JP2019101777A - Abnormality diagnosing apparatus and abnormality diagnosing method - Google Patents

Abstract

To perform learning and diagnosis on a diagnostic target whose driving level greatly changes by using an index indicating a driving level of the diagnostic target.SOLUTION: In a preparation phase, parameter creating means 100 of a diagnostic device creates a diagnostic parameter by learning based on time series data measured in advance from a diagnostic target of an abnormality diagnosis. At this time, a target of learning is divided into driving levels from driving information tagged to time series data, and diagnostic parameters are created for each driving level. In addition, diagnosis means 200 diagnoses an abnormality of a diagnostic target, using the diagnosis parameter with respect to time series data newly measured from the diagnostic target. Here, a driving level is calculated from the driving information tagged with the time series data, and the diagnosis parameter is used to diagnose an abnormality of the diagnostic target according to the driving level.SELECTED DRAWING: Figure 5

Description

  The present invention relates to a technology for analyzing abnormality and analyzing waveform data (time-series data) such as vibration and sound during operation of a diagnosis target (device, facility, etc.) and diagnosing abnormality.

  The abnormality diagnosis method of Patent Document 1 divides waveform data of time series at short time intervals, performs Fourier transform to create many samples of multivariate data of frequency components, and measures many in advance during a period when the device is considered normal. An eigenvector (also referred to as a rotation matrix) for principal component analysis of the multivariate sample of to obtain a principal component score is prepared.

  At the time of diagnosis, measurement data is Fourier-transformed in the same manner as described above to create multivariate data samples of frequency components, and this is transformed with a prepared rotation matrix to obtain principal component scores, which are normal An abnormality diagnosis is performed by comparing with the principal component score of In other words, instead of performing principal component analysis independently for each diagnosis, comparison is made easier by converting the same criteria at the time of diagnosis using results obtained by performing principal component analysis in advance.

  However, when trying to diagnose for a long time with a rotation matrix that has been created in advance using multivariate samples based on Fourier transform, a large amount of data is required for prior analysis. For example, in the diagnosis of a rotating machine, at least in the low frequency range, a frequency resolution on the order of 1 Hz is required.

  In order to do this, each sample needs to be measured at least for about 1 second, and even if the number of samples is increased over time, the sample is essentially similar and lacks diversity. Then, in the long run, it becomes impossible to distinguish between anomalous and mere change over time.

  Therefore, the abnormality diagnosis method of Patent Document 2 has been proposed which uses constant Q conversion instead of Fourier conversion. According to this abnormality diagnosis method, the preparation phase and the diagnosis phase are performed. According to this preparation phase, time series data measured in advance from a diagnosis target of abnormality diagnosis is subjected to constant Q conversion to create a multivariate sample, and diagnostic parameters are created based on the created multivariate sample .

  Further, according to the diagnosis phase, time series data measured from a new diagnosis target of abnormality diagnosis is subjected to constant Q conversion to create a multivariate sample, and an abnormality is generated using the diagnosis parameter for the created multivariate sample. Are diagnosed. In addition, as a component operation method of constant Q conversion, the method described in the nonpatent literature 1-4 is used, for example.

Patent No. 3382240 JP, 2017-198620, A

judith c. brown: calculation of a constant q spectral transform, j. acoust. soc. am. 89 (1): 425-434, 1991 Judith C. Brown and Miller S. Puckette: An efficient algorithm for the calculation of a constant Q transform, J. Am. Acoust. Soc. Am. 92 (5): 2698-2701, 1992 yukara_13: [Python] Constant-Q conversion (logarithmic frequency spectrogram), introduction to music programming (provisional) in Hatena Blog, 2013-12-01, 2013, Internet <URL: http: // yukara-13. hatenablog. com / entry / 2013/12/01/222742>. [2016-02-22 access] yukara_13: [Python] High-speed Constant-Q conversion (with FFT), introduction to music programming (provisional) in Hatena Blog, 2014-01-05, 2014, Internet <URL: http: // yukara-13. hatenablog. com / entry / 2014/01/05/062424>. [2016-02-22 access]

  The abnormality diagnosis method of Patent Document 2 handles all time-series data that the diagnostic target is operating in the learning period of the preparation phase in the same row. Certainly, if the operation to be diagnosed is uniform, it is safe to handle all 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 per hour is plotted for a certain generator to be diagnosed, and it is understood from the graph that an output difference of twice or more occurs depending on the time.

  When the operation level is so different, a large difference may occur in the waveforms of time-series data (vibration, sound, current, etc.) measured at that time. Therefore, if those time-series data are collectively learned, the normal range may become too wide, which may make early detection of abnormality difficult.

  2 (a) to 2 (j) show that, in the generator having the generated power of FIG. 1, a learning period L for the first week in vibration data (time-series data) for each hour of generating time is generated. The principal component analysis is performed by taking constant Q frequency components as a scatter plot of “first principal component × second principal component” (principal component distribution: first component score X axis, second component Y Shown on the axis).

  “P” in FIGS. 2 (a) to 2 (j) indicates the whole principal component distribution, and “Q” in FIGS. 2 (b) to 2 (j) indicates the principal component distribution Q according to the level of the generator. ing. Here, the level of the generator is divided into “Lv1, Lv2 ...” in every generated power “0.8” to “0.2”, and the generated power “2.4” or more is considered as “Lv 9” There is. At this time, according to FIGS. 2 (a) to 2 (j), the entire main component distribution P has a heart shape, but the main component distribution Q for each level of the generated power changes from right to left every time the level increases. It is changing.

  Even if time-series data distributed on the left side of Fig. 2 (b) is obtained when the generated power level is "Lv 1" in such a distribution, those that should be recognized as abnormal in terms of statistical abnormality diagnosis are In the method of Patent Document 1, there is a problem that it is recognized as normal since all the operating conditions are handled in the same line.

  The present invention is made in order to solve such conventional problems, and for diagnosis targets whose operation levels greatly change, learning / diagnosis is performed by using an index indicating the operation level of the diagnosis target. It is considered as a problem to solve.

(1) One aspect of the present invention is
Diagnostic parameter generation means for generating a diagnostic parameter by learning a learning sample based on time series data measured in advance from a diagnosis target of abnormality diagnosis;
Diagnostic means for diagnosing an abnormality of the diagnosis target using time-lapse data newly measured from the diagnosis target using the diagnostic parameter;
An abnormality diagnosis device provided with
The diagnostic parameter creation means divides the learning sample into levels according to the driving information linked to the time series data, learns each level, and creates a diagnostic parameter.
The diagnostic means calculates a level based on driving information linked to the time series data, and diagnoses an abnormality of the diagnostic object using the diagnostic parameter according to the level.

(2) Another aspect of the present invention is
An abnormality diagnosis method executed by a computer,
A diagnostic parameter creation step of creating a diagnostic parameter by learning a learning sample based on time series data measured in advance from a diagnosis target of abnormality diagnosis;
A diagnosis step of diagnosing an abnormality of the diagnosis target using time-lapse data newly measured from the diagnosis target using the diagnostic parameter;
In creating the diagnostic parameter, the learning sample is divided into levels according to the driving information linked to the time series data, and the diagnostic parameter is created by learning for each level.
In the diagnosis step, a level based on the driving information linked to the time series data is calculated, and an abnormality of the diagnosis object is diagnosed using the diagnostic parameter according to the level.

  ADVANTAGE OF THE INVENTION According to this invention, with respect to the diagnostic object to which a driving | operation level changes a lot, it can be divided into levels using the parameter | index which shows the operating level of a diagnostic object, and it can learn and diagnose.

A graph showing an hourly power generation amount of a certain generator. (A) is a graph showing the principal component distribution of the whole operation level (Lv), taking constant Q frequency components with the first week's worth of the generator in FIG. 1 as a learning period, (b) a principal component of the same Lv1 A graph showing the distribution, (c) a graph showing the principal component distribution of the same Lv2, (d) a graph showing the principal component distribution of the same Lv3, (e) a graph showing the principal component distribution of the same Lv4, (f) Is a graph showing the principal component distribution of the same Lv5, (g) a graph showing the principal component distribution of the same Lv6, (h) a graph showing the principal component distribution of the same Lv7, (i) a principal component distribution of the same Lv8 The graph which shows, (j) The graph which shows the principal component distribution of same Lv9. BRIEF DESCRIPTION OF THE DRAWINGS The block diagram of the abnormality-diagnosis apparatus based on 1st Embodiment of this invention. The schematic which shows the level division of the same driving information, and the range of the learning data for every level. The block diagram of the abnormality-diagnosis apparatus which concerns on 2nd Embodiment of this invention. The graph which shows the example of a diagnostic result when not dividing into a level about a certain generator. A graph showing an example of diagnostic results when divided into the same level. The graph which shows the example of a diagnostic result at the time of dividing into the same level and making a learning range overlap. The graph which shows the example of two-dimensional level division.

  Hereinafter, an abnormality diagnosis apparatus according to an embodiment of the present invention will be described. This abnormality diagnosis apparatus 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) of abnormality diagnosis.

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

  In this method, in order to analyze data in a shorter period as the high frequency band, it is necessary to measure a long period (1 second or more) in order to increase the frequency resolution in the low frequency band (for example, about 1 Hz) Because the periods do not overlap in the high frequency band even if multiple samples are overlapped, diversity is secured as a multivariate sample. In addition, the method described in the nonpatent literature 1-4 can be used like the patent document 2 as a component calculating method of constant Q conversion.

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

  The parameter creating unit 100 and the diagnostic unit 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 performs constant Q conversion on time series data D1 collected (measured) from the diagnosis target of abnormality diagnosis in advance to create a multivariate sample (learning sample) to be a target of learning. A transformation unit 101, a principal component analysis unit 102 which performs principal component analysis of the created multivariate samples to obtain a transformation matrix and calculates principal component scores of each multivariate sample, and multivariate sample based on the principal component scores A statistical value calculation unit 103 for obtaining a statistical value that is an index of the above-mentioned, and a normalization unit 104 for normalizing the statistical value to make it an anomaly degree.

  In addition, the diagnosis unit 200 is also obtained by the principal component analysis unit 102 from a constant Q conversion unit 201 that constant Q converts newly measured time series data D2 to create a multivariate sample, and the generated multivariate sample. The principal component calculation unit 202 for obtaining a principal component score using the transformation matrix, the second statistic value calculation unit 203 for obtaining a statistic based on the principal component score, and the normalization unit 104 using the statistic And a normalization unit 204 that calculates the degree of abnormality by normalization using the normalization factor, and diagnoses the abnormality of the abnormality diagnosis target based on the degree of abnormality.

  However, the abnormality diagnosis device 1 is added with operation information for determining the operation level for each of the time series data D1 and D2. This is different from Patent Document 2 in that each of the preparation phase and the diagnosis phase is divided into operation levels classified based on operation information and implemented.

(1) Preparation phase In the preparation phase of the abnormality diagnosis device 1, the constant Q conversion unit 101 performs constant Q conversion on the time series data D1 to create a multivariate sample, but the multivariate sample is tagged with operation information ( Stringing) and performing the subsequent processing separately for each driving level based on the driving information and learning.

  This level division shall be performed on the basis of the value of the operation information tagged with the time series data D1. As the operation information, for example, when diagnosing the time series data D2 of the vibration obtained by measuring the generator with the bearing by the diagnostic means 200, the generated power measured at the same timing can be used. Also, as an example of level division, “Lv01 to Lv18” in FIG. 1 indicates level division by the generated power of the generator.

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

  Then, for the multivariate samples divided according to the driving level, diagnostic parameters are created by the same method as in Patent Document 2. That is, the principal component analysis unit 102 performs principal component analysis of multivariate samples for each operation level to obtain a conversion matrix, and calculate principal component scores of each multivariate sample. Also, in the statistical value calculation unit 103, based on the principal component score calculated for each operation level, a statistic serving as an index such as a hotelling T2 / Q statistic is acquired, and the statistic is calculated in the normalization unit 104. Normalized to be anomalous.

  As a result, according to the preparation phase of the abnormality diagnosis device 1, diagnostic parameters (conversion matrix and normalization coefficient) are created for each driving 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 on the time series data D2 to create a multivariate sample. Next, the driving level is calculated based on the value of the driving information tagged to the time series data D2.

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

  To explain the outline, the principal component score is calculated by using the transformation matrix corresponding to the operation level in the principal component calculation unit 202, and based on this, the statistic value calculation unit 203 calculates the statistic value in the same calculation as 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 the diagnosis target is abnormal or not based on the value of the calculation result.

Example 2
A second embodiment of the abnormality diagnosis apparatus will be described based on 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, the multivariate samples used to learn the driving level are limited to those obtained from the time-series data D1 belonging to the driving level, so each driving level is exclusively independent. doing. On the other hand, the preparation phase of the second embodiment is different in that the range of multivariate samples used for learning is expanded using the value of the driving information.

  That is, although the abnormality diagnosis apparatus 1 according to the first embodiment diagnoses multivariate samples of time series data to be diagnosed by level division based on driving level information, the level division is exclusive, Multivariate samples can not belong to more than one level.

  Therefore, when using the same multivariate sample set for learning, as the level is subdivided, the number of multivariate samples belonging to each level decreases, and appropriate leveling becomes difficult after segmentation, so the level There is a limit to the segmentation of

  For example, as shown in FIG. 1, when dividing the generated power into “Lv01 to Lv18” in steps of “0.8” to “0.1”, some levels of only one or two multivariate samples in the learning period It will come out. So, in Example 2, in the level division subdivided in this way, the range of the learning sample (multivariate sample) of each driving | operation level is expanded, and the number of learning samples is ensured.

  (2) Here, since the driving information is one numerical value information, each driving level is determined based on the magnitude of the numerical value of the driving information by providing a threshold of each driving level. This leveling must be exclusive in order to determine the level uniquely during the diagnostic phase.

  However, in the second embodiment, as shown in FIG. 4, the range of the learning target (multivariate sample) at each driving level is determined so that the value of the driving information is slightly wider than the range of each driving level. For example, the learning object of the driving level 3 (Lv 3) is extended not only to the driving level 3 (Lv 3) range defined from the driving information but also to the driving level 2 (Lv 2) and 4 (Lv 4) range continuous with the driving level 3 .

  Although the range of a learning object overlaps by each driving | operation level by this, the boundary with the non-driving range shall be not included in a learning object. Although the range is shown in FIG. 4 so that the same multivariate sample is not included in the learning targets at non-adjacent levels (for example, Lv1 and Lv3), such a restriction need not necessarily be provided.

Example 3
A third embodiment of the abnormality diagnosis apparatus will be described based on FIG. Here, although the range to be learned is expanded as in the second embodiment, the range is fixed to the range of the adjacent driving level.

  That is, in the second embodiment, while an exclusive operation level is used for diagnosis, duplication of the operation level is allowed for extraction of the learning range. However, according to the second embodiment, it is necessary to reconstruct not only the level division for each driving level but also the value of driving information at the time of extraction of multivariate samples used for learning to be slightly wider than the range of each level. The processing may be complicated.

  So, in Example 3, in order to simplify extraction of the range of study object, the range of study object concerned is fixed to the range of the adjacent operation level. For example, the learning targets of the driving level “Lv1” are multivariate samples of “Lv1” and “Lv2”, and the learning targets of the driving level “Lv2” are multivariate samples of “Lv1” to “Lv3”, and the driving level of “Lv3” It is assumed that the learning target is a multivariate sample of “Lv2” to “Lv4”, and the learning target of the last driving level “Lvn” is “Lv (n−1)” and “Lvn”.

  In this way, it is possible to tag the multivariate sample with the driving level calculated from the driving information as shown in FIG. 5 instead of the driving information, so that the multivariate sample is grouped according to the driving level Can be implemented. Therefore, when learning the preparation phase, the multivariate sample groups of two or three adjacent operation levels are merged, and the processes of the principal component analysis unit 102, the statistic value calculation unit 103, and the normalization unit 104 are performed. Good.

«Functional effect»
According to the first to third embodiments, the learning of the preparation phase and the diagnosis of the diagnosis phase are divided for each operation level and the spread of the main component distribution of the preparation phase is divided for each operation level and made compact It is possible to improve the sensitivity of abnormality detection. This point will be described based on FIGS.

  FIG. 6 shows a diagnostic result in which the first one week of vibration data (time-series data) for each hour of generating time is set as the learning period L in the generator of the generated power shown in FIG. There is. Here, in FIG. 1, as a result of learning / diagnosing without dividing the level of the generated power (Y axis) “0.8” or more as driving, the result of learning / diagnosis is shown in Patent Document 2. The generator that acquired the vibration data in Fig. 1 stops power generation for a while from the 21st day of operation start (refer to X-axis), and resumes power generation on the 25th day. ing.

  FIG. 7 shows a diagnosis result obtained by level-dividing the vibration data with the generated power in the learning period L and learning without overlapping the levels, that is, the diagnosis result of the first embodiment. Here, as in FIG. 6, the results of learning / diagnosis in units of “0.2” with power generation of “0.8” or more as operation are shown.

  FIG. 8 shows diagnostic results obtained by level division of the vibration data by the generated power in the learning period L and learning by the method of the third embodiment. Here, the generated electric power “0.8” or more is divided into levels by “0.1” as operation, and the result is shown in FIG. 3 by expanding the target to learning to the adjacent level of the third embodiment and overlapping and performing learning and diagnosis There is.

  According to the diagnosis results of FIGS. 6 to 8, when the level division of FIG. 6 is not performed, attention is given to the day of the failure (the 25th day of the operation start) because the variation in the learning period L is large. I never crossed the line.

  On the other hand, in the case of level division in FIG. 7 and FIG. 8, the variation in the learning period L is suppressed more than in FIG. 6, and the caution line is exceeded on the ninth day of operation start. Therefore, according to the level division of the first and third embodiments, it is possible to detect an abnormal sign from two weeks or more before the failure. In particular, according to the third embodiment, the variation is suppressed more than in the case where the level is not generally subdivided except for the time beyond the attention line (FIG. 6).

  Therefore, according to the level divided learning in the first to third embodiments, since the range of the normal state can be limited, the sensitivity of the abnormality detection is improved, and further, the level division can be subdivided by overlapping the learning targets in the third embodiment. It is even more effective.

«Others ・ Other examples»
The present invention is not limited to the above embodiment, and can be modified within the scope of the claims. An example will be described below.

  (1) The operation information in the first to third embodiments is one numerical value information, and level division is performed based on the magnitude. However, level division is sufficient as long as multivariate samples to be learned can be appropriately classified, and therefore it is not necessary to level division from only one numerical value. For example, if the diagnosis target is a wind turbine generator, in addition to the generated power, the wind speed at that time may be used to combine them into two-dimensional level division.

  In level division by such driving information of a plurality of numerical values, a range of a learning target similar to that in FIG. 4 is expanded for each numerical value of driving information to create a multivariate sample. Here, when using the adjacent level similar to the third embodiment, the adjacent level multivariate sample groups to be merged are not two or three but four in the two-dimensional level division shown in FIG. In the case of the corner), 6 (in the case of the end), 9 (except the corner and the end).

  (2) In addition, the present invention only needs to separate and learn multivariate samples to be learned for each driving level, and the learning objects are not limited to multivariate samples created by constant Q conversion, but Fourier It may be a multivariate sample by conversion or the like.

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

DESCRIPTION OF SYMBOLS 1 ... abnormality diagnosis apparatus 100 ... preparation phase 101, 201 ... fixed Q conversion part 102, 202 ... principal component analysis part 103, 203 ... statistical value calculation part 104, 204 ... normalization part 200 ... diagnosis phase

Claims (8)

Diagnostic parameter generation means for generating a diagnostic parameter by learning a learning sample based on time series data measured in advance from a diagnosis target of abnormality diagnosis;
Diagnostic means for diagnosing an abnormality of the diagnosis target using time-lapse data newly measured from the diagnosis target using the diagnostic parameter;
An abnormality diagnosis device provided with
The diagnostic parameter creation means divides the learning sample into levels according to the driving information linked to the time series data, learns each level, and creates a diagnostic parameter.
The diagnosis means calculates a level based on driving information linked to the time series data, and diagnoses an abnormality of the diagnosis object using the diagnosis parameter according to the level. apparatus. The abnormality diagnosis device according to claim 1, wherein the level is included in the target of the learning when creating the diagnostic parameter.   The abnormality diagnosis device according to claim 2, wherein a target of the learning is wider than the level. Expand the subject of the learning to the adjacent range of the level,
4. The abnormality diagnosis apparatus according to claim 3, wherein the learning target is duplicated. An abnormality diagnosis method executed by a computer,
A diagnostic parameter creation step of creating a diagnostic parameter by learning a learning sample based on time series data measured in advance from a diagnosis target of abnormality diagnosis;
A diagnosis step of diagnosing an abnormality of the diagnosis target using time-lapse data newly measured from the diagnosis target using the diagnostic parameter;
In creating the diagnostic parameter, the learning sample is divided into levels according to the driving information linked to the time series data, and the diagnostic parameter is created by learning for each level.
In the diagnosis step, a level based on the driving information linked to the time series data is calculated, and an abnormality of the diagnosis object is diagnosed using the diagnosis parameter according to the level. Method. The method according to claim 5, wherein the level is included in the target of the learning when creating the diagnostic parameter.   The abnormality diagnosis method according to claim 6, wherein a target of the learning is wider than the level. Expand the subject of the learning to the adjacent range of the level,
The abnormality diagnosis method according to claim 7, wherein the learning target is duplicated.

Images