JP2010191641A - Method, program, and device for plant monitoring - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To highly precisely detect abnormality in a plant where states change each moment. <P>SOLUTION: There is a case where a factor having strong effects to decrease a Mahalanobis distance is present in the operation data of the plant. In that case, even when the plant is put in an abnormal state, there is a possibility that the Mahalanobis distance takes a small value. As a result, there is a possibility of overlooking the abnormality when simply monitoring the Mahalanobis distance. Then, the factor to decrease the Mahalanobis distance is specified, and the state of the plant is monitored by using the Mahalanobis distance excluding the contribution, so that the overlooking of such abnormality is prevented. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a data processing technique for monitoring a plant.

  In various plants such as gas turbine power plants, nuclear power plants, and chemical plants, plant state quantities such as temperature and pressure are acquired, and whether or not the plant is operating normally is monitored based on the state quantities. In monitoring the plant, it is necessary to monitor a large number of state quantities. Also, skill is required to determine whether or not the plant is operating normally by monitoring the state quantity trend.

  In plant monitoring, a technique using Mahalanobis distance obtained by statistically processing operation data is known. Patent Document 1 describes an example of such a technique regarding an oil refinery plant. According to this document, the oil refining plant has various processes, so that it may be normal even if the Mahalanobis distance is large, or it may be abnormal if it is small. Accordingly, it is an object of the present invention to prevent erroneous determination by evaluating a relative value between a measured value of a state quantity and similar normal operation data.

FIG. 1 is a flowchart showing the monitoring method described in Patent Document 1. In this technique, the following method is generally performed.
(1) Collect plant operation data.
(2) Calculate the Mahalanobis distance of the collected operation data.
(3) Search for data similar to the operation data (2) (hereinafter, similar data) from the normal operation data separately stored in the database.
(4) The Mahalanobis distance of the similar data found in (3) is calculated.
(5) The Mahalanobis distance of the operation data is compared with the Mahalanobis distance of the similar data, and if the difference is large, it is determined as abnormal.

JP 2006-309570 A

  In the invention described in Patent Document 1, it may be determined that the state of the plant is normal even when the Mahalanobis distance increases. However, the original advantage of Mahalanobis distance is that normality / abnormality can be judged by the property that the distance increases when the driving data deviates from the normal population distribution regardless of the distribution state. It has not been. In other words, both the Mahalanobis distance and the driving state are required for a certain state, and the determination of the similarity of the driving state can be a new erroneous determination factor.

  Plant conditions change from moment to moment. In addition, plant operating data is affected not only by operating conditions but also by climate. Therefore, the operation data to be accumulated becomes enormous. Also, the first pseudo-abnormality (when the Mahalanobis distance is large but the plant is normal) cannot be detected because there is no similar data in the database.

  The present invention has been made in view of the above circumstances, and it is possible to detect a plant abnormality with high accuracy without acquiring similar operation data, considering that the state of the plant changes from moment to moment. With the goal.

  Hereinafter, means for solving the problem will be described using the numbers used in [DETAILED DESCRIPTION] in parentheses. These numbers are added to clarify the correspondence between the description of [Claims] and [Mode for Carrying Out the Invention]. However, these numbers should not be used to interpret the technical scope of the invention described in [Claims].

  The plant monitoring method according to one aspect of the present invention includes a step (S2) of acquiring operation data including a plurality of items indicating a plant state, and a Mahalanobis distance of the operation data with respect to a recorded normal state of the plant as a total Mahalanobis distance. Using the orthogonal table having two levels indicating whether each of the plurality of items is used as a factor and each of the plurality of items is used for calculation of the Mahalanobis distance, A step (S6) of specifying an item that contributes most to a decrease in Mahalanobis distance as a reduction factor candidate, and when it is determined that the contribution of the reduction factor candidate to the Mahalanobis distance is large based on a set criterion, Step (S8) to set as a reduction factor, and operation data related to a correction item obtained by removing the reduction factor from a plurality of items For the corrected operation data, the step of calculating the Mahalanobis distance as the corrected Mahalanobis distance (S14), and the step of determining whether the state of the plant is abnormal by comparing the corrected Mahalanobis distance with a predetermined reference (S10). ).

  In the operation data of the plant, there may be a factor having a strong effect of reducing the Mahalanobis distance. In such a case, even if the plant is in an abnormal state, the Mahalanobis distance may take a small value. Therefore, simply monitoring the Mahalanobis distance may overlook anomalies. According to the present invention, it is possible to identify such a reduction factor and monitor the plant state using the Mahalanobis distance excluding the contribution, and thus it is possible to prevent such an abnormality from being overlooked.

  A plant monitoring program according to one aspect of the present invention causes a computer to execute the plant monitoring method according to the present invention.

  A plant monitoring apparatus according to one aspect of the present invention includes an operation data acquisition unit (14) that acquires operation data including a plurality of items indicating a plant state, and a total Mahalanobis distance of the operation data with respect to a normal state of the recorded plant. Using a total Mahalanobis distance calculation unit (16) for calculating the Mahalanobis distance, and an orthogonal table having two levels indicating whether each of the plurality of items is used as a factor and whether each of the plurality of items is used for the calculation of the Mahalanobis distance The reduction factor candidate specifying unit (17) for specifying, as a reduction factor candidate, the item that most contributes to the reduction of the Mahalanobis distance among a plurality of items, and the criterion based on the set criteria for the contribution of the reduction factor candidate to the Mahalanobis distance Is determined, a reduction factor setting unit (1) that sets a reduction factor candidate as a reduction factor ) And a corrected Mahalanobis distance calculation unit (20) that calculates the Mahalanobis distance as a corrected Mahalanobis distance with respect to the corrected operation data that is the operation data related to the correction item obtained by removing the reduction factor from a plurality of items, And a plant state determination unit (28) for determining whether or not the state of the plant is abnormal.

  According to the present invention, it is possible to detect an abnormality of the plant with high accuracy in consideration of the state of the plant changing every moment.

FIG. 1 is a flowchart of a plant monitoring method in the prior art. FIG. 2 shows an example of Mahalanobis distance before and after removal of the reduction factor. FIG. 3 is a diagram for explaining the excluded section and the averaging section. FIG. 4 shows the configuration of the plant monitoring apparatus. FIG. 5 shows a program stored in the storage unit. FIG. 6 is a flowchart showing the operation of the plant monitoring apparatus. FIG. 7 shows the configuration of operation data. FIG. 8 shows normal operation data. FIG. 9 shows an orthogonal table. FIG. 10 shows an example of the excluded section and the averaging section. FIG. 11 shows an example of the excluded section and the averaging section. FIG. 12 shows an example of the excluded section and the averaging section. FIG. 13 is a flowchart showing processing of a very small singular value.

  First, the concept of the plant monitoring method in one embodiment of the present invention will be described. The plant has a number of components. In general, there are factors that increase the Mahalanobis distance, as well as factors that decrease the Mahalanobis distance, such as an element indicating the occurrence of an abnormality. In addition, as described above, the state of the plant changes from moment to moment, so these factors also change with time. For example, a reduction factor at one time may replace another reduction factor at another time.

  If a factor that greatly reduces the Mahalanobis distance is acting, even if an abnormality has occurred, the Mahalanobis distance may be overlooked and the abnormality may be overlooked. In the example shown in FIG. 2, when the Mahalanobis distance (before deletion) calculated for all items is monitored in the area surrounded by a square, no increase in distance is particularly observed. However, if the Mahalanobis distance (after deletion) is deleted, the peak of the Mahalanobis distance can be found.

  Therefore, an orthogonal table analysis is periodically performed to identify items that greatly contribute to the decrease of the Mahalanobis distance as reduction factors, and the Mahalanobis distance is recalculated without the reduction factors. For example, the gain of the desired large S / N ratio used in the Taguchi method is used to determine the reduction factor. However, the present invention is not limited to this, and an index such as a contribution rate may be used.

  It is thought that which factor becomes the reduction factor changes at each time point. Therefore, the calculation using all items is always executed, and the reduction factor is specified again by performing orthogonal table analysis at each time point.

When any of the following conditions is satisfied, it is determined that the state of the plant is abnormal.
(1) When the corrected Mahalanobis distance D1, which is the Mahalanobis distance calculated excluding the reduction factor, exceeds the threshold T1.
(2) When the difference D2 = D1−D0 between the Mahalanobis distance D0 calculated using D1 and all items exceeds the threshold T2.
How to determine the threshold values T1 and T2 will be described later.

  A procedure for selecting monitoring items based on the results of orthogonal table analysis in Taguchi method is generally used. However, in a system such as a plant whose state changes from moment to moment, an item selected at one point may become inappropriate at another point. In the present embodiment, the reduction factor is updated as needed so that it always matches the state of the plant. As a result, by removing factors that greatly contribute to the reduction of Mahalanobis distance at each time point, it is possible to detect abnormalities that may be missed by monitoring using all items.

  Compared with the above-mentioned patent document 1, in this embodiment, the comparison between the operation data and the database is unnecessary. As a result, the monitoring method is simple and there is no risk of misjudgment because there is no need to judge the similarity of operating conditions.

  Next, an example of how to determine the threshold values T1 and T2 will be described with reference to FIG. The vertical axis represents D1 or D2 described above. The horizontal axis indicates time. The operation data is acquired at regular intervals. Operation data includes control signals for controlling plant equipment (example: fuel control valve opening) and sensor detection values (example: generator output, turbine speed, fuel temperature, fuel flow, air compressor inlet) And outlet temperature, blade path temperature, exhaust gas temperature, nitrogen oxide concentration, combustion pressure fluctuation, bearing vibration, bearing metal temperature, rotor cooling air temperature, disk cavity temperature). The point in the figure indicates the value of D1 or D2 of the operation data acquired at each time. The determination time T indicates a time at which operation data that is a target for determining whether the state of the plant is normal or abnormal is acquired. In order to determine the threshold value, an averaging interval and an exclusion interval are set. The excluded section is a section including k pieces of data preceding the determination time. The averaging section is a section including n pieces of data preceding the excluded section.

  When abnormality determination is performed at the determination time T, an average value a1 of D1 for n averaged sections is obtained. The product m1 · a1 of this a1 and a predetermined value m1 is set to the threshold value T1. The average value a2 is similarly obtained for D2. The product m2 · a2 of this a2 and a predetermined value m2 is set to the threshold value T2. Different values may be set for D1 and D2 as n and k that define the averaging interval.

  By setting the exclusion section, the following advantages can be obtained. If the average of data up to immediately before the determination time is used to set the threshold value, when an abnormality occurs and D1 or D2 tends to increase, the threshold value increases due to the abnormality. In this case, even if an abnormality has occurred, there is a possibility that the abnormality is missed because the threshold is large. By providing the excluded section, it is possible to suppress an increase in the threshold value and prevent an abnormality from being overlooked.

  In addition, since the plant including a gas turbine consists of various components, the responsiveness of a state change is also various. Therefore, it is necessary to make the averaging section and the exclusion section suitable for the actual situation of the plant. For example, it is empirically known that when a combustor of a gas turbine fails, the blade path temperature shows an abnormal value in about 15 minutes. Thus, when detecting a failure of a combustor or a device having similar characteristics, it is desirable to set the averaging period to 60 minutes and the exclusion period to about 15 minutes.

  Next, a plant monitoring apparatus for realizing the above monitoring method will be described. FIG. 4 shows the configuration of the plant monitoring apparatus in the present embodiment. The plant monitoring device 2 is a computer and includes a calculation control unit 4, a storage unit 6, an input unit 8, an output unit 10, and a communication unit 12. The communication unit 12 is connected to the measuring device 13. The measuring device 13 measures the physical quantity indicating the operation state of the plant and generates operation data.

  FIG. 5 shows a program stored in the storage unit 6. The storage unit 6 includes an operation data acquisition unit 14, a total Mahalanobis distance calculation unit 16, a reduction factor candidate identification unit 17, a lifting factor setting unit 19, a modified Mahalanobis distance calculation unit 20, a correlation matrix generation unit 22, an inverse matrix calculation unit 24, A plant state determination unit 28, a singular value calculation unit 30, and an under-singular value determination unit 32 are provided. The arithmetic control unit 4 reads out these programs from the storage unit 6 and operates according to the procedure described in the read program. In the following description, for the sake of convenience, the operation executed by the arithmetic control unit 4 in accordance with the procedure of each unit will be described as the operation of each program itself.

  FIG. 6 is a flowchart showing the operation of the plant monitoring device 2. The measuring device 13 measures the state of the plant and generates operation data composed of a plurality of items indicating the measured values. The operation data acquisition unit 14 collects operation data from the measurement device 13 via the communication unit 12 and stores it in the storage unit 6. This collection of operation data is repeatedly performed at predetermined time intervals (step S2).

FIG. 7 shows the configuration of operation data. Each parameter of the plurality of items is indicated by P ₁ , P ₂ ... P _n . In association with the time T indicating the measurement time of the operating data, parameters _P 1 at the time _{T, P 2, P 3 ···} P each observation value _X 1T of _{n, X 2T, X 3T ···} X _nT is collected and recorded.

The storage unit 6 of the plant monitoring device 2 stores time-series operation data generated by the measurement device 13 measuring the plant when the plant is operating in a normal state. This normal operation data is used as a unit space for calculating the Mahalanobis distance. FIG. 8 shows an example of normal operation data. At m times from time T1 to Tm, the measured values during normal operation of parameters P ₁ , P ₂ , P ₃ ... P _n are registered in advance.

Normal operation data is acquired as follows. Various sensors are attached to the plant and the operation state is measured. Time-series operation state data indicating the measured operation state is collected by an analysis computer via the Internet or a dedicated line. Of the collected time-series operation data, a portion satisfying the following conditions is adopted as normal operation data.
(1) It is data in a normal operation data collection section that ends at a time point before a predetermined period from the present time and is between the end and a predetermined period from the end to the past. Since the state of the plant may greatly shift with the passage of time, it is desirable to periodically discard the old data and update the unit space by adding new data by such processing.
(2) The data is not determined to be abnormal based on a predetermined standard as shown in FIG.

  Next, the total Mahalanobis distance calculation part 16 produces | generates the correlation matrix showing the correlation degree between each parameter. In the example of FIGS. 7 and 8, the operation data is composed of n parameters, so the correlation matrix is a square matrix of n rows and n columns. Next, the total Mahalanobis distance calculation unit 16 calculates the Mahalanobis distance of the collected operation data as the total Mahalanobis distance for all items used for monitoring in the operation data, using the normal operation data as a unit space (step S4).

Below, the procedure of the calculation which the total Mahalanobis distance calculation part 16 performs is demonstrated.
(1) Normalization of each parameter The following U _kTj is _obtained for the operation data X _{kTj at} time Tj, where mk is the average of the parameters P _k and σk is the standard deviation. U _kTj is a parameter normalized to a distribution with a mean of 0 and a standard deviation of 1.

(2) Formation of Correlation Matrix The correlation matrix generator 22 obtains the correlation between all parameters according to the definition of the correlation for the data normalized by the above procedure, and forms the following correlation matrix R. r _ij indicates the degree of correlation between the parameters Pi and Pj. The diagonal components are all 1.

逆行列計算部２４は、相関行列の逆行列Ｒ^−１を計算する。マハラノビス距離ＭＤは定義に従い、以下のように求められる。

(3) It is assumed that the operation data at time T is represented by the following vector X.

The inverse matrix calculation unit 24 calculates an inverse matrix R ⁻¹ of the correlation matrix. The Mahalanobis distance MD is determined as follows according to the definition.

The reduction factor candidate specifying unit 17 uses an orthogonal table of Taguchi methods to select an item that contributes most to a decrease in Mahalanobis distance among a plurality of items (parameters P ₁ , P ₂ ... P _n ) that constitute the operation data. It specifies as a reduction factor candidate (step S6).

FIG. 9 shows an example using the L8 orthogonal table. Each of the plurality of items constituting the operation data is used as factors P ₁ , P ₂ ... P _n of the orthogonal table. In this example, the orthogonal table has two levels. This level indicates whether or not to adopt each factor in calculating the Mahalanobis distance for monitoring the plant. In each column, “1” indicates that the factor is adopted, and “2” indicates that the factor is not adopted. For example, combination No. 1 indicates that the Mahalanobis distance is calculated for all factors. Each Y1~Y8 combination of M _D column No. It is a Mahalanobis distance calculated based on a combination of factors defined in 1 to 8.

Then, for example, when focusing on factors _{P 1,} this factor _{P 1} is a combination No. 1 to 4 are used. 5 to 8 is not adopted. Thus, for example, by calculating the following values can be used to assess the contribution to the Mahalanobis distance by the presence of Factor P _1.
{(Y1 + Y2 + Y3 + Y4)-(Y5 + Y6 + Y7 + Y8)} / 4
The evaluation of the degree of contribution of each factor to the Mahalanobis distance is not limited to the above formula, and other indicators such as a gain of the desired SN ratio used in the Taguchi method may be used. The reduction factor candidate specifying unit 17 specifies a factor that contributes most to the decrease in Mahalanobis distance among all the factors as a reduction factor candidate.

  The reduction factor setting unit 18 evaluates the contribution of the reduction factor candidate to the Mahalanobis distance, and sets the reduction factor candidate as a reduction factor to be excluded from the monitoring target when it is determined that the contribution is large based on the set criterion. . The set standard is, for example, the magnitude relationship between the value of {(Y1 + Y2 + Y3 + Y4) − (Y5 + Y6 + Y7 + Y8)} / 4 described with reference to FIG. 9 and a predetermined threshold (step S8).

  A plurality of factors may be set as the reduction factor. In that case, in step S6, a plurality of factors are arranged in descending order of contribution to the decrease in Mahalanobis distance. And it is compared with the set reference | standard in order from the factor with the higher degree of contribution, and it is determined whether it is a reduction factor.

  If the reduction factor setting unit 18 does not determine the reduction factor candidate as a reduction factor to be excluded from monitoring (NO in step S8), the plant state determination unit 28 is normal or abnormal based on the total Mahalanobis distance. Determine whether. Specifically, the total Mahalanobis distance of the operation data is displayed on the display device provided in the output unit 10. Further, when the total Mahalanobis distance is larger than a predetermined standard, the output unit 10 changes the display color of the corresponding data, outputs a warning sound, etc. to the plant monitoring worker. A warning is given (step S10).

  When the reduction factor setting unit 18 determines the reduction factor candidate as a reduction factor (YES in step S8), the corrected Mahalanobis distance calculation unit 20 excludes the reduction factor from the corrected operation data obtained by removing the reduction factor from the operation data and the normal operation data. A correction unit space is created (step S12). The corrected Mahalanobis distance calculation unit 20 calculates the correlation matrix of the corrected unit space in the same manner as the total Mahalanobis distance calculation procedure, and uses this correlation matrix to calculate the Mahalanobis distance of the corrected operation data as the corrected Mahalanobis distance (step S14). ). The plant state determination unit 28 determines whether the state of the plant is normal or abnormal by comparing the corrected Mahalanobis distance with a predetermined reference. Specifically, for example, the plant state is determined by a method of comparing the Mahalanobis distances D1 and D2 described with reference to FIG. 3 with T1 and T2, respectively. When it is determined that the state of the plant is abnormal, the output unit 10 gives a warning to the plant monitoring worker (step S10).

[Combination of multiple averaging intervals and exclusion intervals]
In the description with reference to FIG. 3, a method of determining an abnormality by setting an averaging interval and an exclusion interval is shown. By using a plurality of sets of the averaging interval and the exclusion interval, information for specifying the cause of abnormality can be obtained. For example, in a gas turbine, as described above, there is a value indicating an abnormality about 15 minutes after the occurrence of the abnormality, but there is also a value that changes after several hours from the occurrence of the abnormality. These abnormalities can be identified by setting a plurality of sets of averaging intervals and exclusion intervals.

  FIG. 10 shows an example of the exclusion interval and the averaging interval when the responsiveness of abnormality is high. By setting any section short, an abnormality with high responsiveness can be detected at an early stage.

  FIG. 11 shows an example when the responsiveness of abnormality is low. If such an abnormality is set to a short exclusion section as shown in FIG. 10, the determination threshold value may increase due to the influence of an increase in the operation data value due to the abnormality, and the abnormality may not be detected. By making the exclusion section longer as shown in FIG. 11, an abnormality with low responsiveness can be detected. In addition, since the parameter with low responsiveness often changes slowly, the averaging interval is also set long as shown in FIG.

  FIG. 12 shows another example of the exclusion period and the averaging period. When the state changes gradually due to an abnormality, it can be detected by the exclusion interval and the averaging interval set in this way. In the example of FIG. 12, the parameter gradually decreases in the long term before the parameter value increases rapidly. In such an example, if the averaging interval is set to be long, the value when the parameter value is relatively high is incorporated into the average value, and the value of ma increases. In such a case, setting the averaging period longer will have an adverse effect, so it is desirable to provide an appropriate set value for the upper limit of the length of the averaging period.

  As described above, with a combination of a plurality of exclusion intervals and averaging intervals, for example, a combination of an averaging interval and an exclusion interval when the responsiveness of abnormality is high is pattern A, and a combination when the responsiveness of abnormality is low is pattern B Every two patterns are used together to determine abnormality. At this time, if an abnormality is detected in the pattern A and no abnormality is detected in the pattern B, it can be determined that the abnormality is highly responsive, such as an abnormality in the combustor.

[Determination of abnormal singular values]
The Mahalanobis distance may show an abnormal increase not only when the plant is in an abnormal state but also due to a convergence problem in numerical calculations. In order to calculate the Mahalanobis distance according to the above [Equation 4], it is necessary to obtain the inverse matrix of the correlation matrix shown in [Equation 2] by numerical calculation. In this case, singular value decomposition can be generally used. At this time, as will be described later, if the singular value is too small, the Mahalanobis distance may become unnaturally large even if no abnormality occurs. It is desirable to distinguish such a phenomenon from a plant abnormality.

  FIG. 13 is a flowchart showing a procedure for processing such a very small singular value. The operation data acquisition unit 14 acquires time-series operation data from the measurement device 13 (step S22). This operation data is used as normal operation data. The correlation matrix generation unit 22 forms a correlation matrix from the normal operation data. The singular value calculation unit 30 performs singular value decomposition on the correlation matrix to obtain a singular value (step S24).

  As described above, normal monitoring of the plant is performed using the total Mahalanobis distance or, if there is a reduction factor, the modified Mahalanobis distance (step S26). When the plant state determination unit 28 determines that the plant state is normal (step S28 NO), normal monitoring is continued (step S30). When the plant state determination unit 28 determines that the state of the plant is abnormal based on the Mahalanobis distance (step S28 YES), the process proceeds to a check for the presence or absence of a very small singular value (step S32).

  The under-singular value determination unit 32 compares each of the singular values obtained in step S24 with a predetermined threshold value registered in advance. If there is a singular value smaller than the threshold value (YES in step S32), normal monitoring is continued (step S36). When there is no singular value smaller than the threshold value (NO in step S32), the under singular value determination unit 32 outputs a warning indicating the fact by the output unit 10. In response to the warning, the plant monitoring worker analyzes the cause of the Mahalanobis distance being greater than the reference (step S34). By such processing, when the Mahalanobis distance is a large value due to an excessively small singular value, the cause can be identified without confusion for the monitoring worker.

  In step S34, the plant monitoring worker analyzes the cause of the excessive Mahalanobis distance. The plant monitoring device 2 desirably has a function to support this analysis. This function will be described below.

  In step S6 of FIG. 6, the contribution to the Mahalanobis distance was calculated for each of a plurality of items constituting the driving data. In this calculation, the pulling factor setting unit 19 specifies at least one item that greatly contributes to the increase of the Mahalanobis distance as a pulling factor, and creates a list arranged in descending order of contribution.

  When the plant state determination unit 28 determines that the state of the plant is abnormal, the monitoring worker performs a predetermined operation on the input unit 8 to display the pulling factor list on the display screen of the output unit 10. By referring to this pulling factor list, it is possible to identify the parameter causing the abnormality in the operation data. Once the parameters have been identified, a specific cause analysis of abnormalities such as a gas turbine combustor failure is performed.

  When the plant state determination unit 28 determines that the state of the plant is abnormal and the under-singular value determination unit 32 outputs a warning indicating that a very small singular value exists, Refer to If it can be determined that there is no parameter causing an abnormality based on the pulling-up factor list, the monitoring worker determines that the warning is not caused by a plant abnormality but is caused by the convergence problem of the Mahalanobis distance calculation, and continues normal monitoring. If a warning indicating that a very small singular value exists is issued, but the warning that the plant state determination unit 28 determines that the state of the plant is abnormal is not issued, the monitoring worker performs normal monitoring. to continue.

Ｕはｎ行ｎ列のユニタリ行列、Ｖ^Ｔはｎ行ｎ列のユニタリ行列Ｖの随伴行列である。逆行列は以下のように求められる。

[Supplement about singular value decomposition]
The correlation matrix R is decomposed into the following forms. a1, a2,... an are singular values.

U is a unitary matrix of n rows and n columns, and V ^T is an adjoint matrix of a unitary matrix V of n rows and n columns. The inverse matrix is obtained as follows.

Mahalanobis distance D ^2, when a signal of k data and Y, are expressed as follows.
D ² = YR ⁻¹ Y ^T / k
Therefore, if there is a very small singular value, the Mahalanobis distance becomes very large in inverse proportion to that value. Since the singular value is determined from the unit space that is the operation data selected from the normal state, in other words, the Mahalanobis distance becomes very large even if there is no abnormality in the plant.

  Also, if the singular value is very small, errors such as digit loss due to a computer are likely to occur. The calculation diverges when the singular value becomes zero. Therefore, if the singular value is very small, it is impossible to accurately know the state of the plant based on the calculated Mahalanobis distance. For the above reasons, when a very small singular value exists, the plant monitoring device warns the monitoring worker.

2 Plant Monitoring Device 4 Arithmetic Control Unit 6 Storage Unit 8 Input Unit 10 Output Unit 12 Communication Unit 13 Measuring Device 14 Operation Data Acquisition Unit 16 Total Mahalanobis Distance Calculation Unit 17 Reduction Factor Candidate Identification Unit 18 Reduction Factor Setting Unit 19 Increase Factor Setting Unit 20 corrected Mahalanobis distance calculation unit 22 correlation matrix generation unit 24 inverse matrix calculation unit 28 plant state determination unit 30 singular value calculation unit 32 under-singular value determination unit

Claims (17)

A step of acquiring operation data including a plurality of items indicating the state of the plant;
Calculating the Mahalanobis distance of the operating data for the recorded normal state of the plant as a total Mahalanobis distance;
Using the orthogonal table having two levels indicating whether each of the plurality of items is a factor and each of the plurality of items is used for calculation of the Mahalanobis distance, the most reduced Mahalanobis distance among the plurality of items Identifying an item that contributes to a reduction factor candidate;
When it is determined that the reduction factor candidate has a large contribution to the Mahalanobis distance based on a set criterion, the step of setting the reduction factor candidate as a reduction factor;
Calculating the Mahalanobis distance as the corrected Mahalanobis distance for the corrected operation data that is the operation data related to the correction item obtained by removing the reduction factor from the plurality of items;
Determining whether or not the state of the plant is abnormal by comparing the corrected Mahalanobis distance with a predetermined reference. A plant monitoring method according to claim 1, comprising:
The step of identifying the item that contributes most to the decrease in the Mahalanobis distance is a plant monitoring method that is repeatedly executed at predetermined time intervals. A plant monitoring method according to claim 1 or 2,
The predetermined reference is set with the determination time as a termination from a start time that is a predetermined time preceding a determination time at which a step of determining whether or not the state of the plant is abnormal is executed. A plant monitoring method that is determined based on an average value of the modified Mahalanobis distance in an averaged section that is a section from the start of the excluded section to an end time. A plant monitoring method according to claim 3, wherein
Each of a plurality of averaging intervals is set as the averaging interval,
Each of a plurality of predetermined criteria is set as the predetermined criterion,
In the step of determining whether or not it is abnormal, it is determined which of the plurality of predetermined standards the corrected Mahalanobis distance exceeds. A plant monitoring method according to claim 3 or 4,
The predetermined standard indicates that the corrected Mahalanobis distance has reached a predetermined number of times the average value in the averaging section at the determination time. A plant monitoring method according to claim 3 or 4,
The predetermined criterion indicates that a difference between the corrected Mahalanobis distance and the total Mahalanobis distance has reached a predetermined number of times the average value in the averaging section at the determination time. A plant monitoring method according to any one of claims 1 to 6,
Calculating the modified Mahalanobis distance comprises:
Generating a correlation matrix for the correction item of the operation data in the normal state;
Obtaining an inverse matrix of the correlation matrix by singular value decomposition;
Using the correlation matrix and the inverse matrix, calculating a Mahalanobis distance of the corrected operation data as the corrected Mahalanobis distance,
The plant monitoring method further includes:
Calculating a singular value by performing singular value decomposition on the correlation matrix;
And a step of outputting a warning when an under-singular value smaller than a predetermined threshold exists among the singular values. The plant monitoring method according to claim 7, further comprising:
Each of the correction items is a factor, and an orthogonal table having two levels indicating whether or not each of the correction items is used for calculation of the Mahalanobis distance is used to contribute to an increase in the Mahalanobis distance among the plurality of items. A plant monitoring method comprising a step of specifying and displaying at least one item as an increase factor in the order of contribution.   A plant monitoring program for causing a computer to execute the plant monitoring method according to claim 1. An operation data acquisition unit that acquires operation data including a plurality of items indicating the state of the plant;
A total Mahalanobis distance calculation unit that calculates the Mahalanobis distance of the operation data for the recorded normal state of the plant as a total Mahalanobis distance;
Using the orthogonal table having two levels indicating whether each of the plurality of items is a factor and each of the plurality of items is used for calculation of the Mahalanobis distance, the most reduced Mahalanobis distance among the plurality of items A reduction factor candidate identifying unit that identifies items that contribute to a reduction factor candidate;
A reduction factor setting unit that sets the reduction factor candidate as a reduction factor when it is determined that the reduction factor candidate has a large contribution to the Mahalanobis distance based on a set criterion;
A corrected Mahalanobis distance calculation unit for calculating the Mahalanobis distance as a corrected Mahalanobis distance for the corrected operation data that is the operation data related to the correction item excluding the reduction factor from the plurality of items,
A plant monitoring device comprising: a plant state determination unit that determines whether or not the state of the plant is abnormal by comparing the corrected Mahalanobis distance with a predetermined reference. A plant monitoring apparatus according to claim 10, wherein
The reduction factor identifying unit repeatedly identifies an item that contributes most to the decrease in Mahalanobis distance at predetermined time intervals. A plant monitoring device according to claim 10 or 11,
The predetermined reference is set with the determination time as the end from a start time that is a time preceding the determination time by which the plant state determination unit determines whether or not the state of the plant is abnormal. A plant monitoring device that is determined based on an average value of the modified Mahalanobis distance in an averaging section that is a section between the end time when the excluded section is started. A plant monitoring apparatus according to claim 12, comprising:
Each of a plurality of averaging intervals is set as the averaging interval,
Each of a plurality of predetermined criteria is set as the predetermined criterion,
The plant state determination unit determines which of the plurality of predetermined standards the corrected Mahalanobis distance exceeds. A plant monitoring device according to claim 12 or 13,
The predetermined criterion indicates that the corrected Mahalanobis distance has reached a predetermined number of times an average value in the averaging section at the determination time. A plant monitoring device according to claim 12 or 13,
The predetermined criterion indicates that a difference between the corrected Mahalanobis distance and the total Mahalanobis distance has reached a predetermined multiple of an average value in the averaging section at the determination time. The plant monitoring device according to any one of claims 10 to 15,
Further, a correlation matrix generation unit that generates a correlation matrix for the correction item of the operation data in the normal state;
An inverse matrix calculation unit for obtaining an inverse matrix of the correlation matrix by singular value decomposition,
The corrected Mahalanobis distance calculation unit calculates the Mahalanobis distance of the corrected operation data as the corrected Mahalanobis distance using the correlation matrix and the inverse matrix,
The plant monitoring device further includes
A singular value calculation unit for calculating a singular value by performing singular value decomposition on the correlation matrix;
A plant monitoring apparatus comprising: an under-singular value determination unit that outputs a warning when an under-singular value smaller than a predetermined threshold exists in the singular value. The plant monitoring device according to claim 16, further comprising:
Each of the correction items is a factor, and an orthogonal table having two levels indicating whether or not each of the correction items is used for calculation of the Mahalanobis distance is used to contribute to an increase in the Mahalanobis distance among the plurality of items. A plant monitoring apparatus comprising a lifting factor specifying unit that specifies and displays at least one item as a lifting factor in order of contribution.

Images