetermined distance definition equation that is a reclustering rule among the first regression equations generated for each initial cluster with respect to the time series data regression equation is identified and identified. Re-clustering by generating a first cluster that is a cluster composed of local time-series data that is the basis of a time-series data regression formula that is the same as the first regression formula, corresponding to different first regression formulas, By performing a regression analysis on the union of local time series data belonging to each first cluster, a second regression equation that is representative information is generated for each first cluster. Regression formula generation process is executed for,
Similarly,
The predetermined distance definition among the respective p + 1 regression equations that are representative information generated for each pth cluster (p is an integer of 1 or more) with respect to each of the time series data regression equations of the N local time series data. The p + 1-th cluster which is a cluster composed of local time-series data that identifies the time-series data regression equation that identifies the p + 1-th regression equation with the shortest distance calculated according to the equation Are re-clustered by generating corresponding to each different p + 1-th regression equation, and regression analysis is performed on the union of local time series data belonging to each p + 1-th cluster for each p + 1-th cluster While executing a regression equation generation process for generating the p + 2 regression equation that is representative information,
Each time a new p + 1th regression equation is generated, it is determined whether there is a change from the previously generated pth regression equation. If there is a change, the next new p + 2 regression equation regression expression generation process is continued and the change is made. When there is not, the determination processing for ending the regression equation generation processing is executed without continuing the regression equation generation processing of the next new p + 2 regression equation,
The local time series data extraction unit includes:
A range of data values of designated time-series data specified in advance among a plurality of different types of time-series data is set to K pieces of data that differ from the first data range to the K-th data range (K is an integer of 2 or more). Divide into ranges, and generate N local time series data using a predetermined local time series data generation rule for each of the divided K data ranges,
The local time series data regression equation generation unit
The local time series data extraction unit calculates a local time series data regression equation corresponding to the total number of K × N local time series data generated for each of the divided K data ranges as K × N local times. Generate for each series data,
The local time series data clustering unit
For each of the K data ranges divided by the local time series data extraction unit, a regression formula generation process and a determination process are performed, and the local time series data corresponding to the data range generated by the local time series data extraction unit, A local time series data regression formula generated by the local time series data regression formula generation unit, and a local time series data regression formula in which the data range generated by the local time series data extraction unit is the same as the local time series data Clustering device to execute. The local time series data clustering unit
When it is determined that the regression equation generation processing is to be terminated in the determination processing, the cluster that is the source of the last generated regression equation is determined as a local cluster, and the regression equation corresponding to the determined local cluster is represented by that cluster. The clustering apparatus according to claim 1 , wherein local cluster representative information to be determined is determined. The clustering apparatus includes:
Based on local cluster representative information determined by the local time-series data clustering unit, the evaluation target data separately given as an evaluation target, and a plurality of different types of time-series data from which local time-series data is generated An error in detecting whether the value defined as the distance exceeds any threshold with any local cluster representative information for the evaluation target data consisting of multiple sets of time series data of the same type. clustering apparatus according to claim 1 or claim 2, characterized in that with a value detection unit.   Time series data belonging to the time range is extracted for each N time ranges (N is an integer of 2 or more) from the first time range to the Nth time range from a plurality of different time series data. Local time-series data extraction processing for generating N local time-series data consisting of a plurality of sets of time-series data in a range;
  The N local time-series data extracted by the local time-series data extraction process are divided into initial clusters having a preset number of initial clusters in accordance with an initial cluster division rule set in advance as an initial cluster division rule. N is generated by generating representative information indicating the characteristics of the initial cluster for each initial cluster, and distributing N local time-series data according to the re-clustering rule set in advance as a re-clustering rule for each generated representative information. Re-clustering is performed to divide the local time-series data into clusters, representative information is regenerated for each re-clustered cluster, and the N time-series data extraction processes extracted for each re-generated representative information Re-cluster local time series data,
Similarly,
  Repeat the re-clustering of N local time-series data and the regeneration of the representative information, and every time the representative information is regenerated, is the representative information generated this time changed from the representative information generated immediately before? When there is a change, the next representative information regeneration process is continued, and when there is no change, the next representative information regeneration process is continued without re-clustering and representing the N local time-series data. Local time series data clustering processing to finish the process of information regeneration,
  Local time series data regression equation generation processing for generating a local time series data regression equation corresponding to N local time series data generated by the local time series data extraction processing for each N local time series data When
And execute
  The local time series data clustering process is:
  By performing regression analysis on the union of local time series data belonging to each divided initial cluster, a first regression equation is generated as representative information for each initial cluster, and each of the N local time series data is generated. The first regression equation that has the shortest distance calculated according to a predetermined distance definition equation that is a reclustering rule among the first regression equations generated for each initial cluster with respect to the time series data regression equation is identified and identified. Re-clustering by generating a first cluster that is a cluster composed of local time-series data that is the basis of a time-series data regression formula that is the same as the first regression formula, corresponding to different first regression formulas, By performing a regression analysis on the union of local time series data belonging to each first cluster, a second regression equation that is representative information is generated for each first cluster. Regression formula generation process is executed for,
  Similarly,
  The predetermined distance definition among the respective p + 1 regression equations that are representative information generated for each pth cluster (p is an integer of 1 or more) with respect to each of the time series data regression equations of the N local time series data. The p + 1-th cluster which is a cluster composed of local time-series data that identifies the time-series data regression equation that identifies the p + 1-th regression equation with the shortest distance calculated according to the equation Are re-clustered by generating corresponding to each different p + 1-th regression equation, and regression analysis is performed on the union of local time series data belonging to each p + 1-th cluster for each p + 1-th cluster While executing a regression equation generation process for generating the p + 2 regression equation that is representative information,
  Each time a new p + 1th regression equation is generated, it is determined whether there is a change from the previously generated pth regression equation. If there is a change, the next new p + 2 regression equation regression expression generation process is continued and the change is made. When there is not, the determination processing for ending the regression equation generation processing is executed without continuing the regression equation generation processing of the next new p + 2 regression equation,
  The local time series data extraction process includes:
  A range of data values of designated time-series data specified in advance among a plurality of different types of time-series data is set to K pieces of data that differ from the first data range to the K-th data range (K is an integer of 2 or more). Divide into ranges, and generate N local time series data using a predetermined local time series data generation rule for each of the divided K data ranges,
  The local time series data regression formula generation process is:
  In the local time series data extraction processing, the local time series data regression equation corresponding to the total number K × N of local time series data generated for each of the divided K data ranges is represented as K × N local times. Generate for each series data,
  The local time series data clustering process is:
  For each of the K data ranges divided by the local time-series data extraction process, a regression equation generation process and a determination process are performed, the local time-series data corresponding to the data range generated by the local time-series data extraction process, A local time series data regression formula generated by the local time series data regression formula generation process, and a local time series data regression formula in which the data range generated by the local time series data extraction process is the same as the local time series data Clustering program to execute.

Images