JP7009961B2 - Abnormality diagnosis device and abnormality diagnosis method - Google Patents

Description

The present invention relates to a technique for diagnosing an abnormality by analyzing waveform data (time series data) such as vibration and sound during operation of a diagnosis target (equipment / equipment, etc.).

In the abnormality diagnosis method of Patent Document 1, time-series waveform data is divided into short time intervals, Fourier transform is performed to make a large number of samples of multivariate data of frequency components, and a large number of samples are measured in advance during a period in which the device is considered normal. Prepare an eigenvector (also called a rotation matrix) for performing principal component analysis on the multivariate sample of.

At the time of diagnosis, the measurement data is Fourier transformed in the same manner as described above to create a multivariate data sample of the frequency component, which is transformed by the prepared rotation matrix to obtain the principal component score, which is normally obtained. Abnormality diagnosis is performed by comparing with the principal component score of. In other words, instead of independently analyzing the principal components for each diagnosis, the results of the principal component analysis performed in advance are used, and conversion is performed based on the same criteria at the time of diagnosis to facilitate comparison.

However, when trying to make a long-term diagnosis using a rotation matrix prepared in advance with a multivariate sample based on the Fourier transform, a lot of data is required for the preliminary analysis. For example, in the diagnosis of a rotating machine, a frequency resolution of about 1 Hz is required at least in a low frequency region.

For this purpose, each sample needs to be measured for at least 1 second, and even if the number of samples is increased over time, the samples are essentially similar and lack diversity, so diagnosis is made based on that. Then, in the long run, it becomes impossible to distinguish between abnormalities and mere changes over time.

Therefore, a method for diagnosing anomalies in Patent Document 2 using a constant Q transform instead of a Fourier transform has been proposed. According to this abnormality diagnosis method, a preparation phase and a diagnosis phase are executed. According to this preparation phase, a multivariate sample is created by constant Q conversion of the time series data measured in advance from the diagnosis target of the abnormality diagnosis, and diagnostic parameters are created based on the created multivariate sample. ..

In addition, according to the diagnosis phase, a multivariate sample is newly created by constant Q conversion of the time series data measured from the diagnosis target of the abnormality diagnosis, and the created multivariate sample is abnormal using the diagnostic parameters. Is diagnosed. As the component calculation method for the constant Q conversion, for example, the methods described in Non-Patent Documents 1 to 4 are used.

Japanese Patent No. 3382240 Japanese Unexamined Patent Publication No. 2017-198620

judith c. brown: calculation of a constant q spectral transform, j. acost. soc. am. 89 (1): 425-434, 1991 Judith C.I. Brown and Miller S.A. Puckette: An effective algorithm for the calculation of a constant Q transform, J. Mol. Acoustic. Soc. Am. 92 (5): 2698-2701, 1992 yukara_13: [Python] Constant-Q conversion (logarithmic frequency spectrogram), super-introduction to music programming (provisional) inHatena Blog, 2013-12-01, 2013, Internet <URL: http: // yukara-13. hatenablog. com / entry / 2013/12/01/222742>. [2016-02-22 Access] yukara_13: [Python] Fast Fourier-Q conversion (with FFT), super-introduction to music programming (provisional) inHatena Blog, 2014-01-05, 2014, Internet <URL: http: // yukara-13. hatenablog. com / entry / 2014/01/05/062414>. [2016-02-22 Access]

The abnormality diagnosis method of Patent Document 2 handles all the time-series data in which the diagnosis target is operating during the learning period of the preparation phase in the same row. Certainly, if the operation of the diagnosis target is uniform, there is no problem in treating all the time series data in the same row, but in reality, there are few such diagnosis targets.

FIG. 1 shows a graph in which the generated power for each hour is plotted with a certain generator as a diagnosis target, and it can be seen from the graph that an output difference of more than double occurs depending on the time.

If the operation level is so different, the waveform of the time series data (vibration, sound, current, etc.) measured at that time may differ greatly. Therefore, if the time-series data are collectively learned, the normal range may become too wide and it may be difficult to detect an abnormality at an early stage.

2 (a) to 2 (j) show that in the generator having the generated power of FIG. 1, the learning period L is for the first week in the hourly vibration data (time series data) at the time of power generation. Principal component analysis was performed by taking a constant Q frequency component, and a scatter diagram of "first principal component x second principal component" (principal component distribution: the first principal component score is the X-axis, the second is the Y axis. (Shown on the axis) is shown.

“P” in FIGS. 2 (a) to 2 (j) indicates the overall main component distribution, and “Q” in FIGS. 2 (b) to (j) indicates the main component distribution Q for each level of the generator. ing. Here, the level of the generator is divided into "Lv1, Lv2 ..." for each "0.2" from the generated power "0.8" or more, and the generated power "2.4" or more is regarded as "Lv9". There is. At this time, according to FIGS. 2A to 2J, the overall principal component distribution P has a heart shape, but the principal component distribution Q for each level of generated power changes from right to left as the level increases. It is changing.

Even if time-series data distributed on the left side of FIG. 2 (b) is obtained when the level of generated power is "Lv1" with such a distribution, what should be clearly recognized as an abnormality from the viewpoint of statistical abnormality diagnosis is In the method of Patent Document 1, since all the operating states are treated in the same line, there is a problem that they are recognized as normal.

The present invention has been made to solve such a conventional problem, and it is intended to learn and diagnose a diagnosis target whose driving level changes significantly by classifying the diagnosis target by using an index indicating the driving level of the diagnosis target. It is a solution issue.

(1) One aspect of the present invention is
A diagnostic parameter creation means that creates diagnostic parameters by learning a learning sample based on time-series data measured from the diagnosis target of abnormal diagnosis in advance.
A diagnostic means for diagnosing an abnormality in the diagnostic target using the diagnostic parameters for the time-series data newly measured from the diagnostic target.
It is an abnormality diagnostic device equipped with
The diagnostic parameter creating means divides the learning sample into levels according to the operation information associated with the time-series data, learns each level, and creates diagnostic parameters.
The diagnostic means calculates a level based on the operation information associated with the time-series data, and diagnoses the abnormality of the diagnosis target by using the diagnostic parameter corresponding to the level.

(2) Another aspect of the present invention is
It is a method of diagnosing abnormalities performed by a computer.
A diagnostic parameter creation step that creates diagnostic parameters by learning a learning sample based on time-series data measured from the diagnosis target of abnormal diagnosis in advance, and
It has a diagnostic step of diagnosing an abnormality of the diagnostic target using the diagnostic parameters for the time-series data newly measured from the diagnostic target.
In the diagnostic parameter creation, the learning sample is divided into levels according to the operation information associated with the time-series data, and learning is performed for each level to create diagnostic parameters.
In the diagnosis step, a level based on the operation information associated with the time series data is calculated, and the abnormality of the diagnosis target is diagnosed by using the diagnostic parameter corresponding to the level.

According to the present invention, it is possible to learn and diagnose a diagnosis target whose driving level changes significantly by classifying the diagnosis target using an index indicating the driving level of the diagnosis target.

A graph showing the amount of power generated by a generator every hour. (A) is a graph showing the main component distribution of the entire operation level (Lv) by taking the constant Q frequency component with the first week as the learning period for the generator in FIG. 1, and (b) is the main component of the same Lv1. A graph showing the distribution, (c) is a graph showing the main component distribution of the same Lv2, (d) is a graph showing the main component distribution of the same Lv3, (e) is a graph showing the main component distribution of the same Lv4, (f). Is a graph showing the main component distribution of the same Lv5, (g) is a graph showing the main component distribution of the same Lv6, (h) is a graph showing the main component distribution of the same Lv7, and (i) is the main component distribution of the same Lv8. The graph which shows, (j) the graph which shows the principal component distribution of the same Lv9. The block diagram of the abnormality diagnosis apparatus which concerns on 1st Embodiment of this invention. The schematic diagram which shows the level classification of the driving information and the range of learning data for each level. The block diagram of the abnormality diagnosis apparatus which concerns on 2nd Embodiment of this invention. A graph showing an example of diagnosis results when a certain generator is not classified into levels. A graph showing an example of diagnosis results when the same level is divided. A graph showing an example of diagnosis results when the learning ranges are overlapped by dividing them into the same levels. A graph showing an example of two-dimensional leveling.

Hereinafter, the abnormality diagnostic apparatus according to the embodiment of the present invention will be described. This abnormality diagnosis device uses constant Q conversion as in Patent Document 1 when generating a multivariate sample of frequency components from time-series data of a diagnosis target (for example, a generator) for abnormality diagnosis.

This constant Q transform does not analyze data for the same period in all frequency bands as in the Fourier transform, but refers to each frequency so that the number of cycles is the same in all frequency bands. Analyze by changing the number of data.

In this method, the higher the frequency band, the shorter the period of data is analyzed. Therefore, when a long period (1 second or more) of measurement is required to increase the frequency resolution in the low frequency band (for example, about 1 Hz), the period is long. Since the periods do not overlap in the high frequency band even if a large number of samples are taken by overlapping, diversity is ensured as a multivariate sample. As the component calculation method for the constant Q conversion, the methods described in Non-Patent Documents 1 to 4 can be used as in Patent Document 2.

<< Example 1 >>
The first embodiment of the abnormality diagnosis apparatus will be described with reference to FIG. The abnormality diagnosis device 1 is configured by a computer and includes hardware resources of a normal computer (for example, a main storage device such as a CPU, ROM, and RAM, an auxiliary storage device such as an HDD or SSD). As a result of the collaboration between the hardware resource and the software resource (OS, application, etc.), the parameter creating means 100 for executing the preparation phase and the diagnostic means 200 for executing the diagnostic phase are implemented.

The parameter creating means 100 and the diagnostic means 200 are basically configured in the same manner as in Patent Document 2, and execute the same processing operation as in Patent Document 2. That is, the parameter creating means 100 constantly measures (measures) the time-series data D1 collected from the diagnosis target of the abnormality diagnosis, converts the time-series data D1 into a constant Q, and creates a multivariate sample (learning sample) to be learned. The conversion unit 101, the principal component analysis unit 102 that analyzes the created multivariate sample for principal component analysis to obtain a conversion matrix, and calculates the principal component score of each multivariate sample, and the multivariate sample based on the principal component score. It is provided with a statistic value calculation unit 103 for obtaining a statistic as an index of the above, and a normalization unit 104 for normalizing the statistic to obtain an abnormality degree.

Further, the diagnostic means 200 is also obtained by the constant Q conversion unit 201 that constantly Q-converts the newly measured time-series data D2 to create a multivariate sample, and the principal component analysis unit 102 from the created multivariate sample. The principal component calculation unit 202 that obtains the principal component score using the obtained conversion matrix, the second statistical value calculation unit 203 that obtains the statistic based on the principal component score, and the normalization unit 104 use the statistic. It is provided with a normalization unit 204 that normalizes with the normalized normalization coefficient and calculates the degree of abnormality, and diagnoses the abnormality of the abnormality diagnosis target based on the degree of abnormality.

However, the abnormality diagnosis device 1 is provided with operation information for determining the operation level for each time series data D1 and D2. Here, it differs from Patent Document 2 in that each of the preparation phase and the diagnosis phase is divided into levels according to the operation level classified based on the operation information.

(1) Preparation phase In the preparation phase of the abnormality diagnosis device 1, the time series data D1 is subjected to constant Q conversion by the constant Q conversion unit 101 to create a multivariate sample, and the multivariate sample is tagged with operation information (1). (Association), and the subsequent processing is carried out separately for each operation level based on the operation information and learned.

This leveling shall be performed based on the value of the operation information tagged with the time series data D1. As the operation information, for example, when the time-series data D2 of the vibration measured by the bearing of the generator is diagnosed by the diagnostic means 200, the generated power measured at the same timing can be used. Further, as an example of leveling, "Lv01 to Lv18" in FIG. 1 show leveling according to the generated power of the generator.

Furthermore, if there is a measured amount that can determine the operation level at that time, such as the amount of drainage for a drainage pump and the amount of air blown for a blower, the measured amount can be used as operation information. As the operation information when diagnosing from the time-series data of the measured vibration waveform, a method may be used in which a numerical value calculated from the time-series data itself such as an effective value of the time-series data of the measured vibration waveform is used.

Then, diagnostic parameters are created for the multivariate samples divided according to the operation level by the same method as in Patent Document 2. That is, in the principal component analysis unit 102, the multivariate sample is subjected to principal component analysis for each operation level to obtain a transformation matrix, and the principal component score of each multivariate sample is calculated. Further, the statistical value calculation unit 103 acquires an index statistic such as the Hotelling T2 / Q statistic based on the principal component score calculated for each operation level, and the normalization unit 104 obtains the statistic. Normalize to the degree of anomaly.

As a result, according to the preparation phase of the abnormality diagnosis device 1, diagnostic parameters (transformation matrix and normalization coefficient) are created for each operation level. Each diagnostic parameter created here is stored in the storage device or the like.

(2) Diagnosis phase In the diagnosis phase of the abnormality diagnosis device 1, first, the constant Q conversion unit 201 performs constant Q conversion of the time series data D2 to create a multivariate sample. Next, the operation level is calculated based on the value of the operation information tagged in the time series data D2.

Diagnostic parameters (transformation matrix and normalization coefficient) corresponding to the operation level are acquired from the storage device, and diagnosis is performed on the time series data D2 using the acquired diagnostic parameters. As this diagnostic method, for example, the method of Patent Document 2 can be used.

To explain the outline, the principal component calculation unit 202 calculates the principal component score using the transformation matrix corresponding to the operation level, and based on this, the statistical value calculation unit 203 calculates the statistical value by the same calculation as in the preparation phase. get. Finally, the normalization unit 204 calculates the degree of abnormality using the normalization coefficient corresponding to the operation level, and diagnoses whether or not the diagnosis target is abnormal based on the value of the calculation result.

<< Example 2 >>
The second embodiment of the abnormality diagnosis apparatus will be described with reference to FIG. Here, the device configuration of the second embodiment is the same as that of the first embodiment.

(1) In the preparation phase of Example 1, since the multivariate sample used for learning the operation level is limited to the one obtained from the time series data D1 belonging to the operation level, each operation level is exclusively independent. is doing. On the other hand, in the preparation phase of the second embodiment, the difference is that the range of the multivariate sample used for learning is expanded by using the value of the operation information.

That is, although the abnormality diagnosis device 1 of the first embodiment diagnoses by leveling the multivariate sample of the time series data to be diagnosed by the operation level information, the leveling is exclusive and one. Multivariate samples cannot belong to multiple levels.

Therefore, when the same set of multivariate samples is used for learning, the more subdivided the levels, the smaller the number of multivariate samples belonging to each level, and if the subdivision is too much, proper learning becomes difficult. There is a limit to the subdivision of.

For example, as shown in FIG. 1, when the generated power is divided into levels from "0.8" to "Lv01 to Lv18" in increments of "0.1", there are several levels of only one or two multivariate samples during the learning period. It will come out. Therefore, in the second embodiment, in the level division subdivided in this way, the range of the learning samples (multivariate samples) of each operation level is expanded to secure the number of learning samples.

(2) Since the operation information is one numerical information here, each operation level is determined by setting a threshold value for each operation level and determining the magnitude of the numerical value of the operation information. This leveling must be exclusive to uniquely determine the level during the diagnostic phase.

However, in the second embodiment, as shown in FIG. 4, it is determined that the value of the driving information is slightly wider than the range of each driving level for the range of the learning target (multivariate sample) at each driving level. For example, the learning target of the driving level 3 (Lv3) is expanded not only to the driving level 3 (Lv3) range determined from the driving information but also to the driving levels 2 (Lv2) and 4 (Lv4) continuous with the driving level 3. ..

As a result, the range of the learning target overlaps at each driving level, but the boundary with the non-driving range is not included in the learning target. In FIG. 4, the range is shown so that the same multivariate sample is not included in the learning target of the non-adjacent level (for example, Lv1 and Lv3), but such a limitation does not necessarily have to be provided.

<< Example 3 >>
The third embodiment of the abnormality diagnosis apparatus will be described with reference to FIG. Here, the range of the learning target is expanded in the same manner as in the second embodiment, but the range is fixed to the range of the adjacent operation level.

That is, in the second embodiment, the exclusive operation level is used for the diagnosis, while the overlap of the operation level is allowed for the extraction of the learning range. However, according to the second embodiment, it is necessary not only to divide the level for each operation level but also to reconstruct the value of the operation information to be slightly wider than the range of each level when extracting the multivariate sample used for learning. , Processing may be complicated.

Therefore, in the third embodiment, in order to simplify the extraction of the range of the learning target, the range of the learning target is fixed to the range of the adjacent driving level. For example, the learning target of the driving level "Lv1" is a multivariate sample of "Lv1" and "Lv2", the learning target of the driving level "Lv2" is a multivariate sample of "Lv1" to "Lv3", and the learning target of the driving level "Lv3". The learning target is advanced like a multivariate sample of "Lv2" to "Lv4", and the learning target of the final operation level "Lvn" is "Lv (n-1)" and "Lvn".

As a result, tagging of multivariate samples can be done not by operation information but by operation level calculated from operation information as shown in Fig. 5, and multivariate samples should be summarized by operation level. Can be implemented in. Therefore, when learning the preparation phase, if two or three adjacent multivariate sample groups of operation levels are merged and the processing of the principal component analysis unit 102, the statistical value calculation unit 103, and the normalization unit 104 is executed. good.

≪Action effect≫
According to Examples 1 to 3, by executing the learning of the preparation phase and the diagnosis of the diagnosis phase separately for each operation level, the spread of the principal component distribution of the preparation phase is divided for each operation level and becomes compact. It is possible to improve the sensitivity of abnormality detection. This point will be described with reference to FIGS. 6 to 8.

FIG. 6 shows the diagnosis result in which the first week of the hourly vibration data (time series data) of the power generation time in the generator of the generated power shown in FIG. 1 is set as the learning period L. There is. Here, in FIG. 1, the result of learning / diagnosing the generated power (Y-axis) “0.8” or more as operation without classifying the level, that is, the result of learning / diagnosing in Patent Document 2 is shown. The generator that acquired the vibration data in FIG. 1 stopped power generation for a while from the 21st day (see X-axis) of the start of operation, and after restarting power generation on the 25th day, it stopped urgently due to a failure about 2 hours later. ing.

FIG. 7 shows a diagnostic result in which the vibration data is divided into levels according to the generated power in the learning period L and learned without overlapping the levels, that is, the diagnostic result of Example 1. Here, as in FIG. 6, the results of learning and diagnosis in increments of "0.2" are shown with the generated power of "0.8" or more as the operation.

FIG. 8 shows the diagnostic results obtained by classifying the vibration data into levels according to the generated power in the learning period L and learning by the method of Example 3. Here, the results of learning / diagnosis are shown in which the generated power of "0.8" or more is divided into levels in increments of "0.1", expanded to the adjacent level of Example 3, and duplicated. There is.

According to the diagnosis results of FIGS. 6 to 8, if the levels of FIG. 6 are not divided, the learning period L varies widely, so caution is taken until the day of failure (25th day of the start of operation). I never crossed the line.

On the other hand, when the levels of FIGS. 7 and 8 are divided, the variation in the learning period L is suppressed as compared with FIG. 6, and the caution line is crossed on the 9th day of the start of operation. Therefore, according to the level classification of Examples 1 and 3, the abnormality sign can be detected from two weeks or more before the failure. In particular, according to Example 3, the variation is suppressed as compared with the case where the level is not subdivided as a whole except for the time when the caution line is crossed (FIG. 6).

Therefore, according to the level-divided learning of Examples 1 to 3, the range of the normal state can be limited, so that the sensitivity of abnormality detection is improved, and further, the level division can be subdivided by duplication of the learning objects of Example 3. The effect is further increasing.

≪Other / Other examples≫
The present invention is not limited to the above embodiment, and can be modified and implemented within the scope described in each claim. An example will be described below.

(1) In Examples 1 to 3, the operation information is one numerical information, and the level is divided according to the size. However, it is not always necessary to classify the levels from only one numerical value, as long as the multivariate samples to be learned can be appropriately classified. For example, if the diagnosis target is a wind turbine generator, two-dimensional leveling may be performed by combining them using the wind speed at that time in addition to the generated power.

In such leveling based on the operation information of a plurality of numerical values, the range of the learning target is expanded for each numerical value of the operation information as in FIG. 4, and a multivariate sample is created. Here, when the same adjacent level as in Example 3 is used, the number of the multivariate sample groups of the adjacent level to be merged is not two or three, but four if the two-dimensional leveling shown in FIG. 9 is used. (For corners), 6 (for edges), 9 (excluding corners and edges).

(2) Further, in the present invention, the multivariate sample to be learned may be separately learned and diagnosed for each operation level, and the learning target is not limited to the multivariate sample created by the constant Q transformation, and Fourier. It may be a multivariate sample by conversion or the like.

(3) Further, the learning method of the present invention is not limited to the principal component analysis, and various machine learning algorithms such as "One Class SVM" can be applied.

1 ... Abnormality diagnosis device 100 ... Preparation phase 101, 201 ... Constant Q conversion unit 102, 202 ... Principal component analysis unit 103, 203 ... Statistical value calculation unit 104, 204 ... Normalization unit 200 ... Diagnosis phase

Claims (2)

A diagnostic parameter creation means that creates diagnostic parameters by learning a learning sample based on time-series data measured from the diagnosis target of abnormal diagnosis in advance.
A diagnostic means for diagnosing an abnormality in the diagnostic target using the diagnostic parameters for the time-series data newly measured from the diagnostic target.
It is an abnormality diagnostic device equipped with
The diagnostic parameter creating means divides the learning sample into levels according to the operation information associated with the time-series data, learns each level, and creates diagnostic parameters.
The diagnostic means calculates a level based on the operation information associated with the time-series data, diagnoses the abnormality of the diagnosis target by using the diagnostic parameter corresponding to the level, and diagnoses the abnormality.
When creating the diagnostic parameters, the learning target is extended to the adjacent range of the level.
An abnormality diagnostic device characterized in that the learning targets are duplicated . It is a method of diagnosing abnormalities performed by a computer.
A diagnostic parameter creation step that creates diagnostic parameters by learning a learning sample based on time-series data measured from the diagnosis target of abnormal diagnosis in advance, and
It has a diagnostic step of diagnosing an abnormality of the diagnostic target using the diagnostic parameters for the time-series data newly measured from the diagnostic target.
In the diagnostic parameter creation step, the learning sample is divided into levels according to the operation information associated with the time-series data, and learning is performed for each level to create diagnostic parameters.
In the diagnosis step, a level based on the operation information associated with the time series data is calculated, and the abnormality of the diagnosis target is diagnosed using the diagnostic parameter corresponding to the level.
When creating the diagnostic parameters, the learning target is extended to the adjacent range of the level.
An abnormality diagnosis method characterized in that the learning targets are duplicated .

Images