JP6641678B2 - Fault diagnosis device - Google Patents

Description

  The present invention relates to an abnormality diagnosis device that diagnoses an abnormality.

  In recent years, the use of the MT method (Maharanobis-Taguchi method), which is a type of MT system, has become widespread for diagnosis of abnormalities. In the MT method, first, a reference point in a unit space formed by normal data which is a plurality of parameter values measured in a normal state from a device or system to be diagnosed is determined. Thereafter, the values of the same parameters are measured at the time of the abnormality diagnosis, and the abnormality is diagnosed using the distance between these values and the reference point obtained from the normal data.

  This MT method is a useful method in that it can be used in various fields. On the other hand, there are abnormalities that are difficult to detect by the MT method. For example, when there is multicollinearity (redundancy or strong correlation) between a plurality of different parameter values, an abnormality in which the correlation between these parameters is broken is hard to be detected. Further, when the values of the parameters are constant during normal operation (a constant value item), an abnormality in which the values of these parameters have changed from the values of the parameters at the time of diagnosis at the time of diagnosis cannot be detected theoretically.

  In addition, a reconstruction error is also attracting attention as a technique used for diagnosing abnormalities (for example, see Non-Patent Document 1). When a reconstruction error is used, an abnormality can be detected when there is a break in multicollinearity or a change in a fixed value item, but not all types of abnormalities can be detected. Here, specifically, the multicollinearity and the constant value item are characteristics of data, and are not abnormal per se, but the target to be detected as abnormal is a change in characteristics of data.

Takehisa Yairi and 3 others, "Spacecraft Telemetry Monitoring Method by Dimension Reduction and Clustering", Proceedings of Japan Aerospace Laboratory, Vol. 59, No. 691, 197-205, August 2011

  As described above, in the abnormality diagnosis using the Mahalanobis distance and the abnormality diagnosis using the reconstruction error, there are types of abnormalities that are hard to detect, and there has been a problem that the detection accuracy of the abnormalities is reduced due to these types.

  In view of the above problems, an object of the present invention is to improve detection accuracy in abnormality diagnosis.

In order to achieve the above object, a first aspect of the present invention is an abnormality diagnostic apparatus for diagnosing an abnormality in which a plurality of parameter values to be diagnosed are inputted, and which is formed by a plurality of parameter values measured in a normal state. Of the normal data matrices to be obtained, the parameters of constant value items whose values are constant are replaced with 0 and normalized to obtain a correlation matrix, singular value decomposition is calculated from the obtained correlation matrix, and all predetermined singular values are determined. Preprocessing means for obtaining a pseudo-inverse matrix from the upper singular values of the number of upper singular values obtained from the ratio of the cumulative value of the upper singular values to the total sum of the parameters, and a diagnosis formed by a plurality of values of each parameter measured at the time of diagnosis The normal value is subtracted from the data matrix, and the column corresponding to the parameter of the constant value item is not divided by the standard deviation during normal operation, and the other columns are normalized by dividing by the standard deviation during normal operation. A first calculating means for calculating Mahalanobis distance by using the diagnostic data obtained and the pseudo-inverse matrix obtained by the preprocessing means; and a diagnostic data normalized by the first calculating means. Second calculating means for calculating a reconstruction error of each vector constituting the diagnostic data, and comparing the Mahalanobis distance calculated by the first calculating means with a predetermined first threshold value, and calculating by the second calculating means. The reconstruction error is compared with a predetermined second threshold, and when at least the Mahalanobis distance calculated by the first calculation means is larger than the predetermined first threshold, the reconstruction error calculated by the second calculation means is determined by a predetermined value. And determining means for determining that an abnormality has occurred in any of the cases where the value exceeds the second threshold value.

  ADVANTAGE OF THE INVENTION According to this invention, the detection accuracy of abnormality diagnosis can be improved.

FIG. 1 is a block diagram illustrating an abnormality diagnosis device according to the embodiment. FIG. 2 is a graph displaying the degree of abnormality (the Mahalanobis distance and the reconstruction error) calculated by the abnormality diagnostic device of FIG. FIG. 3 is a diagram illustrating information monitored by each of the Mahalanobis distance and the reconstruction error in the abnormality diagnosis device of FIG. FIG. 4 is a diagram illustrating an example of a diagnosis result displayed by the abnormality diagnosis device of FIG. FIG. 5 is a diagram illustrating the relationship between the Mahalanobis distance and the reconstruction error.

  Hereinafter, an abnormality diagnosis device according to an embodiment of the present invention will be described with reference to the drawings. The abnormality diagnosis device according to the embodiment diagnoses an abnormality of a diagnosis target using each parameter value for a diagnosis target such as a gas turbine, a vacuum furnace, and an aero engine where each parameter value is measured by a large number of sensors. .

  As shown in FIG. 1, an abnormality diagnosis device 1 according to the embodiment includes a preprocessing unit 11 that calculates a pseudo inverse matrix and a principal component matrix to be used for a diagnosis process by using values of normal parameters, A first calculating means 12 for calculating a Mahalanobis distance (MD value) using the value of the above, a second calculating means 13 for calculating a reconstruction error (RE value) using the value of each parameter, and The apparatus includes a determination unit that determines whether there is an abnormality using the Mahalanobis distance and the reconstruction error, and an output unit that outputs a determination result.

  As shown in FIG. 1, the abnormality diagnosis device 1 is an information processing device including a CPU 10, a storage device 20, an input device 30 used for inputting an operation, an output device 40 used for outputting a processing result, and the like. is there. When the abnormality diagnosis program P stored in the storage device 20 is executed, the CPU 20 executes processing as the preprocessing unit 11, the first calculation unit 12, the second calculation unit 13, the determination unit 14, and the output unit 15.

  The storage device 20 stores, in addition to the abnormality diagnosis program P, normal data D1 used in the preprocessing, operation data D2 obtained as a result of the preprocessing and used in the operation of the abnormality diagnosis, and measured during the target period of the abnormality diagnosis. Diagnostic data D3, which is obtained data, and result data D4, which is a result of abnormality diagnosis.

  The normal data D1 is each parameter value in a normal state. Here, the normal data D1 includes a plurality of measured parameter values in a normal state. The diagnostic data D3 is each parameter value measured during the diagnosis target period. The diagnostic data D3 also includes a plurality of parameter values measured during the target period. The abnormality diagnosis device 1 diagnoses an abnormality of a diagnosis target using the normal data D1 and the diagnosis data D3. For example, the abnormality diagnosis device 1 is connected to sensors that measure each parameter of a device or system to be diagnosed, and accumulates parameter values input from the plurality of sensors to generate normal data D1 and diagnosis data D3. Can be.

  The preprocessing unit 11 reads out the normal data D1 from the storage device 20 at the timing of executing the preprocessing, and obtains a pseudo inverse matrix and a principal component matrix used for the diagnosis process. Specifically, it is assumed that the number of parameters (the number of sensors) used in the abnormality diagnosis is k, and the normal data D1 includes each parameter value for n times of measurement. Let x be a matrix having a row of measured unit data samples, and X be a normalized unit data matrix.

  First, the preprocessing unit 11 obtains a normalized normal data matrix X using Expression (1-1). The normal data matrix X has n rows and k columns. Conventionally, the column of the parameter value of the constant value item has been removed from the unit data, but the abnormality diagnosis device 1 does not need to remove the column of the constant value item.

  Subsequently, the preprocessing means 11 calculates the correlation matrix of the normalized normal data matrix X by using the equation (1-3).

^{R = (X T X) /} n ··· (1-3)
Specifically, the correlation matrix obtained by Expression (1-3) is a pseudo-correlation matrix. The calculated matrix component is a correlation coefficient for a parameter that is not a constant value item, and is 0 for a parameter of a constant value item.

  Next, using the singular value decomposition, R is expressed by Expression (1-4).

R = USU ^T ··· (1-4)
Here, U is an orthonormal matrix having the main components of R in columns. S is a diagonal matrix having a singular value of R as a diagonal component.

Thereafter, the preprocessing means 11 calculates a diagonal matrix _Sq having q higher-order singular values of R as diagonal components from S. Here, the number q of the upper singular values is obtained from “the ratio of the cumulative value of the upper singular values to the sum total of all the singular values” input through the input device 30 in advance. For example, it is assumed that a condition is defined in which the sum of upper singular values and the number of singular values whose sum of all singular values exceeds 99% is set as the number of upper singular values. In this case, suppose that the number of all singular values is 1000 (singular values arranged in descending order are a1, a2,..., A1000, respectively), The ratio ((a1 + a2 + ... + a134) / (a1 + a2 + ... + a1000)) is 98.5%, and the ratio of the sum of the singular values up to the 135th in descending order and the sum of all the singular values ((a1 + a2 + ... + a135) / (a1 + a2 + ... + a1000) )) Is 99.1%, the upper-order singular value number q is 135.

Further, the preprocessing unit 11 obtains a pseudo inverse matrix R _q ^-1 of R using Expression (1-6).

  The first calculation means 12 reads the operation data D2 and the diagnosis data D3 from the storage device 20 at the timing of executing the diagnosis processing, and obtains the Mahalanobis distance (MD value). When the number of parameters (the number of sensors) used in the abnormality diagnosis is k, the diagnostic data D3 includes m parameter values each, and the matrix of the diagnostic data is y, first, the first calculation unit 12 Using the equation (2-1), a normalized diagnostic data matrix Y is obtained.

After that, the first calculating unit 12 uses the diagnostic data matrix Y to calculate the Mahalanobis distance MD ² (Y _i ) of each vector Y _i of the diagnostic data matrix Y according to Expression (2-2).

The second calculation means 13 reads the operation data D2 and the diagnosis data D3 from the storage device 20 at the timing of executing the diagnosis processing, and obtains a reconstruction error (RE value). Specifically, the second calculating means 13, by the equation (3-1), obtaining the data Y ^p reconstructed.

続いて、第２算出手段１３は、式（３−２）を用いて再構成誤差ベクトルＹ^hを求める。

Subsequently, the second calculation means 13 obtains the reconstructed error vector Y ^h using Equation (3-2).

^{Y h = Y-Y p ···} (3-2)

In addition, the second calculating means 13 calculates the reconstruction error RE ² (Y _i ) of each vector Y _i of the diagnostic data matrix Y by using the obtained reconstruction error vector Y ^h and Expression (3-3). .

  The judging means 14 compares each Mahalanobis distance calculated by the first calculating means 12 with a predetermined first threshold value. If the Mahalanobis distance is equal to or more than the first threshold value as shown in FIG. It is determined that an abnormality has occurred. Further, the determination unit 14 compares each reconstruction error calculated by the second calculation unit 13 with a predetermined second threshold, and as shown in FIG. 2B, the reconstruction error is equal to or larger than the second threshold. In this case, it is determined that an abnormality has occurred in the diagnosis target.

  The first threshold value is a value determined in advance for detecting an abnormality of a diagnosis target from the Mahalanobis distance. The second threshold value is a value that is determined in advance to detect an abnormality of a diagnosis target from a reconstruction error. The first threshold value and the second threshold value may be determined and set by the pre-processing unit 11 at the timing of pre-processing, or may be set by values input via the input device 30.

  The determination unit 14 detects an abnormality by combining the determination result using the Mahalanobis distance and the determination result using the reconstruction error. Specifically, if it is determined that an abnormality has occurred in any of the determination results, it is determined that an abnormality has occurred in the diagnosis target, and if it is determined that no abnormality has occurred in any of the determination results, the abnormality has occurred. It is determined that there is not. In addition, the determination unit 14 stores the determination result in the storage device 20 as the result data D4.

The output unit 15 outputs the result of the determination by the determination unit 14 to the output device 40. For example, the output unit 15 outputs a result using a graph as shown in FIG . As shown in FIG. 4A, abnormal data that cannot be detected by the Mahalanobis distance can be detected by a reconstruction error as shown in FIG. 4B .

For example, when there is multicollinearity between parameters or when there is a parameter of a fixed value item, the normal data D1 is distributed in a linear subspace smaller than the number of dimensions k of the parameter. When the singular value decomposition is performed, the upper principal component becomes a basis of the subspace, and the lower principal component becomes a basis of a space orthogonal to the subspace. Therefore, as shown in FIG. 5, the Mahalanobis distance (MD value) Y ^p based on the pseudo inverse matrix means the distance between the distribution of the normal data D1 and the diagnostic data D3 in the space where the normal data D1 is distributed. I do. In contrast, the reconstruction error (RE value) Y ^h is a subspace normal data D1 is distributed means the vertical distance between the diagnostic data D3. When an abnormality occurs, one or both of the Mahalanobis distance and the reconstruction error increase depending on the type of the abnormality.

  FIG. 5 is a diagram illustrating the geometric meaning of the Mahalanobis distance (MD value) and the reconstruction error (MD value) when k = 3. In FIG. 5, it is assumed that the dimension of the subspace in which the normal data D1 is distributed is 2, and the dimension of the subspace orthogonal thereto is 1. When the distance between the normalized diagnostic data matrix Y and the normal state is divided into two components, the distance in the subspace in which the normal data is distributed is represented by the Mahalanobis distance, and the subspace in which the normal data is distributed Is represented by a reconstruction error.

  For example, when the parameter value that has been constant during the normal period (the parameter value of the constant value item) changes, or when the correlation that appears in the lower principal component of the correlation matrix is broken, the reconstruction error increases. On the other hand, when the correlation appearing in the upper principal component of the correlation matrix is broken, or when the state of the parameter value to be diagnosed deviates from the equilibrium point included in the normal data due to aging, climate change, or the like, Mahalanobis The distance increases.

As described above, in the abnormality diagnosis apparatus 1, as shown in FIG. 3, an abnormality that cannot be detected using the Mahalanobis distance (MD value) is detected using the reconstruction error (RE value). Can be. Therefore, by using the Mahalanobis distance and the reconstruction error together, it is possible to detect all the abnormalities, and the abnormality detection efficiency is improved.
As described above, the present invention has been described in detail with reference to the embodiments. However, the present invention is not limited to the embodiments described in this specification. The scope of the present invention is determined by the description of the claims and the scope equivalent to the description of the claims.

Reference Signs List 1 abnormality diagnosis apparatus 11 preprocessing means 12 first calculation means 13 second calculation means 14 determination means 15 output means 20 storage device D1 normal data D2 operation data D3 diagnosis data D4 result data P abnormality diagnosis program 30 input device 40 output device

Claims (2)

An abnormality diagnosis device in which a plurality of parameter values to be diagnosed are input and an abnormality is diagnosed,
In the normal data matrix formed by the values of the parameters for a plurality of times measured in the normal state, the parameters of the constant value items whose values are constant are replaced with 0 and normalized to obtain the correlation matrix, and the obtained correlation matrix Preprocessing means for calculating a singular value decomposition from, and calculating a pseudo-inverse matrix from the upper singular values of the number of upper singular values obtained from the ratio of the cumulative value of the upper singular values to the sum of all predetermined singular values,
Subtract the average value at normal time from the diagnostic data matrix formed by the values of each parameter for multiple measurements measured at the time of diagnosis, and in the normal condition, do not divide the column corresponding to the parameter of the constant value item by the standard deviation, other columns Is normalized by dividing by the normal time standard deviation, the first calculating means for calculating the Mahalanobis distance using the normalized diagnostic data and the pseudo-inverse matrix obtained by the preprocessing means,
A second calculating unit that calculates a reconstruction error of each vector constituting the normalized diagnostic data using the diagnostic data normalized by the first calculating unit;
The Mahalanobis distance calculated by the first calculating means is compared with a predetermined first threshold, and the reconstruction error calculated by the second calculating means is compared with a predetermined second threshold, and at least the first calculating means calculates It is determined that an abnormality has occurred when the Mahalanobis distance is larger than a predetermined first threshold value or when the reconstruction error calculated by the second calculation means is larger than a predetermined second threshold value. Means for determining
An abnormality diagnosis device comprising: The abnormality diagnosis device according to claim 1 , further comprising an output unit that outputs a result of the determination by the determination unit to an output device.

Images